foundinghires/ai_researchers_atlas

Atlas of AI Research

Where the people who do AI research are, what they work on, and when they entered the field. Four tables of counts, measured on an index of 790 725 researchers read off 496 295 AI papers.

These are aggregates. The dataset holds no names, no affiliations at the person level, no contact details and no individual dates. Every number is a count of people, and the smallest cell is a count of people on one subject in one country.

Built and published by Founding Hires, which maintains an index of the published AI research and the public records that date the end of a doctorate. That index is a commercial product; these aggregates are not, and they are released under CC BY 4.0 because the questions they answer come up constantly and nobody else publishes the answer.

The canonical copy, with a manifest, lives at foundinghires.tech/data.html, and the method behind every field is set out at foundinghires.tech/method.html.

Quick start

from datasets import load_dataset

countries = load_dataset("foundinghires/ai_researchers_atlas", "countries")["train"]
countries.to_pandas().sort_values("researchers", ascending=False).head(10)

The four configs are countries, subjects, country_subjects and years. Each is small enough to hold in memory, and the same tables are served as plain CSV and JSON at foundinghires.tech/data.html if you would rather not depend on this library.

What it answers

• Where the people who publish AI research actually sit, and how that compares once you divide by population.
• Which subjects are growing, measured by where newcomers to the field start rather than by where the volume already is.
• How many people entered AI research in a given year, worldwide or in one country, which is the closest public proxy for the size of the incoming cohort.
• How concentrated a subject is, and which countries carry it.

It has been used to write country and subject reports, and it is the same data behind the Atlas of AI research.

The four tables

countries (40 rows)

One row per country with at least a few hundred indexed researchers.

column	meaning
country_code	ISO 3166-1 alpha-2
country	English name
researchers	Indexed researchers whose affiliation sits in that country
early_career	Of those, the ones early in a research career
in_industry	Of those, the ones whose declared employer is a company
newcomers_recent_window	First indexed paper in 2024 or later
newcomers_previous_window	First indexed paper between 2021 and 2023

subjects (108 rows)

Worldwide, one row per topic.

column	meaning
subject	OpenAlex topic name
researchers	Indexed researchers carrying that topic, worldwide
newcomers_recent_window	Of those, a first indexed paper in 2024 or later
newcomers_previous_window	Of those, a first indexed paper between 2021 and 2023

A person can carry several topics, so the column sums to more than the number of people. These counts cover the whole index, while country_subjects covers the countries published here, so the two do not reconcile.

country_subjects (698 rows)

column	meaning
country_code	ISO 3166-1 alpha-2
country	English name
subject	OpenAlex topic name
researchers	Indexed researchers on that subject in that country

years (496 rows)

Entries into the field, by year, for the world and for each country and subject.

column	meaning
scope	world, country or subject
key	Country code or subject name, world for the world row
name	Display name
year	Year of the person's first indexed paper
newcomers	People whose first indexed paper falls in that year

The series stops at the last complete year. The current year is always incomplete, and publishing it would show a collapse that is an artefact of the calendar. It covers the world, every country in countries and every subject in subjects, so it joins cleanly to both.

How it is measured

The index is built from public sources: arXiv for the papers, OpenAlex, ORCID and DBLP for the people behind them. A researcher enters the index by authoring an AI paper, which means the population is authors of published research rather than everyone who works on AI.

• Country is where the institution sits, taken from the most recent affiliation on a paper. Someone who has moved and not yet published from the new address still counts at the old one.
• Subject follows the OpenAlex topic taxonomy. A person can carry several, so shares of subjects are shares of assignments rather than of people.
• Entry year is the year of the first paper the index holds for that person, which approximates but does not equal the start of a career.
• Early career is inferred from how recently the first paper appeared.
• Industry counts a declared employer that resolves to a company, which is available for a minority of profiles and undercounts the rest.

The method, with its limits stated at greater length, is at foundinghires.tech/method.html.

Coverage

Of 790 725 indexed researchers, 553 497 carry a country, and the 40 countries published here hold 534 577 of those. The rest sit in countries too small to publish a count for without pointing at individuals. The manifest carries these three numbers so the arithmetic can be checked against any copy of the files.

Known limits

They measure what is indexed, not what exists. Research published outside the sources above, or in a language the sources cover poorly, is missing. Counts move between two republications because the index grows every night. Comparisons across countries carry the bias of each country's publishing culture, and the share of a country's researchers who publish in English.

Population rates, where they appear on the site, use UN 2024 estimates.

Licence and citation

Creative Commons Attribution 4.0. Use it commercially, republish it, build on it. The one condition is attribution.

Founding Hires, Atlas of AI research, https://foundinghires.tech/data.html

BibTeX:

@misc{foundinghires_atlas,
  title  = {Atlas of AI Research},
  author = {{Founding Hires}},
  year   = {2026},
  url    = {https://foundinghires.tech/data.html},
  note   = {Aggregate counts of AI researchers by country, subject and year of entry}
}

Updates

Rebuilt from the live index and republished whenever the numbers move, which is several times a week. Figures on this card are as of 2026-09-22. The manifest at foundinghires.tech/data/index.json carries the date each table was built and the row counts, which is the quickest way to tell whether the copy you are reading is the current one.