MicheleStingo/kindynos-causal-risks

Kindynos causal risks

AI risk descriptions labeled with the three axes of the MIT Causal Taxonomy of AI Risks: who or what causes the risk (entity), whether the harm is intended (intent) and when it occurs (timing). Two sources, with the same core fields:

• kindynos: 605 risks in English and Italian written by Kindynos, an open-source AI risk assessment tool, from simulated questionnaires about AI systems. The texts were generated with Gemini 2.5 Flash. The train and validation labels are LLM-annotated (two independent annotators and an adjudicator, following written guidelines); the test labels were verified by a human annotator.
• mit: the 1,473 risks of the MIT AI Risk Repository v4 that carry causal labels: excerpts of research papers, labeled by the repository's annotators (CC BY 4.0).

These are the data of laya-kindynos-causal, split exactly as they were used for training, model selection and testing. They are also on Kaggle.

Config	Split	Risks	Languages	Labels
kindynos	train	363	174 EN, 189 IT	LLM-annotated: two independent annotators and an adjudicator
kindynos	validation	97	45 EN, 52 IT	LLM-annotated: two independent annotators and an adjudicator
kindynos	test	145	85 EN, 60 IT	verified by a human annotator
mit	train	1,116	EN	MIT AI Risk Repository annotators
mit	validation	197	EN	MIT AI Risk Repository annotators
mit	test	160	EN	MIT AI Risk Repository annotators

from datasets import load_dataset

area = load_dataset("MicheleStingo/kindynos-causal-risks", "kindynos")
mit = load_dataset("MicheleStingo/kindynos-causal-risks", "mit")

Fields

Both configs:

• id: for Kindynos <run>-<subdomain>-<n>; for MIT the repository's Ev_ID, which the source repeats for a few different risks (62.31.01–03 each mark two).
• subdomain_id, subdomain_name: the risk's subdomain in the MIT Domain Taxonomy, named as in the MIT AI Risk Repository.
• title, text: the risk's title and description.
• entity (ai, human, other), intent (intentional, unintentional, other), timing (pre-deployment, post-deployment, other).

kindynos only; every split has every column, empty where it does not apply:

• language (en, it), run, severity (the severity Kindynos gave the risk: low, medium, high).
• profile: the simulated user, beginner, intermediate or expert; empty for the Kindynos example assessment.
• system: the AI system the simulated user was asked to describe (train and validation).
• text_model: the LLM that wrote the text, gemini-2.5-flash; empty for the Kindynos example assessment, whose model was not recorded.
• label_source (train and validation): for each axis, agreed when both annotators gave the label, adjudicated when the adjudicator chose it.
• source (test): generated for the six simulated assessments, example for the Kindynos example assessment.
• changed_in_review and harmonized (test): the axes the human annotator corrected, and those harmonized across near-identical risks, with the reason.

mit only: paper_id.

How the Kindynos risks were made

1. Questionnaire. Kindynos asks 24 questions about an AI system. Gemini 2.5 Flash answered them in the role of a beginner, intermediate or expert user. Each of the 28 train and validation runs describes a given fictional system, for example a CV screening and candidate ranking system used by a recruitment agency or un sistema di gestione adattiva dei semafori per il traffico di una città, in English and Italian alternately. In the six test runs (two languages, three profiles) the simulated user chose the system to describe; the 28 train and validation systems were chosen to differ from those. The seventh test run is the Kindynos example assessment.
2. Risks. The Kindynos domain analyzer (Gemini) turned each set of answers into risks, grouped by MIT subdomain, each with a title, a description, a severity and a mitigation. The dataset keeps all but the mitigation.
3. Splits. Whole runs go to one split: 22 runs to train, 6 to validation (three per language, drawn at random with a fixed seed), 7 to test. Six entries in which the analyzer states that a subdomain carries no significant risk were removed, because they are not risks.

Labeling

The labels follow the MIT definitions and the decision rules in LABELING.md, written for the Kindynos texts: label what the text states, and use other when it does not settle an axis; entity is whose decision or action produces the harm, as the text frames it; intent is about the harm, not the action; timing is when the harm occurs, so the harms of a system in use are post-deployment even when introduced earlier.

• Train and validation (460 risks): two LLM annotators labeled each risk independently, following the written guidelines; an adjudicator resolved their disagreements (49 labels on 42 risks), with a reason for each. Agreement between the annotators: Cohen's κ 0.93 / 0.89 / 0.97 (entity / intent / timing). The prompts and the raw annotations are in annotations/.
• Quality check: on a sample of 60 risks with human-verified labels, the protocol's labels matched 95 / 85 / 100% of them (quality_check/).
• Test (145 risks): pre-annotated by a model with the same guidelines, then verified one by one by a human annotator, who corrected 25 of the 435 labels; then 17 labels of near-identical risks were harmonized, to the MIT annotators' convention for the subdomain where it is clear.

Split	Entity: ai / human / other	Intent: intentional / unintentional / other	Timing: pre / post / other
kindynos train	48 / 39 / 13%	20 / 67 / 13%	0 / 93 / 7%
kindynos validation	45 / 46 / 8%	24 / 61 / 15%	0 / 93 / 7%
kindynos test	54 / 35 / 10%	23 / 58 / 19%	1 / 91 / 8%
mit train	44 / 36 / 20%	32 / 37 / 31%	12 / 61 / 27%

The mit config

The rows of the MIT AI Risk Repository v4 (December 2025) with causal labels, extracted by data_prep/extract_dataset.py of Kindynos: classifiable rows only, labels normalized to the values above, title and description kept, subdomain added from the repository's domain coding. The split keeps each risk category with its subcategories: the test split holds 160 rows from groups not used for training; the validation split holds whole groups making up 15% of the remaining rows, which laya-kindynos-causal left out of its training data.

Uses

• Training and evaluating classifiers for the MIT Causal Taxonomy, especially on short risk statements in English and Italian.
• Studying labeling protocols with LLM annotators against human-verified labels.

Provenance and license

CC BY 4.0. The mit config and the taxonomy definitions come from the MIT AI Risk Repository (Slattery et al., 2025, CC BY 4.0). The Kindynos texts were generated with Gemini 2.5 Flash by the Kindynos pipeline. The train and validation labels are LLM-annotated (two independent annotators and an adjudicator), with the protocol checked against human-verified labels on a sample; the test labels were verified by a human annotator.

Citation

@misc{stingo2026kindynoscausalrisks,
  title  = {Kindynos causal risks: AI risk descriptions labeled with the MIT Causal Taxonomy},
  author = {Stingo, Michele},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MicheleStingo/kindynos-causal-risks}
}

@article{slattery2025airisk,
  title   = {The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence},
  author  = {Slattery, Peter and Saeri, Alexander K. and Grundy, Emily A. C. and Graham, Jess and Noetel, Michael and Uuk, Risto and Dao, James and Pour, Soroush and Casper, Stephen and Thompson, Neil},
  journal = {arXiv preprint arXiv:2408.12622},
  year    = {2025}
}