Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
9
9
date_observed
stringdate
1995-09-29 00:00:00
2026-11-11 00:00:00
date_basis
stringclasses
5 values
date_precision
stringclasses
4 values
benefit_status
stringclasses
5 values
title
stringlengths
4
383
summary
stringlengths
94
200
headline
stringlengths
35
72
metric
stringlengths
2
98
โŒ€
brief
stringclasses
42 values
confidence
stringclasses
4 values
claim_status
stringclasses
12 values
design
stringclasses
6 values
certainty
stringclasses
4 values
category
stringclasses
5 values
unexpected
bool
2 classes
system
stringlengths
3
475
model_version
stringclasses
230 values
developer
stringlengths
2
267
โŒ€
developer_country
stringclasses
139 values
deployer
stringlengths
3
288
โŒ€
deployer_country
stringclasses
166 values
location
stringlengths
4
309
โŒ€
case_jurisdictions
stringclasses
241 values
domain
stringlengths
3
109
segment
stringclasses
293 values
sector
stringclasses
17 values
policy_area
stringclasses
21 values
policy_subtopic
stringclasses
124 values
kind_of_ai
stringclasses
34 values
people_helped
stringclasses
126 values
global_goals
stringclasses
84 values
income_groups
stringclasses
11 values
broad_benefits
stringclasses
113 values
scsp_example
stringclasses
31 values
deepest
float64
1
5
โŒ€
reach
float64
1
5
โŒ€
duration
float64
1
5
โŒ€
distribution
float64
1
5
โŒ€
probability
float64
2
5
โŒ€
frequency
float64
1
5
โŒ€
magnitude
float64
0.11
0.68
โŒ€
groups
stringclasses
14 values
sources
int64
1
33
selected_by
stringclasses
12 values
echoes
int64
0
16
comparator_against
stringlengths
7
256
โŒ€
comparator_before
stringlengths
2
226
โŒ€
comparator_after
stringlengths
2
208
โŒ€
comparator_basis
stringclasses
2 values
later_status
stringclasses
8 values
later_date
stringlengths
10
10
โŒ€
learns_from_data
stringclasses
2 values
learns_from_data_source
stringlengths
22
280
โŒ€
value_amount
stringclasses
63 values
value_what
stringclasses
64 values
value_estimate
bool
2 classes
benefits
stringclasses
2 values
AIBD-2302
1995-09-29
source publication
day
tested
FDA approval review of the AutoPap 300 QC System: a statistical classifier's choice of Pap smears for quality control recovered up to five times the missed abnormal slides of a random choice
In the study at six laboratories that FDA reviewed, rescreening the tenth of normal Pap smears that the AutoPap 300 QC selected recovered up to five times the missed abnormal slides of a random tenth.
*AutoPap 300 QC* found up to five times more missed abnormal Pap smears
50%
null
found
Found by a regulator
controlled comparison
very low
program
false
AutoPap 300 QC System
null
NeoPath, Inc. (BD Diagnostics is the applicant in FDA's database today)
United States
NeoPath (study at three academic institutions and three commercial laboratories)
United States
United States
United States
Healthcare (cervical cytology)
null
Health
Health
Disease Prevention
Images and video
Health workers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
3
null
0
A random selection of a tenth of the smears judged normal, the quality control then required
A tenth of the missed abnormal slides recovered
Up to half of the missed abnormal slides recovered
measured
unknown
null
shown
https://pubmed.ncbi.nlm.nih.gov/9022727/
null
null
null
[]
AIBD-2307
1995-11-08
source publication
day
tested
FDA approval review of the PAPNET Testing System: a neural network's rescreening found abnormal cells in earlier negative Pap smears of 31.6% of women with later cervical lesions
In the trial that FDA reviewed, rescreening with PAPNET found abnormal cells in earlier smears, read as negative, of 72 of 228 women (31.6%) who later had a cervical lesion confirmed by biopsy.
*PAPNET* found abnormal cells that Pap smear readers had missed
31.6%
null
found
Found by a regulator
tested against a standard
very low
program
false
PAPNET Testing System
null
Neuromedical Systems, Inc. (BD Diagnostics is the applicant in FDA's database today)
United States
Neuromedical Systems (trial on stored smears from 10 medical institutions)
United States
United States
United States
Healthcare (cervical cytology)
null
Health
Health
Disease Prevention
Images and video
Women and mothers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
2
null
0
Manual screening and routine rescreening, which had reported the same smears as negative
All 487 earlier smears of the 228 women reported as negative
Abnormal cells found in earlier smears of 72 of the 228 women (31.6%)
measured
withdrawn
2009-06-04
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf/P940029.pdf
null
null
null
[]
AIBD-2304
1998-06-26
source publication
day
tested
FDA approval review of the ImageChecker M1000: the first detection aid for mammography marked 81.3% of cancers that radiologists judged visible on earlier screening films
In the studies that FDA reviewed, the ImageChecker M1000 marked 91 of 112 earlier mammograms (81.3%) on which radiologists, looking back, judged that a later cancer already called for action.
*ImageChecker M1000* marked cancers that screening readers had missed
81.3%
null
found
Found by a regulator
tested against a standard
very low
program
false
ImageChecker M1000
null
R2 Technology, Inc. (Hologic is the applicant in FDA's database today)
United States
R2 Technology (studies on screening mammograms from 13 institutions and a prospective study at five institutions)
United States
United States
United States
Healthcare (breast screening)
null
Health
Health
Disease Prevention
Images and video
Women and mothers; Health workers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
4
null
0
Screening readings without the system, which had not acted on the same earlier films
112 earlier films judged actionable by most of five radiologists, none of them acted on at the time
91 of the 112 (81.3%) correctly marked by the system
measured
challenged
2007-04-05
shown
https://spie.org/news/computer-aided-diagnosis-becomes-a-reality-in-mammography
null
null
null
[]
AIBD-2305
1998-09-25
source publication
day
reported
FDA approval review of the Logicon Caries Detector: dentists' sensitivity for decay in the dentin rose from 70.3% to 90.5% with a neural network's reading of X-rays
In the clinical study that FDA reviewed, 18 dentists found decay that had reached the dentin with 90.5% sensitivity after running the Logicon Caries Detector, against 70.3% before.
*Logicon Caries Detector* helped dentists find more decay on X-rays
90.5%
null
found
Found by a regulator
before and after
very low
aggregate result
false
Logicon Caries Detector
null
Logicon RDA, a Northrop Grumman company (Carestream Dental is the applicant in FDA's database today)
United States
18 dentists in 16 private practice offices in the United States
United States
United States
United States
Healthcare (dentistry)
null
Health
Health
Disease Prevention
Images and video
Health workers; Patients
3
High income
physical_health
Medical diagnosis
2
3
1
null
3
1
null
organizational
2
null
0
The same dentists' diagnoses of the same surfaces before they ran the software
Mean sensitivity 70.3%, specificity 88.6%
Mean sensitivity 90.5%, specificity 88.3%
measured
persisted
2011-03-01
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf/P980025.pdf
null
null
null
[]
AIBD-0798
2001-07-12
source publication
day
tested
FDA found that with the RapidScreen RS-2000, radiologists detected lung cancers of 9.5-27.5 mm on chest X-rays better than without it
In the reader study that FDA reviewed for the RapidScreen RS-2000, 15 radiologists raised their area under the ROC curve for lung cancers from 0.8288 to 0.8654 with the system.
*RapidScreen RS-2000* helped radiologists spot small lung cancers
the area under the ROC curve rose from 0.8288 to 0.8654
null
found
Found by a regulator
controlled comparison
very low
aggregate result
false
RapidScreen RS-2000
null
Deus Technologies, LLC (listed as Riverain Medical Group in the FDA database)
United States
Deus Technologies (reader study with 15 radiologists at Georgetown University Medical Center)
United States
Washington, D.C., United States
United States
Healthcare
null
Health
Health
Disease Prevention
Images and video
Health workers
3
High income
physical_health
Medical diagnosis
null
2
1
null
null
1
null
organizational
2
null
0
The same radiologists reading the same films without the system
Area under the ROC curve 0.8288, sensitivity 71%
Area under the ROC curve 0.8654, sensitivity 78%
measured
challenged
2006-05-04
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf/P000041B.pdf
null
null
null
[]
AIBD-0797
2002-01-31
source publication
day
tested
FDA concluded that Second Look would have cut missed cancers on earlier screening mammograms by at least 26.2%, with work-ups rising from 16.6% to 17.2%
FDA concluded that Second Look's marks would have cut missed cancers on earlier screening mammograms by at least 26.2%. In a prospective test, work-ups rose from 16.6% to 17.2%.
*Second Look* would have cut missed breast cancers by at least a quarter
cut missed cancers on earlier screening mammograms by at least 26.2%
null
found
Found by a regulator
no comparison
very low
aggregate result
false
Second Look
null
CADx Medical Systems, Inc. (listed as iCAD, Inc. in the FDA database)
Canada
CADx Medical Systems (ROSE-1 study on earlier screening mammograms from 17 institutions and ROSE-2 study at 5 institutions)
Canada
null
null
Healthcare
null
Health
Health
Disease Prevention
Images and video
Health workers
3
null
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
2
null
0
Unaided readings of the same earlier mammograms, which had missed every one of the cancers
306 retrospectively visible false negatives at most
At least 80.3 of them estimated to be worked up with the marks
assumed
scaled
2005-03-15
not shown
https://www.accessdata.fda.gov/cdrh_docs/pdf/P010034B.pdf
null
null
null
[]
AIBD-0846
2002-05-01
event
month
reported
EPA approves a software monitor for NOx on Braintree's Potter II Unit 3 after finding that it matched reference measurements on both of its fuels
EPA found that a predictive emission monitor on Braintree's Potter II Unit 3 matched reference NOx measurements on gas and oil. It estimated that it would cost less than half as much as an analyzer.
Braintree's *predictive emission monitor* passes EPA's tests for NOx
cost less than half as much as an analyzer
null
found
Found by a regulator
tested against a standard
low
aggregate result
false
Predictive emission monitoring system (PEMS) on Potter II Unit 3
null
Not named in the letter
null
Braintree Electric Light Department (BELD), Norton P. Potter II Station Unit 3
United States
Braintree, Massachusetts, United States
United States
Environment and climate
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
Value above replacement
1
1
4
null
3
5
null
organizational
1
null
0
Hourly NOx readings from a temporary reference method system over the same hours, judged by EPA's three statistical tests
A linear correlation coefficient (r) above 0.8 required
r = 0.8524 on natural gas and r = 0.9211 on fuel oil
measured
unknown
null
shown
https://www.epa.gov/sites/default/files/2015-05/documents/potter.pdf
$12,950
Estimated yearly saving of the PEMS over an analyzer even if outside contractors run the extra audits
true
[]
AIBD-1036
2004-01-01
event
year
reported
GAO finds that FHA's TOTAL Scorecard gave lenders more consistent underwriting decisions than the scorecards that it replaced
In April 2006 GAO found that FHA's TOTAL Scorecard gave lenders less paperwork than the two scorecards that it replaced. Borrowers had less. Underwriting decisions were more consistent.
*TOTAL Scorecard* makes FHA's underwriting more consistent
null
null
found
Found by an auditor
no comparison
very low
program
false
Technology Open to Approved Lenders (TOTAL) Scorecard
null
Federal Housing Administration (HUD), with contractor Unicon Research Corporation and subcontractor Fair Isaac Corporation
United States
Federal Housing Administration and its direct endorsement lenders
United States
United States (national)
United States
Housing finance: mortgage insurance underwriting
null
Finance
Domestic Commerce
Consumer Finance
Prediction from data
Consumers
10; 11
High income
operational_improvement
Value above replacement
1
null
4
4
3
5
null
organizational; personal
1
null
0
The FHA versions of Freddie Mac's Loan Prospector and Fannie Mae's Desktop Underwriter scorecards, used before 2004
at times the results of their assessments differed, which resulted in the same loan being accepted by one scorecard and referred by the other
TOTAL limits the number of loans that need to be approved manually because it provides consistent automatic underwriting decisions
assumed
persisted
2022-09-27
shown
https://www.govinfo.gov/content/pkg/GAOREPORTS-GAO-06-435/pdf/GAOREPORTS-GAO-06-435.pdf
null
null
null
[]
AIBD-0843
2004-07-01
event
month
reported
EPA finds that SmartCEM-75 software tracked NOx at Dearborn Industrial Generation closely enough to stand in for hardware analyzers
EPA found that SmartCEM-75 software matched a NOx analyzer on a Dearborn boiler over 720 hours. It estimated that on a turbine there the software would cost under half as much as an analyzer.
*SmartCEM-75* stands in for NOx analyzers at a Dearborn plant
matched a NOx analyzer on a Dearborn boiler over 720 hours
null
found
Found by a regulator
tested against a standard
low
aggregate result
false
SmartCEM-75 predictive emission monitoring system (PEMS)
null
CMC Solutions, L.L.C.
United States
Dearborn Industrial Generation (DIG), on boiler Unit BL1100 and turbine Unit GTP1
United States
Dearborn, Michigan, United States
United States
Environment and climate
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
Value above replacement
1
1
4
null
3
5
null
organizational
2
null
0
Hourly readings from the unit's certified NOx analyzer over the same 720 operating hours, judged by EPA's three statistical tests
A linear correlation coefficient (r) above 0.8 required
r = 0.9678 over all 720 hours and r = 0.9286 over 98 startup and shutdown hours
measured
unknown
null
shown
https://www.epa.gov/system/files/documents/2024-09/dig-55088-original-petition-response-bl1100-2006.9.6.pdf
$12,950
Estimated yearly saving of the PEMS on Unit GTP1 over an analyzer even if outside contractors run the extra audits
true
[]
AIBD-0679
2005-01-01
event
year
reported
ANAO finds that the Passport Office's facial recognition matching caught many identity frauds that would otherwise have gone undetected
The ANAO found that facial recognition matching of passport applications detected many identity frauds that would not otherwise have been found, about 22 recorded a year.
Passport photo *FR matching* exposes otherwise hidden identity fraud
about 22 recorded a year
null
found
Found by an auditor
no comparison
very low
aggregate result
false
Australian Passport Office facial recognition (FR) matching of applicants' photos against its image database (one-to-many)
null
A commercial supplier of the biometric algorithm (not named in the report)
null
Australian Passport Office, Department of Foreign Affairs and Trade
Australia
Australia (national)
Australia
Identity documents (passport issuing)
null
Public sector
International Affairs
Diplomats
Images and video
Public servants
16
High income
privacy_protection
Fraud detection
2
1
4
null
3
3
null
organizational; societal
1
null
0
null
null
null
null
persisted
2025-11-18
not shown
https://www.anao.gov.au/sites/default/files/201112%20Audit%20Report%20No%2033.pdf
null
null
null
[]
AIBD-0796
2006-03-16
source publication
day
mixed
FDA finds that adding the LUMA Cervical Imaging System after colposcopy found 22.0% more women with high-grade lesions, with extra biopsies
In a study of 193 women that FDA reviewed, biopsies directed by LUMA after colposcopy found high-grade lesions in 9 more women, a 22.0% relative gain.
*LUMA* imaging found more high-grade cervical lesions after colposcopy
a 22.0% relative gain
null
found
Found by a regulator
controlled comparison
very low
aggregate result
false
LUMA Cervical Imaging System
null
MediSpectra, Inc.
United States
Seven colposcopy centers with 16 colposcopists in the intended use study
null
null
null
Healthcare
null
Health
Health
Disease Prevention
Images and video
Health workers; Women and mothers
3
null
physical_health
Medical diagnosis
2
2
1
null
2
1
null
personal
2
null
0
Colposcopy alone in the same women
41 women with CIN 2,3+ found by colposcopy (21.2%)
9 more women found with LUMA (4.7%)
measured
withdrawn
null
not shown
https://www.accessdata.fda.gov/cdrh_docs/pdf4/P040028B.pdf
null
null
null
[]
AIBD-0895
2006-04-05
source publication
day
reported
EPA approved a neural-network model of emissions at the Mahrt Paper Mill after its NOx output passed all three federal statistical tests
EPA found that a neural network's NOx estimates for a mill turbine passed all three federal statistical tests. It estimated that two such systems cost less than half as much as hardware monitors.
EPA approves *Software CEM* to monitor NOx from a paper mill's turbine
less than half the cost
null
found
Found by a regulator
no comparison
very low
aggregate result
false
Pavilion Technologies Software CEM predictive emission monitoring system (PEMS)
null
Pavilion Technologies
United States
MeadWestvaco Coated Board, Inc. (Mahrt Mill)
United States
Cottonton, Alabama, United States
United States
Air emissions monitoring
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
Value above replacement
1
1
4
null
3
5
null
organizational
1
null
0
Two continuous emission monitoring systems (CEMS), the next least costly option under EPA's cost model
$125,400 a year for two CEMS
approximately $51,700 a year for two PEMS with in-house audits
assumed
unknown
null
shown
https://www.epa.gov/sites/default/files/2015-05/documents/meadwestvaco_coated_board_inc.pdf
$36,650
Estimated yearly saving of two PEMS over two CEMS at the mill, even if outside contractors run the audits
true
[]
AIBD-0795
2007-01-01
event
year
tested
FDA finds that MelaFind detected 98.3% of melanomas among lesions chosen for biopsy, at a low specificity that still beat dermatologists'
In FDA's review of MelaFind, the device detected 112 of 114 melanomas (98.3%), with 10.6% specificity against 5.5% for the study dermatologists.
*MelaFind* caught nearly all melanomas at low specificity
112 of 114 melanomas (98.3%)
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
MelaFind
null
MELA Sciences, Inc.
United States
Seven dermatology sites in the blinded study (Protocol 20061) and physicians in a web-based reader study (Protocol 20063)
United States
Seven dermatology sites, United States
United States
Healthcare
null
Health
Health
Disease Prevention
Images and video
Health workers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
2
null
0
Study dermatologists' specificity on the same lesions
5.5%
10.6%
measured
withdrawn
2024-06-15
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf9/P090012B.pdf
null
null
null
[]
AIBD-0845
2007-01-01
event
year
reported
EPA finds that SmartCEM-75 models matched NOx analyzers on all eight T.H. Wharton turbines, where NRG has reported emissions with them since 2009
EPA found that SmartCEM-75 models matched NOx analyzers over 720 hours each on eight NRG turbines in Houston. NRG has reported NOx with them since 2009 for six units. Two more followed in 2011.
*SmartCEM-75* models match NOx analyzers on eight Houston turbines
matched NOx analyzers over 720 hours each
null
found
Found by a regulator
tested against a standard
low
aggregate result
false
SmartCEM-75 predictive emission monitoring systems (PEMS), one model per turbine
null
CMC Solutions
United States
NRG Texas Power (NRG Energy), T.H. Wharton Generating Station Units THW31 to THW34 and THW41 to THW44
United States
Houston, Texas, United States
United States
Environment and climate
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
null
1
1
4
null
3
5
null
organizational
2
null
0
Hourly readings from certified NOx analyzers on the same units over 720 operating hours each, judged by EPA's three statistical tests
A linear correlation coefficient (r) above 0.8 required
r = 0.936 to r = 0.999 across the eight units
measured
persisted
2025-05-21
shown
https://www.epa.gov/sites/default/files/2015-05/documents/t.h._wharton.pdf
null
null
null
[]
AIBD-0813
2007-02-06
source publication
day
tested
FDA found that the 70-gene signature of MammaPrint separated patients with node-negative breast cancer by risk of metastasis in an independent validation
In FDA's review of MammaPrint, 95% of low risk patients were free of distant metastasis at five years, against 78% of high risk patients. The validation covered 302 European patients.
*MammaPrint* separated node-negative breast cancers by risk of spread
95% of low risk and 78% of high risk patients were free of distant metastasis
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
MammaPrint
null
Agendia BV
Netherlands
TRANSBIG consortium (independent validation on tumors from five European centers)
null
Villejuif and Saint-Cloud, France; Stockholm, Sweden; London and Oxford, United Kingdom
France | Sweden | United Kingdom
Healthcare
null
Health
Health
Disease Prevention
Prediction from data
Patients
3
High income
physical_health
null
null
null
null
null
null
null
null
organizational
1
null
0
Clinical risk assessment with Adjuvant on-line
Unadjusted hazard ratio of 2.32 (95% CI: 1.35 to 4.00)
Hazard ratio of 2.13 (95% CI: 1.19 to 3.82) after adjustment for clinical risk
measured
persisted
2017-03-21
shown
https://www.accessdata.fda.gov/cdrh_docs/reviews/K062694.pdf
null
null
null
[]
AIBD-0844
2007-05-01
event
month
reported
EPA finds that SmartCEM-75 software matched a NOx analyzer on Great River Energy's Cambridge Station turbine, including startups
EPA found that SmartCEM-75 software, trained on a Minnesota turbine's own data, matched a certified NOx analyzer over 720 hours, including startups and shutdowns. EPA approved it on 2 June 2009.
*SmartCEM-75* matches a NOx analyzer on a Minnesota turbine
matched a certified NOx analyzer over 720 hours
null
found
Found by a regulator
tested against a standard
low
aggregate result
false
SmartCEM-75 predictive emission monitoring system (PEMS)
null
CMC Solutions
United States
Great River Energy, Cambridge Station Unit 2
United States
Cambridge, Minnesota, United States
United States
Environment and climate
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
null
1
1
null
null
3
5
null
organizational
1
null
0
Hourly readings from the unit's certified NOx analyzer over the same 720 operating hours, judged by EPA's three statistical tests
A linear correlation coefficient (r) above 0.8 required
r = 0.9521 over all 720 hours and r = 0.8541 over the 542 non-DLN startup and shutdown hours
measured
unknown
null
shown
https://www.epa.gov/sites/default/files/2015-05/documents/cambridge_station.pdf
null
null
null
[]
AIBD-0726
2008-01-01
event
year
reported
CrimeSolutions finds that Philadelphia probationers forecast as low risk by a random forests model did no worse in caseloads of about 400
In a 2007-08 Philadelphia trial rated Promising by CrimeSolutions, people whom a random forests model forecast as low risk were supervised in caseloads of about 400 with no rise in reoffending.
A risk model lets Philadelphia supervise low-risk probationers lightly
a caseload of about 400 people
null
found
Found by an assessment body
randomized trial
moderate
aggregate result
false
Random forests risk forecasting model (modified from Berk et al. 2009)
null
Richard Berk and colleagues (Berk et al. 2009), with the Philadelphia Adult Probation and Parole Department
United States
Philadelphia Adult Probation and Parole Department, First Judicial District of Pennsylvania
United States
Philadelphia, Pennsylvania, United States (western and northeastern regions)
United States
Criminal justice (probation and parole)
null
Public sector
Law and Crime
Prisons
Prediction from data
Public servants
16
High income
operational_improvement
Enhanced productivity
1
3
4
null
3
null
null
organizational; personal
1
null
0
Standard regional supervision of people forecast as low risk
Almost the same contact as in the previous year
Contacts reduced by about 45 percent
measured
persisted
null
shown
https://crimesolutions.ojp.gov/ratedprograms/philadelphia-pa-low-intensity-community-supervision-experiment
null
null
null
[]
AIBD-0673
2008-03-01
event
year
reported
SSA's Inspector General finds that 93% of claims picked by the Quick Disability Determinations model were allowed. Decisions took 18 days on average
SSA's Inspector General found that 537 of 575 sampled claims picked by the QDD model, about 93%, were allowed. It found that decisions took 18 days on average.
*Quick Disability Determinations* speed up likely disability claims
537 claims, about 93%, were allowed
null
found
Found by an auditor
no comparison
very low
aggregate result
false
Quick Disability Determination (QDD) predictive model
null
IBM, which originally designed the model according to SSA, with later changes by the Social Security Administration
United States
Social Security Administration with state Disability Determination Services
United States
United States (national)
United States
Social security (disability benefits)
null
Accessibility
Social Welfare
Disabled Assistance
Prediction from data
People with disabilities
1; 10
High income
opportunity_access
Insurance and benefit access
1
4
1
2
3
1
0.16
personal
2
null
0
SSA's recommended timeframe for QDD claims
20 days or fewer
18 days on average for all sampled claims and 15 days for allowed claims
measured
persisted
2026-05-18
not shown
https://web.archive.org/web/20250718071131id_/https://oig-files.ssa.gov/audits/full/A-01-10-21080.pdf
null
null
null
[]
AIBD-1040
2008-04-24
source publication
day
reported
The ANAO concludes that the ATO's analytics group improved compliance with new quantitative methods
In April 2008 the ANAO concluded that the ATO's analytics group had made a valuable contribution. Its methods improved compliance. They improved services. The ATO used its resources better.
Auditors credit the ATO's *advanced analytics* with better compliance
null
null
studied
Independently studied
no comparison
very low
program
false
ATO advanced analytics: data-mining and predictive models of the analytics group in the Office of the Chief Knowledge Officer
null
Australian Taxation Office, Office of the Chief Knowledge Officer (analytics group)
Australia
Australian Taxation Office (ATO)
Australia
Australia (national)
Australia
Tax administration: compliance risk
null
Public sector
Government Operations
Tax Administration
Prediction from data
null
16
High income
operational_improvement
Enhanced productivity
1
1
4
null
3
null
null
organizational
1
null
0
null
null
null
null
unknown
null
shown
https://web.archive.org/web/20221013050901/http://www.anao.gov.au/work/performance-audit/australian-taxation-offices-use-data-matching-and-analytics-tax
null
null
null
[]
AIBD-0808
2008-07-30
source publication
day
tested
FDA found that the machine-learned Pathwork Tissue of Origin Test agreed with the known diagnosis for most of 545 tumor specimens
In FDA's review of the Pathwork Tissue of Origin Test, results agreed with the available diagnosis for 86.3% of metastatic and 92.0% of primary tumor specimens among 545.
*Pathwork Tissue of Origin Test* matched the known origin of most tumors
86.3% of metastatic and 92.0% of primary tumor specimens
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
Pathwork Tissue of Origin Test
null
Pathwork Diagnostics, Inc.
United States
Pathwork Diagnostics (clinical validation at four processing laboratories)
United States
null
null
Healthcare
null
Health
Health
Drug Industry
Prediction from data
Patients
3
null
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
2
null
0
The available diagnosis for each specimen
Available diagnosis (reference)
Agreement for 86.3% of metastatic and 92.0% of primary tumor specimens
measured
persisted
2013-08-26
shown
https://www.accessdata.fda.gov/cdrh_docs/reviews/K080896.pdf
null
null
null
[]
AIBD-0794
2008-09-01
event
month
mixed
FDA finds that surgeons using the MarginProbe System removed more cancer at lumpectomy margins, with 5 percentage points fewer re-operations
In a randomized trial that FDA reviewed, MarginProbe raised complete re-excision of positive margins from 22.4% to 71.8%. A 5-point drop in re-operations was not statistically significant.
*MarginProbe* helped surgeons remove cancer left at lumpectomy margins
from 22.4% to 71.8%
null
found
Found by a regulator
randomized trial
moderate
aggregate result
false
MarginProbe System
null
Dune Medical Devices, Inc.
United States
Surgical sites in the randomized trial in Israel and the United States (596 women undergoing lumpectomy)
Israel and United States
Israel and United States
Israel | United States
Healthcare
null
Health
Health
Disease Prevention
Prediction from data
Women and mothers; Health workers
3
High income
physical_health
Medical diagnosis
2
3
1
2
2
1
0.12
organizational; personal
2
null
0
Standard care in the randomized control arm
25.8% of patients had a further excision
20.8%
measured
persisted
2018-07-31
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf11/P110014B.pdf
null
null
null
[]
AIBD-0214
2009-01-01
event
year
mixed
U.S. high schools using Cognitive Tutor Algebra I see a second-year gain that loses significance after correction
Students using Cognitive Tutor Algebra I scored 0.21 standard deviations higher in the second year, with no statistically significant difference after correction for multiple comparisons.
Algebra software shows an uncertain gain in a trial's second year
scored 0.21 standard deviations higher in the second year
null
found
Found by an assessment body
randomized trial
moderate
aggregate result
false
Cognitive Tutor Algebra I, with an ACT-R-based cognitive model, classroom instruction, and curriculum materials
null
Carnegie Learning
United States
Participating public and Catholic parochial schools in 51 districts across seven states (study sites)
null
Seven states (147 schools), United States
United States
Education
Secondary-school algebra
Education
Education
Excellence
Prediction from data
Students
4
High income
incentivization
Increased/personalized knowledge
2
3
3
null
2
4
null
personal
9
null
0
Schools using their usual algebra curriculum
Comparison schools in the same implementation year
Second high-school cohort: +0.21 standard deviations; p = .04 before correction, not significant at the Clearinghouseโ€™s corrected threshold of .025
measured
challenged
2016-12-21
not shown
https://doi.org/10.3758/BF03194060
null
null
null
[]
AIBD-0894
2009-01-01
event
year
mixed
IIHS researchers found that Google's self-driving test cars had a lower crash rate than human drivers, but the gap was not statistically significant
IIHS researchers found that Google's test cars had 2.19 police-reportable crashes per million miles in Mountain View, against 6.06 for human drivers. The gap was not statistically significant.
Google's *self-driving cars* had a lower crash rate in Mountain View
2.19 police-reportable crashes per million miles
null
studied
Independently studied
controlled comparison
low
aggregate result
false
Google self-driving car (test fleet under employee supervision)
null
Google
United States
Google (test fleet on public roads)
United States
Mountain View, California, United States
United States
Road vehicle crash avoidance / automated driving
null
Safety and rescue
Transportation
Highways
Machines that act
null
3; 11
High income
physical_health
Hazard reduction
null
null
4
null
2
null
null
societal
2
null
0
Police-reported crash rates of human drivers in Mountain View
6.06 police-reportable crashes per million vehicle miles traveled
2.19 police-reportable crashes per million vehicle miles traveled
measured
scaled
2023-08-10
shown
https://web.archive.org/web/2016id_/https://static.googleusercontent.com/media/www.google.com/en//selfdrivingcar/files/reports/report-0616.pdf
null
null
null
[]
AIBD-0168
2009-11-01
event
month
reported
NCAR's DICast cuts the error of Xcel Energy's wind forecasts to 10.8%, with savings that Xcel estimates at US$49 million
NCAR's adaptive system for blending forecasts (DICast), used by Xcel Energy since 2009, cut mean wind forecast error from about 18% to 10.8% by 2014. Xcel estimated that it saved US$49 million.
NCAR's DICast cuts the error of Xcel's wind forecasts to 10.8%
cut mean wind forecast error from about 18% to 10.8% by 2014
null
reported
Independently reported
before and after
very low
aggregate result
false
NCAR wind power forecasting system with the DICast adaptive forecast blending engine
null
National Center for Atmospheric Research (Research Applications Laboratory)
United States
Xcel Energy
United States
United States
United States
Energy
Utility wind generation forecasting
Energy and infrastructure
Energy
Alternative & Renewable
Prediction from data
Consumers
7
High income
environmental_impact; operational_improvement
Weather and environmental forecasting
2
1
4
null
3
5
null
organizational; societal
5
null
0
Xcel's wind forecasts in 2008, before NCAR's system
about 18 percent for 18 to 42 hours ahead
10.8 percent
measured
persisted
2017-12-04
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC7898128/
US $49 million
saved through forecasting improvements, cumulative to 2015
true
[]
AIBD-0724
2010-01-01
event
year
reported
The What Works Clearinghouse finds that the ITSS tutor has positive effects on reading comprehension in two randomized trials with 6,724 students
The What Works Clearinghouse rates ITSS as having positive effects on comprehension. Across two randomized studies with 6,724 students in grades 4, 5 and 7, the improvement index was +9.
*ITSS* tutoring software raises reading comprehension in two trials
The improvement index was +9 percentile points
null
found
Found by an assessment body
randomized trial
moderate
aggregate result
false
Web-Based Intelligent Tutoring for the Structure Strategy (ITSS)
null
ITSS developers led by Kausalai Wijekumar (contact at Texas A&M University)
United States
Rural and suburban schools in two US states (grades 4, 5 and 7) in the two rated studies
United States
United States (two states that the report does not name)
United States
Education
null
Education
Education
Elementary & Secondary
null
Students
4
High income
incentivization
Increased/personalized knowledge
2
3
3
4
3
4
0.42
personal
2
null
0
Classrooms using the usual language arts curriculum
null
Improvement index of +9 percentile points
measured
persisted
2015-08-01
not shown
https://ies.ed.gov/ncee/WWC/Docs/InterventionReports/wwc_ALitss_IR_apr2020.pdf
null
null
null
[]
AIBD-0905
2010-05-12
source publication
day
tested
FDA found that white cell differentials from Medica's EasyCell correlated 0.93 to 0.99 with manual microscopy after a neural network preclassified the cells
In FDA's review, technologists reviewed EasyCell's white cell differentials on 304 specimens. These correlated with manual microscopy at 0.93 to 0.99, including 0.99 for neutrophils.
*EasyCell*'s white cell counts track manual microscopy in FDA's review
correlation coefficients of 0.99 for neutrophils
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
EasyCell Cell Locator (EasyCell)
null
Medica Corporation
United States
Medica Corporation (method comparison study at three sites)
United States
Three study sites (not named in the decision summary)
null
Healthcare (laboratory hematology)
null
Health
Health
Drug Industry
Images and video
Health workers
3
null
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
2
null
0
Manual WBC differential by microscopy on the same specimens
Manual 200-cell differential (reference)
Correlation coefficient 0.99 for neutrophils, 0.98 for lymphocytes, 0.97 for eosinophils and 0.93 for monocytes
measured
persisted
2023-04-24
shown
https://www.accessdata.fda.gov/cdrh_docs/reviews/K092116.pdf
null
null
null
[]
AIBD-0910
2010-07-01
event
month
tested
EPA approves trained Pavilion8 models in place of hardware analyzers for NOx on seven turbines in Arkansas after they passed its tests of accuracy
EPA found that trained Pavilion8 models matched hardware analyzers for NOx on seven turbines in Arkansas, with relative accuracy of 1.38% to 3.15%. It approved them in their place.
EPA lets *Pavilion8 PEMS* stand in for NOx analyzers in Arkansas
relative accuracy of 1.38% to 3.15%
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
Pavilion8 Software PEMS (predictive emission monitoring systems) on Units G1 through G7
null
Rockwell Automation (maker of Pavilion8, which the letter names only as a product)
United States
Arkansas Electric Cooperative Corporation (AECC), Harry L. Oswald Generating Station
United States
Pulaski County, Arkansas, United States
United States
Air emissions monitoring (power generation)
null
Environment
Environment
Air Pollution
Prediction from data
null
null
High income
operational_improvement
null
null
1
1
null
null
1
null
organizational
1
null
0
Certified hardware NOx analyzers (CEMS) on the same turbines
null
Relative accuracy of 1.38% to 3.15% across the seven units, with all three statistical tests passed
measured
persisted
2015-03-16
shown
https://www.epa.gov/sites/default/files/2015-05/documents/harry_l._oswald_generating_station_0.pdf
null
null
null
[]
AIBD-0716
2011-02-01
event
month
reported
The What Works Clearinghouse finds that DreamBox Learning had potentially positive effects on math for students from kindergarten to first grade
The What Works Clearinghouse rated DreamBox Learning as having potentially positive effects on math, an improvement index of +4, in a trial with 557 students from kindergarten to first grade.
*DreamBox Learning* lifts early math scores in a San Jose trial
an improvement index of +4
null
found
Found by an assessment body
randomized trial
moderate
aggregate result
false
DreamBox Learning (adaptive online K-5 mathematics program)
null
DreamBox Learning, Inc.
United States
Rocketship Education (three charter elementary schools in San Jose, California)
United States
San Jose, California, United States
United States
Education
null
Education
Education
Elementary & Secondary
null
Students; Children
4
High income
incentivization
Increased/personalized knowledge
2
3
3
4
3
5
0.48
personal
1
null
0
Students randomly assigned to an online literacy program during the same lab time
156.20 (comparison group mean, MAP overall math score)
159.00 (DreamBox group mean)
measured
persisted
2023-10-12
not shown
https://ies.ed.gov/ncee/WWC/Docs/InterventionReports/wwc_dreambox_121013.pdf
null
null
null
[]
AIBD-2228
2011-12-01
source publication
month
reported
Mortality reduction by heart rate characteristic monitoring in very low birth weight neonates: a randomized trial
In a randomized trial of 3,003 premature infants in nine U.S. intensive care units, showing clinicians a heart rate index cut mortality from 10.2% to 8.1%.
*HeRO* heart rate monitoring cuts deaths of premature infants in a trial
from 10.2% to 8.1%
null
studied
Independently studied
randomized trial
moderate
aggregate result
false
HeRO heart rate characteristics monitor
null
University of Virginia and Medical Predictive Science Corporation
United States
Nine neonatal intensive care units in the United States
United States
United States
United States
Neonatal intensive care (early warning of sepsis from heart rate)
null
Health
Health
Infants and Children
Prediction from data
Children; Health workers
3
High income
physical_health
Medical diagnosis
5
3
5
null
3
5
null
personal
4
null
0
Infants whose index was masked from clinicians, by individual randomization
10.2% mortality
8.1% mortality
measured
challenged
2023-12-14
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC3215822/
null
null
null
[]
AIBD-0725
2012-01-01
event
year
reported
CrimeSolutions finds that patrols guided by the ETAS crime-forecasting model, later sold as PredPol, cut crime volume in a Los Angeles trial
In a 2011-2013 LAPD trial rated Promising by CrimeSolutions, patrols sent by the ETAS model saw 7.4 percent less crime per 1,000 patrol minutes than patrols sent by analysts.
*PredPol* maps cut property crime per patrol minute in an LAPD trial
an average 7.4 percent reduction in crime volume per thousand minutes of patrol time
null
found
Found by an assessment body
randomized trial
high
aggregate result
false
Epidemic-type aftershock sequence (ETAS) crime forecasting model, known as PredPol (later Geolitica)
null
PredPol (later Geolitica), from the ETAS model of George Mohler and colleagues
United States
Los Angeles Police Department (Foothill, North Hollywood and Southwest divisions)
United States
Los Angeles, California, United States
United States
Policing
null
Safety and rescue
Law and Crime
Police
Prediction from data
Public servants
16
High income
physical_health
Crime reduction or prevention
2
null
3
null
3
5
null
societal
2
null
0
Patrols sent to boxes chosen by LAPD crime analysts
null
An average 7.4 percent reduction in crime volume per thousand minutes of patrol time
measured
withdrawn
null
shown
https://crimesolutions.ojp.gov/ratedprograms/predictive-policing-model-los-angeles-calif
null
null
null
[]
AIBD-0828
2012-02-24
source publication
day
intended
A New York federal court approved predictive coding with Axcelerate for over three million emails, finding it superior to manual review
In February 2012 a US magistrate judge approved predictive coding with Recommind's Axcelerate for over three million emails. He found that it beat manual review.
Federal court approves *Axcelerate* for three million emails
over three million emails
null
found
Found by a court
null
very low
benefit created
false
Axcelerate software by Recommind (predictive coding)
null
Recommind
United States
MSL Group (defendant), with its counsel Jackson Lewis LLP
United States
New York, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
Linear manual review or keyword searches
null
null
assumed
litigated
null
shown
https://static.case.law/frd/287/html/0182-01.html
null
null
null
[]
AIBD-0988
2012-04-23
source publication
day
intended
The Circuit Court for Loudoun County, Virginia, let the Landow defendants use predictive coding in the Dulles Jet Center cases over objections
In April 2012 a Virginia state court let defendants use predictive coding over objections. The suits concerned a hangar collapse at Dulles Jet Center. The order gives no reasons.
Virginia court lets defendants use *predictive coding* over objections
null
null
found
Found by a court
null
very low
benefit created
false
Predictive coding
null
Not named in the order or in the defendants' memorandum
null
Landow Aviation Limited Partnership, Landow Aviation I, Inc. and Landow & Company Builders, Inc. (defendants)
United States
Loudoun County, Virginia, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
2
null
0
null
null
null
null
persisted
2013-01-21
shown
https://www.ediscoverylaw.com/wp-content/uploads/2013/11/MemoSupportPredictiveCoding.pdf
null
null
null
[]
AIBD-2296
2012-07-01
event
month
reported
Discovering foodborne illness in online restaurant reviews
New York City's health department credits a classifier of Yelp restaurant reviews with a part in identifying 10 outbreaks and 8,523 complaints of foodborne illness since July 2012.
*Yelp review classifier* helps New York City find ten outbreaks
10 outbreaks
null
studied
Independently studied
no comparison
very low
protective response
false
Yelp review classifier for foodborne illness
null
Columbia University, with the New York City Department of Health and Mental Hygiene
United States
New York City Department of Health and Mental Hygiene
United States
New York City, United States
United States
Public health (surveillance of foodborne illness)
null
Health
Agriculture
Food Inspection & Safety
Text and language
Consumers; Public servants
3
High income
physical_health; operational_improvement
Hazard reduction
2
null
null
null
3
null
null
organizational; societal
2
null
0
null
null
null
null
unknown
null
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC7647154/
null
null
null
[]
AIBD-1201
2012-10-01
event
month
mixed
An NIJ-funded evaluation shows that ShotSpotter alerts brought Kansas City police to gunfire sooner and raised evidence recovery in both cities, with no gain in crime control
In an NIJ-funded evaluation, Kansas City officers reached ShotSpotter alerts in 266.82 s against 390.87 s for citizen calls. NIBIN collection ran 30.4% above control. Crime did not fall.
*ShotSpotter* got police to gunfire sooner, but crime did not fall
266.82 s against 390.87 s
null
found
Found by an evaluation office
controlled comparison
low
aggregate result
false
ShotSpotter gunshot detection technology (GDT), with alerts confirmed at the Real-Time Incident Review Center
null
ShotSpotter, Inc. (renamed SoundThinking, Inc. in 2023)
United States
Kansas City Police Department and Chicago Police Department
United States
Kansas City, Missouri and Chicago, Illinois, United States
United States
Policing (gunfire detection)
null
Safety and rescue
Law and Crime
Police
Speech and sound
Public servants
16
High income
physical_health; operational_improvement
Crime reduction or prevention
2
2
4
4
2
5
0.42
organizational; societal
2
null
0
Citizen calls for service about shots fired in Kansas City
390.87 s to arrive on scene, stopping 54.62 m from the gunfire
266.82 s to arrive on scene, stopping 35.66 m from the gunfire
measured
withdrawn
2026-11-02
shown
https://www.soundthinking.com/law-enforcement/gunshot-detection-technology/
null
null
null
[]
AIBD-0987
2013-02-01
source publication
day
reported
A California federal court found that H5's algorithm-driven review of almost 12 million records was more efficient, holding its fees reasonable
In February 2013 a US district judge found that H5's algorithm-driven review of almost 12 million records was a more efficient method. The judge held its $2,829,349.10 cost reasonable.
Court credits *H5* review of almost 12 million records as efficient
almost 12 million records
null
found
Found by a court
no comparison
very low
benefit created
false
H5's computer-assisted, algorithm-driven document review (proprietary technology)
null
H5
United States
Qualcomm Incorporated and its co-defendants, through their counsel Cooley LLP, with contract reviewers from Black Letter Discovery
United States
San Diego, California, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
1
2
1
null
3
1
null
organizational
1
null
0
Manual review of the records
null
null
assumed
litigated
2014-03-18
not shown
https://www.govinfo.gov/content/pkg/USCOURTS-casd-3_08-cv-01992/pdf/USCOURTS-casd-3_08-cv-01992-9.pdf
null
null
null
[]
AIBD-0054
2013-03-01
event
month
reported
Myoelectric pattern recognition with augmented reality relieves phantom limb pain of 48 years in a treatment-resistant amputee
BioPatRec decoded the signals of the muscles in a man's stump to move a limb in augmented reality. His phantom limb pain of 48 years eased to periods without pain.
Amputee with 48 years of phantom limb pain reaches pain-free periods
after ten weeks near-absent pain developing into completely pain-free periods
null
reported
Independently reported
before and after
very low
benefit created
false
Phantom motor execution: BioPatRec myoelectric pattern recognition (LDA and MLP classifiers) driving augmented/virtual reality and a racing game
BioPatRec, LDA one-vs-one and MLP per degree of freedom
Chalmers University of Technology, Biomedical Engineering Division (academic lab; study part-funded by Integrum AB)
Sweden
Chalmers University of Technology and Sahlgrenska University Hospital, Gothenburg (inferred from author affiliations; the paper does not name the treatment site)
Sweden
Gothenburg, Sweden (inferred from author affiliations)
null
Rehabilitation medicine / chronic pain
Upper-limb amputees with phantom limb pain
Health
Health
Long-term Care
Prediction from data
Patients; People with disabilities
3
null
physical_health
null
2
1
3
null
5
4
null
personal; societal
5
null
3
the patient's pain after 48 years of conventional treatments (mirror therapy, drugs, acupuncture, hypnosis)
constant burning pain of an intensity of 3 on a scale from 0 to 10 (SF-MGPQ)
gradually reduced to complete pain-free periods
measured
challenged
2024-09-05
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC3935120/
null
null
null
[]
AIBD-0982
2013-04-18
source publication
day
reported
An Indiana federal court found that Biomet's review complied fully with the discovery rules, using predictive coding after keyword culling
In April 2013 a US district judge found that Biomet's review fully met the discovery rules. Biomet culled by keyword before predictive coding of 2.5 million documents. He refused to order a restart.
Court upholds Biomet's *predictive coding* review under discovery rules
2.5 million documents
null
found
Found by a court
no comparison
very low
benefit created
false
Axelerate predictive coding service (the order's spelling of Axcelerate)
null
Recommind, maker of Axcelerate (the order names the service as Axelerate and names no company)
United States
Biomet (defendant), with its e-discovery vendor and eight contract attorneys
United States
South Bend, Indiana, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
1
1
1
null
3
1
null
organizational
1
null
0
null
null
null
null
litigated
2013-08-21
shown
https://storage.courtlistener.com/recap/gov.uscourts.innd.71286/gov.uscourts.innd.71286.396.0.pdf
null
null
null
[]
AIBD-0787
2013-08-21
source publication
day
tested
FDA finds that VITEK MS correctly identified 87.5% of 7068 isolates from patient cultures with a machine-learned classifier of spectra
In a study in five US clinical laboratories that FDA reviewed, VITEK MS gave a correct single identification for 87.5% (6184/7068) of isolates from patient cultures.
*VITEK MS* identified microbes from patient cultures in five labs
87.5% (6184/7068)
null
found
Found by a regulator
no comparison
very low
aggregate result
false
VITEK MS (MALDI-TOF mass spectrometry with the Advanced Spectrum Classifier)
VITEK MS V2.0 Knowledge Base
bioMรฉrieux, Inc.
United States
Five US clinical microbiology laboratories in the prospective clinical study
United States
Five clinical microbiology laboratories, United States
United States
Healthcare
null
Health
Health
Drug Industry
Prediction from data
Health workers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
1
null
0
null
null
null
null
scaled
2017-07-22
shown
https://www.accessdata.fda.gov/cdrh_docs/reviews/K124067.pdf
null
null
null
[]
AIBD-1150
2013-10-01
event
month
mixed
Ofcom finds that live TV subtitles in the UK were generally accurate but arrived late, with respeakers using speech recognition alongside stenographers
In Ofcom's sampling of live subtitles from late 2013, accuracy was generally good. For news it was above 98%. But the median delay of 5.6 seconds exceeded Ofcom's 3-second guideline.
UK live subtitles by *respeaking* are accurate but late, Ofcom finds
the average accuracy rate was 98.5%
null
found
Found by a regulator
tested against a standard
low
aggregate result
false
Live subtitling by respeakers using speech recognition software, alongside stenographers
null
Not named (speech recognition software used by subtitling providers)
null
BBC, ITV, Channel 4, Channel 5 and Sky and their subtitling providers
United Kingdom
United Kingdom
United Kingdom
Broadcasting / live subtitles for deaf and hard-of-hearing viewers
null
Accessibility
Technology
Broadcast
Speech and sound
People with disabilities
10
High income
accessibility
null
2
null
4
4
2
5
null
personal
2
null
0
The 98% accuracy threshold that Ofcom's external reviewers treat as acceptable
98%
generally higher for news than 98%
measured
persisted
2023-12-13
shown
https://web.archive.org/web/20180424104217id_/https://www.ofcom.org.uk/__data/assets/pdf_file/0019/45136/sampling-report.pdf
null
null
null
[]
AIBD-0729
2014-01-01
event
year
reported
Evaluators for Energy Trust of Oregon find that Nest thermostats cut heat-pump homes' electricity use by 781 kWh a year, 12% of heating load
In Energy Trust of Oregon's 2013-14 pilot, homes given a Nest thermostat to run their heat pumps used 781 kWh a year less, 4.7% of their electricity, than comparison homes.
*Nest* thermostats trim electricity use in Oregon heat-pump homes
781 kWh per year
null
claimed
The evaluator's account
controlled comparison
low
aggregate result
false
Nest Learning Thermostat (heat pump control)
null
Nest Labs
United States
Energy Trust of Oregon pilot, installed by CLEAResult in 185 heat-pump-heated homes
United States
Oregon (I-5 corridor), United States
United States
Energy (residential efficiency)
null
Energy and infrastructure
Energy
Conservation
Planning and control
Consumers
7
High income
environmental_impact
Environmental impacts
1
3
3
4
3
null
null
personal; societal
1
null
0
299 comparison homes from the same pool without a Nest
null
781 kWh a year less, 4.7% of total electric usage
measured
withdrawn
2026-01-01
shown
https://www.energytrust.org/wp-content/uploads/2016/12/Nest_Pilot_Study_Evaluation_wSR.pdf
null
null
null
[]
AIBD-0670
2014-05-01
event
month
reported
eBay Machine Translation raises U.S. exports on eBay to Spanish-speaking Latin America (Brynjolfsson, Hui and Liu)
eBay's machine translation, introduced for Latin America in May 2014, raised exports on eBay by 10.9% in the published study. In their working paper, the authors estimated 17.5% to 20.9%.
Machine translation lifts cross-border sales on *eBay*
raised exports on eBay by 10.9% in the published study
null
studied
Independently studied
controlled comparison
low
aggregate result
false
eBay Machine Translation (eMT), an in-house statistical machine-learning translation system
null
eBay
United States
eBay
United States
eBay marketplace: U.S. sellers and buyers in Spanish-speaking Latin American countries
United States
E-commerce and international trade
null
Work and industry
Foreign Trade
Exports
Text and language
Consumers
8
High income
operational_improvement; accessibility
Increased revenue
2
null
4
5
3
5
null
personal; organizational
2
null
0
U.S. exports on eBay to destinations that did not receive the change, before and after eBay replaced Bing Translator with eMT
Exports with Bing Translator before May 2014
10.9% more exports (published); 17.5% to 20.9% more (working paper)
measured
unknown
null
shown
https://www.nber.org/system/files/working_papers/w24917/w24917.pdf
null
null
null
[]
AIBD-1733
2014-06-09
source publication
day
intended
Digital Summit: First Emotion-Reading Apps for Kids with Autism
Affectiva's chief science officer said in 2014 that the first apps using its face-reading software to teach children with autism to read emotions were nearing release. No result is reported.
Affectiva says emotion-reading apps for children with autism are near
null
null
claimed
The makers' claim
null
very low
program
false
Affdex
null
Affectiva
United States
App developers (not named)
null
United States
United States
Assistive technology (emotion recognition for autism)
null
Accessibility
Social Welfare
Disabled Assistance
Images and video
Children; People with disabilities
4; 10
High income
emotional_health
Emotional analysis/intelligence
null
null
null
null
null
null
null
personal
1
null
0
null
null
null
null
unknown
2021-05-25
shown
https://www.technologyreview.com/s/528191/digital-summit-first-emotion-reading-apps-for-kids-with-autism/
null
null
null
[]
AIBD-0983
2014-07-22
source publication
day
intended
A Tennessee federal court let Bridgestone review over two million documents with predictive coding over IBM's objection, citing efficient discovery
In July 2014 a US magistrate judge let Bridgestone review over two million documents with predictive coding, over IBM's objection. He cited Rule 26's call for efficient discovery.
Court lets Bridgestone use *predictive coding* over IBM's objection
over two million documents
null
found
Found by a court
null
very low
benefit created
false
Predictive coding
null
Not named in the order
null
Bridgestone Americas, Inc. (plaintiff), with its attorneys, experts and the companies doing the work
United States
Nashville, Tennessee, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
null
null
null
null
challenged
2015-02-05
shown
https://storage.courtlistener.com/recap/gov.uscourts.tnmd.57186/gov.uscourts.tnmd.57186.89.0.pdf
null
null
null
[]
AIBD-0812
2014-08-11
source publication
day
tested
FDA found that Cologuard detected 92.3% of colorectal cancers against 73.8% for a fecal immunochemical test, at lower specificity
In FDA's review of Cologuard, the stool DNA test detected 92.3% of colorectal cancers against 73.8% for FIT in 9,989 adults. Its specificity was 86.6% against 94.9%.
*Cologuard* caught more colorectal cancers than a standard stool test
detected 92.3% of colorectal cancers against 73.8% for FIT
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
Cologuard
null
Exact Sciences Corporation
United States
Exact Sciences (DeeP-C study at 90 sites in the United States and Canada)
United States
90 sites in the United States and Canada
United States | Canada
Healthcare
null
Health
Health
Disease Prevention
Prediction from data
Patients
3
High income
physical_health
Medical diagnosis
null
null
null
2
null
null
null
organizational
2
null
0
A commercial fecal immunochemical test (OC FIT-CHEK) in the same adults
Sensitivity 73.8% for colorectal cancer, specificity 94.9%
Sensitivity 92.3%, specificity 86.6%
measured
scaled
2024-10-04
not shown
https://www.accessdata.fda.gov/cdrh_docs/pdf13/P130017B.pdf
null
null
null
[]
AIBD-0190
2014-09-01
event
month
tested
In a double-blind retrodiction of 2014 inspections, Chicago's model for forecasting food inspections would have found critical violations 7.5 days earlier
In a double-blind retrodiction of 1,637 Chicago food inspections from 2014, the city's Food Inspection Forecasting model would have found critical violations on average 7.5 days earlier.
Chicago's model would have found food violations 7.5 days earlier
would have found 178 (69%) in the first half of the work, on average 7.5 days earlier
null
studied
Independently studied
tested against a standard
very low
aggregate result
false
Food Inspection Forecasting (predictive analytics model, open-sourced in R)
null
City of Chicago Department of Innovation and Technology with the Department of Public Health, Allstate Insurance data scientists and the Civic Consulting Alliance
United States
Chicago Department of Public Health
United States
Chicago, United States
United States
Food safety / public health
Restaurant inspection targeting
Agriculture and food
Agriculture
Food Inspection & Safety
Prediction from data
Public servants
3
High income
operational_improvement; physical_health
Faster data management, decision-making
null
null
null
null
null
null
null
organizational; societal
4
null
0
the Department of Public Health's normal inspection order during the September-October 2014 pilot
55 percent (141 establishments) of those with critical violations found in the first month during normal operations
69 percent (178 establishments) found in the first half of work; violations found on average 7 and a half days earlier
measured
persisted
2016-11-28
shown
https://arxiv.org/pdf/1910.04906
null
null
null
[]
AIBD-1740
2014-09-12
source publication
day
reported
Digital banking: Credit scoring goes social
Euromoney reported in 2014 that Lenddo, which profiles borrowers from their social network accounts, had made over 100,000 credit decisions and supplied banks with borrowers in three countries.
Lenddo scores borrowers from their social networks in three countries
null
null
claimed
The makers' claim
no comparison
very low
aggregate result
false
Lenddo social data credit scoring
null
Lenddo
null
Lenddo and partner banks
null
Mexico, the Philippines and Colombia
Mexico | Philippines | Colombia
Consumer lending (social data scoring)
null
Finance
Domestic Commerce
Consumer Finance
Prediction from data
Consumers
8
Upper middle income
economic_access
Credit access
1
null
null
null
3
null
null
personal
1
null
0
null
null
null
null
unknown
2017-10-01
not shown
null
null
null
null
[]
AIBD-0833
2014-09-17
source publication
day
intended
The US Tax Court let two taxpayers answer a discovery request from the IRS with predictive coding, crediting an estimate of $80,000 to $85,000 for review
In September 2014 the US Tax Court allowed predictive coding over the IRS's objection. It credited an estimate of 200,000 to 400,000 documents to review at $80,000 to $85,000.
Tax Court lets taxpayers use *predictive coding* though the IRS objects
200,000 to 400,000 documents to review at $80,000 to $85,000
null
found
Found by a court
null
very low
benefit created
false
Predictive coding
null
Not named in the opinion
null
Dynamo Holdings Limited Partnership and Beekman Vista, Inc. (petitioners)
United States
United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
null
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
Review under the Commissioner's approach, a privilege search of all restored data
3.5 million to 7 million documents at a cost of $500,000 to $550,000
200,000 to 400,000 documents at a cost of $80,000 to $85,000
assumed
litigated
2016-07-13
shown
https://static.case.law/tc/143/html/0183-01.html
null
null
null
[]
AIBD-0786
2014-11-17
source publication
day
tested
FDA finds that the BrainScope Ahead 100 missed its endpoints for accuracy but ruled out brain injury on CT with an 89.5% negative predictive value
In FDA's review of the BrainScope Ahead 100, the EEG test missed its endpoints for accuracy in 552 patients with head injury. Its negative predictive value for injury on CT was 89.5%.
*Ahead 100* EEG test missed its goals but helped rule out injury on CT
negative predictive value for injury on CT was 89.5%
null
found
Found by a regulator
tested against a standard
very low
aggregate result
false
BrainScope Ahead 100 (Models CV-100 and M-100)
null
BrainScope Company, Inc.
United States
BrainScope (B-AHEAD II study at 11 US emergency department sites)
United States
Eleven emergency department sites, United States
United States
Healthcare
null
Health
Health
Disease Prevention
Prediction from data
Patients; Health workers
3
High income
physical_health
Medical diagnosis
null
null
null
null
null
null
null
organizational
2
null
0
Pre-specified endpoints for the lower confidence limits of sensitivity and specificity
Lower limits above 78% for sensitivity and 50% for specificity
Sensitivity 78.5% and specificity 48.6%, with the endpoints missed
measured
persisted
2015-05-15
shown
https://www.brainscope.com/sciencetechnology
null
null
null
[]
AIBD-1037
2014-12-31
event
day
mixed
The FDIC Inspector General finds that the SCOR, SCOR-Lag and GMS models flagged 67 of 244 well-rated banks downgraded from 2014 to 2017
The FDIC's Inspector General found that SCOR, SCOR-Lag and GMS had flagged 67 of the 244 banks rated 1 or 2 that were downgraded from late 2014 to late 2017.
Three FDIC models including *SCOR* flag 67 of 244 banks later downgraded
67 of the 244 banks
null
found
Found by an auditor
no comparison
very low
aggregate result
false
Statistical CAMELS Offsite Rating (SCOR), SCOR-Lag and Growth Monitoring System (GMS) offsite monitoring models
null
Federal Deposit Insurance Corporation
United States
Federal Deposit Insurance Corporation, Division of Risk Management Supervision (Case Managers in six Regional Offices)
United States
United States (national)
United States
Bank supervision
null
Finance
Domestic Commerce
Banking
Prediction from data
null
16
High income
operational_improvement
Faster data management, decision-making
2
1
4
null
2
2
null
organizational
1
null
0
null
null
null
null
unknown
null
not shown
https://www.oversight.gov/sites/default/files/documents/reports/2023-06/20-002EVAL.pdf
null
null
null
[]
AIBD-0174
2015-01-01
event
year
reported
Bank F's tool for network analytics links a third company to a suspected network for trade-based money laundering in Hong Kong
Bank F's unnamed tool for network analytics linked a third company to a suspected money laundering network through shared phone numbers.
Bank's network tool links a third company to suspected money laundering
linked a third company to a suspected money laundering network
null
reported
Independently reported
no comparison
very low
protective response
false
Bank F network (graph) analytics solution implemented in Hong Kong in 2013
null
Bank F (anonymized subsidiary and regional head office of a large global bank); the report names no vendor
null
Bank F, Hong Kong
null
Hong Kong, China
Hong Kong | China
Anti-money laundering
Trade finance and trade-based money laundering
Finance
Law and Crime
White Collar Crime
Prediction from data
null
16
Upper middle income; High income
privacy_protection
Fraud detection
2
1
4
null
5
null
null
organizational; societal
2
null
3
Bank F's earlier manual review of documentation for signatories, before the network analytics tool
manual approaches to reviewing documentation for signatories, which the report calls the only means to make these connections
the tool linked Company C to Company A through shared phone numbers and email addresses
assumed
scaled
2021-01-21
not shown
https://www.hkma.gov.hk/media/eng/doc/key-information/guidelines-and-circular/2021/20210121e2a1.pdf
null
null
null
[]
AIBD-0176
2015-01-01
event
year
reported
Some subprime borrowers get lower-priced credit at LendingClub
LendingClub's rating grades relied increasingly on alternative data. They moved some borrowers who are subprime by traditional criteria into better grades with lower-priced credit.
Some subprime borrowers get lower-priced credit at LendingClub
about 8 percent of A-rated and 28 percent of B-rated borrowers had FICO scores below 680
null
studied
Independently studied
tested against a standard
low
aggregate result
false
LendingClub rating-grade model (grades A to G)
null
LendingClub
United States
LendingClub
United States
United States
United States
Consumer credit
Marketplace consumer lending
Finance
Domestic Commerce
Consumer Finance
Prediction from data
Consumers
8; 10
High income
economic_access
Fair pricing
1
null
4
4
3
3
null
personal; societal
3
null
1
classification of borrowers by traditional credit criteria
classified as subprime by traditional criteria
slotted into โ€œbetterโ€ loan grades, which allowed them to get lower-priced credit
measured
litigated
2021-07-14
shown
https://web.archive.org/web/2026id_/https://www.sec.gov/Archives/edgar/data/1409970/000140997020000212/a201910-k.htm
null
null
null
[]
AIBD-0186
2015-01-01
event
year
reported
Medicare's Fraud Prevention System helps catch $654.8 million in inappropriate payments in 2015
The Centers for Medicare and Medicaid Services count $654.8 million in inappropriate Medicare payments identified or prevented in 2015 with help from their Fraud Prevention System.
Screening Medicare claims helps catch $654.8 million in 2015
$654.8 million in inappropriate Medicare payments identified or prevented in 2015
null
found
Found by an auditor
no comparison
very low
aggregate result
false
Fraud Prevention System (FPS)
null
Contractors engaged by CMS (the report lists 'Developing Contractor Costs' and 'Modeling Contractor Costs' without naming firms)
null
Centers for Medicare and Medicaid Services (CMS)
United States
United States (national)
United States
Health insurance fraud and improper payments
Medicare program integrity
Finance
Health
Medical Liability
Prediction from data
null
16
High income
privacy_protection
Fraud detection
1
1
4
null
5
5
null
organizational; societal
4
null
3
the FPS's identified savings in the previous year, 2014
identified savings from the previous year ($454.0 million in Table 1)
$654.8 million
measured
persisted
2017-08-30
not shown
https://web.archive.org/web/20260831112755/https://www.gao.gov/products/gao-17-710
$654.8 million
inappropriate Medicare payments identified or prevented in calendar year 2015
false
[]
AIBD-0196
2015-01-01
event
year
reported
Global Fishing Watch data lets Kiribati fine a vessel US$1 million for fishing in the Phoenix Islands Protected Area
Kiribati used Global Fishing Watch to document a vessel fishing illegally in the Phoenix Islands Protected Area. The owner faced a US$1 million fine plus a US$1 million goodwill grant.
Global Fishing Watch helps Kiribati fine a vessel's owner US$1 million
a US$1 million fine plus a US$1 million goodwill grant
null
reported
Independently reported
no comparison
very low
benefit created
false
Global Fishing Watch (AIS-based fishing detection using machine learning)
null
Oceana, SkyTruth and Google
United States
Government of Kiribati
Kiribati
Phoenix Islands Protected Area, Kiribati
Kiribati
Fisheries enforcement
Marine protected area enforcement
Agriculture and food
Agriculture
Fisheries & Fishing
Prediction from data
null
14
Lower middle income
environmental_impact
Environmental impacts
2
1
1
null
3
1
null
organizational; societal
6
null
1
null
null
null
null
scaled
2017-06-07
shown
https://globalfishingwatch.org/transparency/indonesia-shares-vms-with-global-fishing-watch/
$1 million
fine secured by Kiribati against the owner of Marshalls 203 for fishing illegally in the Phoenix Islands Protected Area, plus an additional $1 million paid as a grant under a goodwill arrangement
false
[]
AIBD-0635
2015-01-01
event
year
reported
TIGTA finds that the IRS Return Review Program's identity-theft models alone selected $1.88 billion of confirmed identity-theft refunds in the 2015 filing season
TIGTA found that IRS Return Review Program identity-theft models alone selected $1.88 billion, or 47.8%, of $3.92 billion in confirmed identity-theft refunds in 2015.
*Return Review Program* catches refund fraud that older filters missed
selected $1.88 billion, or 47.8%, of $3.92 billion in confirmed identity-theft refunds
null
found
Found by an auditor
tested against a standard
low
aggregate result
false
IRS Return Review Program (RRP) identity-theft models
null
Internal Revenue Service (Information Technology organization and Office of Research, Applied Analytics, and Statistics)
United States
Internal Revenue Service
United States
United States (national)
United States
Tax administration
null
Public sector
Government Operations
Tax Administration
Prediction from data
null
16
High income
privacy_protection
Fraud detection
2
4
4
null
4
5
null
organizational; societal
6
null
1
The older Electronic Fraud Detection System identity-theft models
$60 million (1.5%) of confirmed identity-theft refunds selected only by the older system
$1.88 billion (47.8%) selected only by the Return Review Program models
measured
persisted
2018-07-24
shown
https://www.tigta.gov/sites/default/files/reports/2022-02/201720080fr.pdf
null
null
null
[]
AIBD-0731
2015-01-01
event
year
reported
Evaluators for Energy Trust of Oregon find that self-installed Nest thermostats cut gas heating use about 6%, while Honeywell's Lyric raised it
In Energy Trust of Oregon's 2014-15 pilot, gas-heated homes that self-installed a Nest saved about 34 therms a year, 6% of heating. Homes given Honeywell's Lyric used 4-5% more.
*Nest* thermostats cut gas heating use in Oregon homes by about 6%
Nest homes saved about 34 therms a year
null
claimed
The evaluator's account
controlled comparison
low
aggregate result
false
Nest Learning Thermostat (second generation)
null
Nest Labs
United States
Energy Trust of Oregon pilot with NW Natural customers, run by CLEAResult (self-installed thermostats in homes with gas furnaces)
United States
Oregon, United States
United States
Energy (residential efficiency)
null
Energy and infrastructure
Energy
Conservation
Planning and control
Consumers
7
High income
environmental_impact
Environmental impacts
1
3
3
1
3
null
null
personal; societal
1
null
0
About 1,000 randomly selected NW Natural customers who were not contacted
null
About 34 therms a year saved, 6.0% of heating usage
measured
scaled
2016-04-06
shown
https://www.energytrust.org/documents/gas-smart-thermostat-pilot-evaluation-report/
null
null
null
[]
AIBD-1737
2015-01-01
event
year
tested
Artificial Neural Networks Applied to Taxi Destination Prediction
Researchers report that their neural network ranked first of 381 teams in a 2015 contest to predict a taxi's destination from the start of its route. No taxi service is reported to use it.
Neural network ranks first of 381 teams at predicting taxi destinations
null
null
claimed
The makers' claim
no comparison
very low
program
false
Taxi destination prediction network
null
The contest's winning team (de Brรฉbisson and others)
null
No user is reported
null
Not stated in the account
null
Transport (taxi destination prediction)
null
Safety and rescue
Transportation
General
Prediction from data
null
null
null
knowledge_discovery
null
null
null
null
null
null
null
null
organizational
1
null
0
null
null
null
null
unknown
null
shown
https://arxiv.org/abs/1508.00021
null
null
null
[]
AIBD-0981
2015-03-02
source publication
day
intended
A New York federal court approved Rio Tinto and Vale's agreed TAR protocol, stressing that TAR must face no higher standard than manual review
In March 2015 a US magistrate judge approved Rio Tinto and Vale's protocol for predictive coding. He wrote that TAR must face no higher standard than manual review. He wrote the same of keywords.
Federal court approves a *TAR* protocol agreed by Rio Tinto and Vale
null
null
found
Found by a court
null
very low
benefit created
false
Relativity Assisted Review (Rio Tinto) and Deloitte's Dynamic Review (Vale), under a stipulated predictive coding protocol
null
Relativity (Relativity Assisted Review) and Deloitte (Dynamic Review), as the parties' letter attached to the opinion names the tools
United States (Relativity); not verified for Deloitte
Rio Tinto plc (plaintiff), with its vendor Precision Discovery, and Vale S.A. (defendant)
United Kingdom and Brazil
New York, United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
2
null
0
Keyword searches or manual review
null
null
assumed
persisted
2015-07-15
shown
https://static.case.law/frd/306/html/0125-01.html
null
null
null
[]
AIBD-0829
2015-03-03
source publication
day
intended
The High Court of Ireland approved technology assisted review for 680,809 documents, finding predictive coding at least as accurate as manual review
In March 2015 the High Court of Ireland approved predictive coding for 680,809 documents over the defendants' objection. It found that the method was at least as accurate as manual review.
Irish court backs *predictive coding* for 680,809 bank documents
less than 10% of the 680,809 documents
null
found
Found by a court
null
very low
benefit created
false
Technology assisted review using predictive coding
null
Not named in the judgment (the bank's expert took his target f-measure from Symantec, the provider of the Clearwell system)
null
Irish Bank Resolution Corporation and its solicitors, in discovery
Ireland
Dublin, Ireland
Ireland
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
2
null
0
Manual (linear) review of the 680,809 documents
9 months at a cost of โ‚ฌ2 m by the bank's solicitor's estimate, or 113 days at โ‚ฌ220,000 by the defendants' expert's
less than 10% of the 680,809 documents would need to be manually reviewed
assumed
persisted
2016-03-01
shown
https://web.archive.org/web/20240428025752if_/https://www.bailii.org/ie/cases/IEHC/2015/H175.html
null
null
null
[]
AIBD-0188
2015-03-31
event
day
reported
nEmesis ranks food venues in Las Vegas from tweets to guide inspections that find 11 venues graded no better than C against 7 under the standard protocol
In a blinded Las Vegas trial from 2 January 2015, inspectors sent to venues that nEmesis flagged from tweets found 9 demerits per visit versus 6. They found 11 venues graded no better than C versus 7.
Inspections guided by nEmesis find more venues graded no better than C
found 11 venues graded C or worse versus 7 for controls
null
studied
Independently studied
controlled comparison
low
aggregate result
false
nEmesis (SVM language model over geotagged tweets with a web interface for inspectors)
null
University of Rochester (Adam Sadilek, Henry Kautz and colleagues)
United States
Southern Nevada Health District
United States
Las Vegas, United States
United States
Food safety / public health
Restaurant inspection targeting
Agriculture and food
Agriculture
Food Inspection & Safety
Text and language; Prediction from data
Public servants
3
High income
operational_improvement; physical_health
Enhanced productivity
2
1
null
null
5
null
null
organizational; societal
2
null
3
Venues inspected under the standard SNHD protocol in the same three months, with inspectors blind to which venues nEmesis chose
6 demerits per inspection and 7 venues graded C or worse in the control arm
9 demerits per inspection and 11 venues graded C or worse in the nEmesis arm
measured
unknown
null
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC11764790/
null
null
null
[]
AIBD-0057
2015-07-01
event
day
reported
OrCam computer-vision readers improve scores on daily tasks in clinical studies from 2015 to 2026, though most patients in a Dutch study stop using them
With OrCam readers, people in studies of 6 to 100 participants scored higher on daily tasks such as reading. In a Dutch study, only three of 20 patients kept using the device.
People with low vision complete more daily tasks with OrCam readers
improved on a 10-item daily-living test from 2.5 to 9.5
null
studied
Independently studied
controlled comparison
low
aggregate result
false
OrCam MyEye, MyEye 2 and OrCam Read (camera-based optical character and object recognition with audio output)
OrCam (2015 study; model not named in the abstract); OrCam MyEye (2023 multicentre study); OrCam MyEye 2 (NIH); OrCam Read (UC Davis)
OrCam Technologies
Israel
Study clinics: an ophthalmology department in the United States (2015); Bartimรฉus and Royal Dutch Visio, two low-vision rehabilitation centers in Amsterdam; five vision rehabilitation centers; UC Davis Ophthalmology; NIH Clinical Center
United States
null
null
Low-vision rehabilitation
People with low vision, acquired brain injury or CLN3 disease
Accessibility
Health
Long-term Care
Images and video; Speech and sound
People with disabilities; Patients
10
null
accessibility
null
2
3
2
4
4
null
null
personal; societal
6
null
1
the same patients' best-corrected visual acuity without the device
a mean of 2.5 of 10 daily-living tasks completed
a mean of 9.5 of 10 tasks with the device at the first visit
measured
persisted
2026-04-03
not shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC10863346/
null
null
null
[]
AIBD-2292
2015-10-01
event
month
reported
Using Twitter to Identify and Respond to Food Poisoning: The Food Safety STL Project
In seven months at the City of St Louis Department of Health, a dashboard that classifies tweets about food poisoning captured 193 relevant tweets, and replies to them drew new reports.
*HealthMap Foodborne Dashboard* brings St Louis more illness reports
193 relevant tweets
null
studied
Independently studied
no comparison
very low
protective response
false
HealthMap Foodborne Dashboard (Food Safety STL)
null
HealthMap researchers at Boston Children's Hospital
United States
City of St Louis Department of Health
United States
St Louis, Missouri, United States
United States
Public health (surveillance of foodborne illness)
null
Health
Health
Disease Prevention
Text and language
The public; Public servants
3
High income
physical_health; operational_improvement
Hazard reduction
1
null
null
null
3
null
null
organizational
2
null
0
null
null
null
null
unknown
null
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC5540821/
null
null
null
[]
AIBD-0882
2015-12-01
source publication
month
reported
Evaluators at Oak Ridge National Laboratory find that socially driven HVAC control raised comfort for 83% of occupants in GSA's Phoenix courthouse
At the federal courthouse in Phoenix, 83% of occupants reported more thermal comfort with socially driven HVAC control. Calls to maintenance about temperature fell 59%.
Occupants vote to adjust *socially driven HVAC* to their comfort
83% of occupants reported more thermal comfort
null
claimed
The evaluator's account
before and after
very low
aggregate result
false
Socially driven HVAC optimization software
null
Building Robotics, LLC (named in the evaluation as the technology provider)
United States
U.S. General Services Administration (Federal Building and U.S. Courthouse, Phoenix)
United States
Phoenix, Arizona, United States
United States
Building energy management
null
Energy and infrastructure
Energy
Conservation
Planning and control
Public servants
7
High income
operational_improvement
Customer personalization
1
3
4
2
3
4
0.35
personal; organizational; societal
2
null
0
The building's previous fixed-range temperature control
null
83% of occupants reported being more satisfied with their thermal environment
measured
unknown
null
shown
https://www.gsa.gov/system/files/GPG025-Socially%20Driven%20HVAC-Findings-508R.pdf
null
null
null
[]
AIBD-0105
2016-01-01
event
year
reported
Subscriptions to GLAD deforestation alerts lower the probability of deforestation in Africa by 18%
Four researchers found that subscriptions to GLAD Landsat alerts cut the probability of deforestation in Africa by 18%. Two of them come from the organizations behind the alerts.
Subscribing to GLAD alerts lowers deforestation risk in Africa by 18%
decreased the probability of deforestation in Africa by 18 percent
null
studied
Independently studied
controlled comparison
low
aggregate result
false
GLAD (Global Land Analysis and Discovery) Landsat deforestation alerts on Global Forest Watch
null
University of Maryland GLAD laboratory (Hansen et al.)
United States
World Resources Institute (Global Forest Watch) and alert subscribers
United States
null
null
Conservation and biodiversity
Forest monitoring and enforcement
Environment
Environment
Species & Forest
Images and video
null
15; 13
null
environmental_impact
Environmental impacts
2
null
4
2
4
4
null
personal; organizational; societal
6
null
2
average 2011โ€“2016 deforestation levels, with areas lacking alert subscribers as the counterfactual across 22 tropical countries
average 2011โ€“2016 levels
probability of deforestation in Africa decreased by 18%; no effect on other continents
measured
persisted
2026-01-13
shown
https://iopscience.iop.org/article/10.1088/1748-9326/11/3/034008
between US$149 million and US$696 million
value of avoided deforestation in Africa, priced at the social cost of carbon
true
[]
AIBD-0166
2016-01-01
event
year
reported
ActiveRemediation helps Flint locate hazardous water pipes
Researchers at the University of Michigan built ActiveRemediation, a model that ranked Flint homes by the chance of a hazardous service line, such as one of lead. Its held-out accuracy was 91.6%.
Flint uses a prediction model to find homes with hazardous water pipes
held-out prediction accuracy was 91.6%
null
reported
Independently reported
no comparison
very low
benefit created
false
ActiveRemediation, combining XGBoost predictions, a hierarchical Bayesian spatial model, and active learning
null
University of Michigan (Jacob Abernethy, Alex Chojnacki, Arya Farahi, Eric Schwartz, Jared Webb)
United States
City of Flint (FAST Start service-line replacement program)
United States
Flint, United States
United States
Water infrastructure
Lead service-line replacement
Energy and infrastructure
Environment
Drinking Water
Prediction from data
The public
6
High income
physical_health; operational_improvement
Hazard reduction
2
3
5
4
3
1
0.28
personal; organizational; societal
8
null
0
The recorded sequence of inspection and replacement work in a simulation
Estimated US$5,818 per replacement
Estimated US$5,196 per replacement
assumed
persisted
2021-01-14
shown
https://arxiv.org/abs/1806.10692
null
null
null
[]
AIBD-0820
2016-01-01
event
month
mixed
New York City's Taxi and Limousine Commission finds that tailgating alerts from Mobileye's system fell in fitted cabs during its safety trial
In a TLC trial, tailgating alerts from Mobileye's system fell from 100 per 100 miles in July 2015 to between 20 and 40 from January 2016. Crash rates in the trial did not clearly fall.
Tailgating alerts from *Mobileye* fall in New York taxis
fell from 100 per 100 miles in July 2015 to between 20 and 40
null
found
Found by a regulator
before and after
very low
aggregate result
false
Mobileye aftermarket warning system, with a black box recorder added for the trial
null
Mobileye
Israel
A fleet of mostly yellow taxis licensed by the TLC, in the TLC's Vehicle Safety Technology Pilot
United States
New York City, New York, United States
United States
Road safety / for-hire vehicles
null
Safety and rescue
Transportation
Highways
Images and video
Drivers and travelers
11
High income
physical_health
Hazard reduction
2
2
4
null
2
5
null
personal; societal
1
null
0
Alert rates in the same fleet early in the trial
100 alerts for each 100 miles traveled (or one alert per mile) in July 2015
between 20 and 40 alerts per 100 miles starting in January 2016
measured
unknown
null
shown
https://ims.mobileye.com/fleets/uk/blog/what-does-ai-have-to-do-with-collision-avoidance-systems/
null
null
null
[]
AIBD-0834
2016-01-01
event
month
mixed
The US Tax Court held that predictive coding of about 406,000 documents gave the IRS a complete response, at a precision of 3%
In July 2016 the US Tax Court held that a response to the IRS by predictive coding was complete. The model sorted about 406,000 documents into a production set of about 180,000, at 3% precision.
Tax Court finds a response to the IRS by *predictive coding* complete
about 406,000 documents into a production set of about 180,000
null
found
Found by a court
no comparison
very low
benefit created
false
Predictive coding algorithm
null
Not named in the order
null
Dynamo Holdings Limited Partnership and Beekman Vista, Inc. (petitioners)
United States
United States
United States
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
null
16
High income
operational_improvement
Enhanced productivity
1
2
1
null
2
1
null
organizational
1
null
0
null
null
null
null
unknown
null
shown
https://www.ediscoverylaw.com/wp-content/uploads/2016/08/Dynamo-II-Order.pdf
null
null
null
[]
AIBD-2058
2016-01-01
event
year
reported
SophieBot
UNFPA says that SophieBot, an AI chatbot for young people in Kenya, has been asked over 38,000 questions on sexual and reproductive health by over 3,000 users since 2016.
Kenya's *SophieBot* answers 38,000 questions on sexual health
38,000
null
claimed
The deployer's account
no comparison
very low
aggregate result
false
SophieBot
null
Sophie Bot (Kenyan start-up)
Kenya
Sophie Bot, with UNFPA Kenya
Kenya
Kenya
Kenya
Healthcare (information on sexual and reproductive health)
null
Health
Health
Disease Prevention
Text and language
The public
3; 5
Lower middle income
physical_health; knowledge_discovery
null
1
3
null
null
3
null
null
personal
1
ITU, United Nations Activities on Artificial Intelligence
0
null
null
null
null
unknown
null
not shown
null
null
null
null
[]
AIBD-2157
2016-01-01
event
year
reported
LAURA
LAURA reads charts in 13 Brazilian hospitals and alerts teams to sepsis. Its medical director reported 25% fewer deaths in a before-and-after study, and one hospital a fall from 4.33% to 1.64%.
Brazilian hospitals report fewer deaths after *LAURA*'s sepsis alerts
25% fewer deaths
null
claimed
The deployer's account
before and after
very low
aggregate result
false
LAURA (alerts of sepsis and clinical deterioration from patient charts)
null
Laura (Brazil)
Brazil
Thirteen hospitals in three Brazilian states, by the count of October 2019
Brazil
Paranรก, Minas Gerais and Rio Grande do Sul, Brazil
Brazil
Health care (early warning of sepsis and clinical deterioration)
null
Health
Health
Disease Prevention
Prediction from data
Patients; Health workers
3
Upper middle income
physical_health
null
5
4
5
null
3
null
null
personal
4
Inter-American Development Bank, fAIr LAC
1
The six months before the hospitals began to use the platform
null
mortality 25% lower and stays 10% shorter
measured
unknown
null
shown
https://fairlac.iadb.org/piloto/laura
null
null
null
[]
AIBD-0830
2016-02-16
source publication
day
intended
The High Court in London approved predictive coding software for over 3 million documents, finding no evidence that it gives less accurate disclosure
In February 2016 Master Matthews approved predictive coding for over 3 million documents. He found no evidence that it gives less accurate disclosure. He judged it far cheaper than manual review.
English court approves *predictive coding* for 3 million documents
over 3 million documents
null
found
Found by a court
null
very low
benefit created
false
Predictive coding software
null
Not named in the judgment
null
MWB Business Exchange Ltd (second claimant), with the agreement of all parties
United Kingdom
London, United Kingdom
United Kingdom
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
2
null
0
A full manual review of over 3 million documents
several million pounds at least
ยฃ181,988 plus monthly hosting costs of ยฃ15,717, to ยฃ469,049 plus monthly hosting costs of ยฃ20,820
assumed
unknown
null
shown
https://caselaw.nationalarchives.gov.uk/ewhc/ch/2016/256
null
null
null
[]
AIBD-2301
2016-03-29
source publication
day
tested
Supplementing Public Health Inspection via Social Media
A model built on 71,360 Yelp reviews predicted health code violations in 78% of San Francisco restaurants with serious citations in a pilot of 440 restaurants. No health department used it.
*Yelp review model* flags San Francisco restaurants with violations
78% of the restaurants with serious citations
null
studied
Independently studied
tested against a standard
very low
program
false
Model of restaurant health code violations from Yelp reviews (University of California, Irvine)
null
University of California, Irvine
United States
Not deployed (tested on inspection data of San Francisco and New York City)
United States
San Francisco, California, United States
United States
Public health (restaurant inspection)
null
Health
Agriculture
Food Inspection & Safety
Prediction from data; Text and language
null
3
High income
physical_health; operational_improvement
Hazard reduction
null
null
null
null
null
null
null
organizational
2
null
0
null
null
null
null
unknown
null
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC4811425/
null
null
null
[]
AIBD-0049
2016-05-01
event
month
reported
Exomiser's ranking of variants contributed to diagnoses of rare disease in the pilot of the 100,000 Genomes Project
In a pilot of England's National Health Service, Exomiser ranked candidate variants for 2,183 families. A quarter of probands got a genetic diagnosis. A quarter of those were actionable at once.
Exomiser ranks gene variants to help diagnose rare disease in England
Genetic diagnoses were found in 25% of probands, rising to 35% for monogenic disorders
The 100,000 Genomes Project ran a preliminary national study in rare disease. From May 2016 through April 2019, the study returned results to the Genomic Medicine Centers of the English National Health Service. Across 2,183 families, genetic diagnoses were found in 25% of probands. Clinicians reported that a quarter of...
studied
Independently studied
tested against a standard
low
aggregate result
false
Exomiser phenotype-driven variant prioritization (machine-learning-assisted scoring and ranking)
null
Queen Mary University of London (Exomiser team led by Damian Smedley)
United Kingdom
Genomics England and NHS Genomic Medicine Centers
United Kingdom
England, United Kingdom
United Kingdom
Clinical genomics / rare disease diagnosis
Families with undiagnosed rare disease in a national health system
Health
Health
Disease Prevention
Prediction from data
Patients; Children
3
High income
physical_health
Medical diagnosis
4
3
5
null
5
1
null
personal; organizational; societal
4
null
2
the virtual panel-based pipeline (applied gene panels) alone
A total of 322 (66%) of the 490 diagnoses that were based on SNVs or indels from the genomes were made with the virtual panelโ€“based pipeline
diagnoses were detected in 77% of the top-ranked candidate variants, in 86% of the top three candidates, and in 88% of the top five candidates
measured
scaled
2024-12-18
not shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC5467691/
null
null
null
[]
AIBD-0831
2016-05-17
source publication
day
intended
The High Court in London ordered predictive coding over a petitioner's objection, finding that it would cost about ยฃ132,000 against at least ยฃ250,000
In May 2016 Mr Registrar Jones ordered disclosure by predictive coding over the petitioner's objection. He found that it would cost about ยฃ132,000, against at least ยฃ250,000 for a keyword search.
Court orders *predictive coding* at ยฃ132,000 over a keyword search
about ยฃ132,000, against at least ยฃ250,000
null
found
Found by a court
null
very low
benefit created
false
Predictive coding
null
Not named in the judgment
null
BCA Trading Ltd and the other respondents
United Kingdom
London, United Kingdom
United Kingdom
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
A keyword search approach to disclosure
at least ยฃ250,000 (and indeed I believe that figure may even reach ยฃ338,000 on a worst case scenario)
in the region of ยฃ132,000
assumed
persisted
2018-08-28
shown
https://web.archive.org/web/20251018042102if_/https://www.austlii.edu.au/cgi-bin/viewdoc/au/cases/vic/VSC/2016/734.html
null
null
null
[]
AIBD-0058
2016-07-01
event
month
reported
Stanford's Superpower Glass gives children with autism cues about emotions on Google Glass and raises a socialization score that fades after treatment
In a trial of 71 children, Stanford's Superpower Glass, which classifies emotions on Google Glass, added 4.58 points on a socialization scale. Spectrum reported that the gain was gone six weeks later.
Superpower Glass briefly raises social scores of children with autism
gained 4.58 points more on the Vineland socialization subscale than controls (P = .005)
null
studied
Independently studied
randomized trial
moderate
aggregate result
false
Superpower Glass: Google Glass with a real-time machine-learning facial emotion classifier and companion smartphone app
null
Stanford University (Wall lab, Division of Systems Medicine and Department of Computer Science)
United States
Participating families at home under the Stanford studies
null
Palo Alto, United States
United States
Autism intervention / digital therapeutics
Children with autism spectrum disorder aged 6-12
Health
Health
Mental Health
Images and video
Children; People with disabilities
3
High income
emotional_health; accessibility
Emotional analysis/intelligence
2
2
2
4
3
4
0.35
personal; societal
4
null
0
applied behavioral analysis therapy alone (treatment as usual)
null
mean [SD] treatment impact, 4.58 [1.62]; P = .005
measured
challenged
2019-05-02
shown
https://europepmc.org/article/PMC/PMC6550272
null
null
null
[]
AIBD-0215
2016-08-01
event
month
reported
Pounce helps incoming students complete college enrollment
AdmitHubโ€™s text-message assistant Pounce raised enrollment by 3.3 percentage points among students already committed to Georgia State in 2016.
A texting assistant helps more admitted students enroll at Georgia State
raised enrollment by 3.3 percentage points
null
reported
Independently reported
controlled comparison
low
aggregate result
false
Pounce, AdmitHubโ€™s SMS assistant with supervised learning and a Georgia State knowledge base
null
AdmitHub
United States
Georgia State University
United States
Atlanta, Georgia, United States
United States
Higher education / admissions and enrollment
College transition support for admitted students
Education
Education
Higher
Text and language
Students
4
High income
opportunity_access
Educational access
2
null
1
4
4
null
null
personal
3
null
2
Usual enrollment support among students already committed to Georgia State
84.6% enrolled in the control subgroup
Enrollment increased by 3.3 percentage points
measured
scaled
2023-09-01
shown
https://files.eric.ed.gov/fulltext/EJ1194134.pdf
null
null
null
[]
AIBD-2230
2016-08-01
event
month
reported
Automated Identification of Adults at Risk for In-Hospital Clinical Deterioration
At Kaiser Permanente hospitals in Northern California, patients who set off the Advance Alert Monitor died less often within 30 days where the program ran than where it did not yet run.
Kaiser's *Advance Alert Monitor* is tied to fewer deaths after alerts
an adjusted relative risk of 0.84
null
studied
Independently studied
controlled comparison
low
aggregate result
false
Advance Alert Monitor
null
Kaiser Permanente Northern California
United States
Kaiser Permanente Northern California (19 hospitals)
United States
Northern California, United States
United States
Inpatient care (early warning of deterioration from the electronic health record)
null
Health
Health
Disease Prevention
Prediction from data
Patients; Health workers
3
High income
physical_health; operational_improvement
Medical diagnosis
5
4
5
null
3
5
null
personal
4
null
0
Patients who reached the alert threshold at hospitals where the program did not yet run
the comparison cohort's mortality within 30 days
an adjusted relative risk of 0.84
measured
unknown
null
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC7787261/
null
null
null
[]
AIBD-1738
2016-09-26
source publication
day
intended
CreditVidya's Mantra: Credit At Fair Rates For All
CreditVidya's founders said in 2016 that their analysis of alternative data lets lenders consider Indians with no credit history, and that half of rejected loan applicants are new to credit.
CreditVidya scores first-time borrowers in India from alternative data
null
null
claimed
The makers' claim
null
very low
program
false
CreditVidya alternative data credit scoring
null
CreditVidya
India
Lenders in India (not named)
India
India
India
Consumer lending (alternative data scoring)
null
Finance
Domestic Commerce
Consumer Finance
Prediction from data
Consumers
8
Lower middle income
economic_access
Credit access
null
null
null
null
null
null
null
personal
1
null
0
null
null
null
null
unknown
null
not shown
null
null
null
null
[]
AIBD-0296
2016-11-01
event
month
reported
A seizure alert from an Empatica Embrace reaches the mother of a 26-year-old teacher, who calls paramedics after a seizure left her daughter not breathing (Fulton, Missouri)
An Embrace wristband detected a seizure that stopped a 26-year-old teacher's breathing. It alerted her mother, who called paramedics. IEEE Spectrum reports that she survived.
Embrace wristband alerts a mother, who calls paramedics
null
In November 2016 a 26-year-old teacher from Fulton, Missouri, had an epileptic seizure. It left her unresponsive and not breathing.[1] Her Embrace wristband detected the seizure and alerted her mother, who called paramedics. The teacher survived. Her rescue rests on a single IEEE Spectrum report. The U.S. Food and Drug...
reported
Independently reported
no comparison
very low
harm prevented
false
Empatica Embrace wrist-worn seizure detection (electrodermal activity and accelerometer, on-board algorithm) with the Alert smartphone app
Embrace (the article does not state the generation); echoes concern Embrace2
Empatica
United States
null
null
Fulton, Missouri, United States
United States
Healthcare (epilepsy; seizure monitoring)
Prescription wearable seizure monitor used at home
Health
Health
Disease Prevention
Prediction from data
Patients
3
High income
physical_health
null
5
1
5
null
5
1
null
personal
5
null
3
seizures recorded during more than 6,500 hours of monitoring of 135 patients (device test reported by IEEE Spectrum)
null
the device correctly identified nearly every single seizure, with a false alarm rate below that of the seizure frequency for most individuals
measured
persisted
2026-01-07
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC8418082/
null
null
null
[]
AIBD-0793
2016-11-09
source publication
day
tested
FDA finds that radiologists read automated breast ultrasound a third faster with QView's QVCAD, with non-inferior cancer detection
In a reader study that FDA reviewed, 18 radiologists took 144 seconds per ultrasound case with QVCAD against 215 seconds without it, with non-inferior AUC.
*QVCAD* cut reading time for breast ultrasound by a third
a 71-second or 33% reduction
null
found
Found by a regulator
controlled comparison
very low
aggregate result
false
QVCAD System
null
QView Medical, Inc.
United States
QView Medical (reader study with 18 radiologists on 185 automated breast ultrasound cases)
United States
United States
United States
Healthcare
null
Health
Health
Disease Prevention
Images and video
Health workers
3
High income
operational_improvement
Enhanced productivity
null
2
1
null
null
1
null
organizational
2
null
0
The same radiologists reading the same cases without QVCAD
215 seconds per case
144 seconds per case
measured
scaled
2017-12-21
shown
https://www.accessdata.fda.gov/cdrh_docs/pdf15/P150043B.pdf
null
null
null
[]
AIBD-0955
2016-12-01
event
month
tested
FRA and TTCI report that KLD Labs' machine vision system captured measurements of railcar trucks with more than 97 percent reliability in revenue service
An FRA-sponsored report shows that KLD Labs' machine vision system measured railcar trucks with more than 97 percent reliability at a CSX site. Uptime averaged 80% against a 99% target.
*KLD truck component inspection system* measures railcar parts at CSX
more than 97 percent reliability
null
claimed
The evaluator's account
no comparison
very low
aggregate result
false
KLD truck component inspection system (AxleScan and TruckScan modules)
null
KLD Labs, Inc.
United States
FRA and Association of American Railroads test installation at CSX Transportation's Hague site south of Waycross, Georgia (revenue service, December 2016 to December 2018)
United States
Hague site south of Waycross, Georgia, United States
United States
Railway safety / freight car inspection
null
Safety and rescue
Transportation
Railroad Travel
Images and video
null
9
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
null
null
null
null
unknown
null
shown
https://railroads.dot.gov/sites/fra.dot.gov/files/2020-08/In-service%20Performance%20Truck%20Component.pdf
null
null
null
[]
AIBD-2231
2016-12-01
event
month
reported
Effect of a machine learning-based severe sepsis prediction algorithm on patient survival and hospital length of stay: a randomised clinical trial
In a randomized trial of 142 intensive care patients in San Francisco, alerts from a machine learning sepsis predictor cut in-hospital mortality from 21.3% to 8.96%.
Dascena's sepsis prediction algorithm cut hospital deaths in a trial
from 21.3% to 8.96%
null
studied
Independently studied
randomized trial
moderate
aggregate result
false
Dascena machine learning severe sepsis prediction algorithm
null
Dascena
United States
University of California, San Francisco Medical Center
United States
San Francisco, California
null
Intensive care (early warning of severe sepsis)
null
Health
Health
Disease Prevention
Prediction from data
Patients
3
null
physical_health
Medical diagnosis
5
2
5
null
3
5
null
personal
3
null
0
Patients monitored only by the hospital's existing sepsis detector, by random allocation
21.3% in-hospital mortality
8.96% in-hospital mortality
measured
unknown
2021-09-23
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC5687546/
null
null
null
[]
AIBD-0832
2016-12-02
source publication
day
intended
The Supreme Court of Victoria ordered technology assisted review for about 1.4 million documents that would take over 23,000 hours to review by hand
In December 2016 the Supreme Court of Victoria ordered technology assisted review for about 1.4 million documents. It said that manual review would take a junior solicitor over 23,000 hours.
Victorian court orders *TAR* for 1.4 million pipeline documents
over 23,000 hours
null
found
Found by a court
null
very low
benefit created
false
Predictive coding technology (TAR or Technology Assisted Review)
null
Not named in the judgment
null
McConnell Dowell Constructors (Aust) Pty Ltd (plaintiff), with the agreement of all parties
Australia
Melbourne, Australia
Australia
Legal services / e-discovery
null
Work and industry
Law and Crime
Court Administration
Text and language
Other workers
16
High income
operational_improvement
Enhanced productivity
null
null
null
null
null
null
null
organizational
1
null
0
Manual review of 1,400,000 documents at one minute each
over 23,000 hours, equating to over 583 working weeks
null
assumed
persisted
null
shown
https://web.archive.org/web/20251018042102if_/https://www.austlii.edu.au/cgi-bin/viewdoc/au/cases/vic/VSC/2016/734.html
null
null
null
[]
AIBD-0222
2016-12-09
source publication
day
tested
Karolinska eye-tracking classifier flags 9โ€“10-year-olds at risk of persistent reading difficulty with 95.6% cross-validated accuracy (PLoS ONE 2016)
A support-vector machine classifier from Karolinska Institutet identified 9โ€“10-year-olds at risk of persistent reading difficulty from under one minute of eye tracking with 95.6% accuracy.
Software flags Swedish children at risk of dyslexia from eye movements
reached 95.6% (ยฑ 4.5%) accuracy with balanced sensitivity and specificity
null
studied
Independently studied
tested against a standard
very low
aggregate result
false
Eye-tracking during reading with a support-vector-machine classifier (SVM-RFE, 48 features)
null
Karolinska Institutet (Nilsson Benfatto, ร–qvist Seimyr and colleagues); two authors hold equity in Optolexia
Sweden
None in the sources (research study)
null
null
null
Education / learning-disability screening
Dyslexia risk screening of 9-10-year-olds
Accessibility
Education
Special
Prediction from data
Children; People with disabilities
4
null
incentivization
null
null
null
null
null
null
null
null
personal
7
null
1
chance-level classification on the same cross-validation (and random feature selection)
Chance: 49.1% ยฑ 13%
95.6% ยฑ 4.5%
measured
scaled
2025-03-30
shown
https://europepmc.org/article/PMC/PMC5147795
null
null
null
[]
AIBD-0331
2017-01-01
event
year
reported
TCU auditors identify overlapping tenders for maintenance of BR-153 after an email from Alice
Alice, the federal audit court's robot that reads tenders, emailed tender details to auditors. They found road-maintenance notices that overlapped a concession. The court ordered them annulled.
Auditors use an Alice email to stop overlapping road tenders in Brazil
R$39 million for maintenance of the BR-153
null
reported
Independently reported
no comparison
very low
harm prevented
false
Alice (Anรกlise de Licitaรงรตes e Editais), TCU/CGU text-analysis and data cross-checking robot
null
Tribunal de Contas da Uniรฃo (TCU) with the Ministรฉrio da Transparรชncia e Controladoria-Geral da Uniรฃo (CGU), in-house
Brazil
Tribunal de Contas da Uniรฃo (TCU), Goiรกs unit
Brazil
Goiรกs, Brazil (BR-153 highway)
Brazil
Public administration: external audit and procurement control
Federal court of accounts (anti-waste and anti-corruption control)
Public sector
Government Operations
Procurement & Contractors
Text and language
Public servants
16
Upper middle income
social_improvement
Transparency and accountability
2
1
1
null
5
1
null
personal; organizational; societal
5
null
2
TCU auditors reading the roughly 200 federal tender notices published each day by hand
seria muito custoso porque sรฃo, em mรฉdia, 200 editais por dia
Os auditores do Tribunal de Contas da Uniรฃo recebem pontualmente ร s 19h um e-mail de Alice. Sรฃo os resumos das centenas de contrataรงรตes federais publicadas naquele dia.
assumed
scaled
2018-03-18
shown
https://pmc.ncbi.nlm.nih.gov/articles/PMC10039446/
R$ 39 milhรตes
value of the two DNIT tender notices for BR-153 maintenance that a TCU acรณrdรฃo blocked
false
[]
AIBD-0681
2017-01-01
event
year
reported
Cour des comptes finds that the ClaudIA chatbot absorbed a fivefold rise in Chorus Pro user requests without extra staff
The Cour des comptes found that the ClaudIA chatbot absorbed Chorus Pro user requests without extra staff. Requests rose from 20,000 to 100,000 a month.
*ClaudIA* chatbot absorbs a fivefold rise in e-invoicing questions
from 20,000 to 100,000 a month
null
found
Found by an auditor
no comparison
very low
aggregate result
false
ClaudIA, the Chorus Pro chatbot
null
Agence pour l'informatique financiรจre de l'ร‰tat (AIFE), using IBM Watson technology according to a 2018 trade report
France
AIFE, for suppliers using the Chorus Pro e-invoicing portal
France
France (national)
France
Public finance (e-invoicing support)
null
Finance
Government Operations
Bureaucracy
Text and language
Consumers
16
High income
operational_improvement
Enhanced productivity
1
1
4
null
3
5
null
organizational
2
null
0
null
null
null
null
unknown
null
shown
https://www.lemondeinformatique.fr/actualites/lire-chorus-pro-active-un-chatbot-sous-ibm-watson-71287.html
1,3 Mโ‚ฌ
Annual savings planned under the transformation contract for ClaudIA, of which the AIFE reports 55% achieved
true
[]
AIBD-0713
2017-01-01
event
year
reported
Sweden's Social Insurance Inspectorate finds that claims picked by Fรถrsรคkringskassan's risk model had errors in 55% of checks, against 34% at random
Sweden's Social Insurance Inspectorate found that 55% of parental-benefit claims picked by a risk model in 2017 had errors, against 34% of claims checked at random.
Swedish risk model finds more errors in parental-benefit claims
55% of parental-benefit claims picked by a risk model
null
found
Found by an auditor
controlled comparison
low
aggregate result
false
Fรถrsรคkringskassan's risk-based selection profile (urvalsprofil) for checks of temporary parental benefit, run through the RaKuR infrastructure
null
Fรถrsรคkringskassan (Swedish Social Insurance Agency), Urvalsanalys unit
Sweden
Fรถrsรคkringskassan (Swedish Social Insurance Agency)
Sweden
Sweden
Sweden
Social insurance control
null
Finance
Social Welfare
Child Care
Prediction from data
null
16
High income
operational_improvement
Enhanced productivity
2
1
null
2
3
null
null
organizational
2
null
0
Claims picked at random for checks in 2017
34 procent
55 procent
measured
withdrawn
null
shown
https://www.imy.se/globalassets/dokument/beslut/2025/beslut-tillsyn-forsakringskassan.pdf
null
null
null
[]
AIBD-0727
2017-01-01
event
year
mixed
Stanford evaluators find that the Allegheny Family Screening Tool made screen-in decisions more accurate, with a possible small loss for screen-outs
Stanford evaluators found that the Allegheny Family Screening Tool, with related policies, made screen-ins more accurate, with roughly 24 more children a month screened in accurately.
Screen-in calls more accurate with the *Allegheny Family Screening Tool*
Roughly 24 more children each month were screened in accurately
null
reported
Independently reported
before and after
very low
aggregate result
false
Allegheny Family Screening Tool (AFST), a predictive risk model that scores child welfare hotline referrals
null
Not named in the evaluation (built for the Allegheny County Department of Human Services)
null
Allegheny County Department of Human Services, Office of Children, Youth and Families (hotline call screeners)
United States
Allegheny County, Pennsylvania, United States
United States
Child welfare
null
Public sector
Law and Crime
Family Issues
Prediction from data
Children; Public servants
16
High income
operational_improvement; physical_health
Faster data management, decision-making
2
4
4
2
2
5
0.42
organizational
6
null
0
General protective service referrals before the tool, January 2015 to July 2016
null
Roughly 24 more children each month screened in accurately
measured
challenged
2023-01-31
shown
https://www.alleghenycountyanalytics.us/wp-content/uploads/2024/05/24-ACDHS-04-Predictive-Risk-Algorithms.pdf
null
null
null
[]
End of preview. Expand in Data Studio

Artificial Intelligence Benefits Database (AIBD)

The Artificial Intelligence Benefits Database (AIBD) documents the benefits of artificial intelligence around the world. It contains 2,205 benefits, located in 146 countries, dated from September 1995 to November 2026. Cite the AIBD by its DOI: https://doi.org/10.7910/DVN/HZ8IIB.

The AIBD as it stands today is at petervartanian.xyz/aibd.

Files

File What it has
records.json every benefit, with its sources
records.csv, ratings.csv, echoes.csv the same as tables: the benefits, their ratings, the attempts to repeat them
schema.json the definition of every field
profiles.json, profile-schema.json, requirements.json the public profile of each benefit, with its definitions
measurements.csv, measurements.json, research-schema.json the measured outcomes
fructus.csv, fructus.json, fructus-schema.json the amounts of money
collections.json, with a JSON and a CSV file for each collection the collections
for-*.csv, readers.json the benefits under the questions of five documents
manifest.json a checksum of every file
datapackage.json, csv-metadata.json, ro-crate-metadata.json descriptions for programs
DATASHEET.md how the AIBD was made

Citation

The citation in BibTeX is as follows.

@misc{vartanian_aibd,
  author = {Vartanian, Peter H.},
  title = {Artificial Intelligence Benefits Database},
  year = {2026},
  url = {https://petervartanian.xyz/aibd/},
  doi = {10.7910/DVN/HZ8IIB},
  note = {Edition 2026-11-16-bb8aa6fb86ef}
}
Downloads last month
1,162