# AISF (ai safety funds)

> Public philanthropic grants to AI safety, classified by research area and method, updated weekly. Built to show which areas are underfunded relative to the rest of the field.

Data covers Coefficient Giving (formerly Open Philanthropy), EA Funds (Long-Term Future Fund, EA Infrastructure Fund), the Survival and Flourishing Fund, and Manifund. It excludes frontier lab internal spending, venture capital, government funding and private donors.

## Data

- [aisf.json](data/aisf.json): every grant as one record. `meta.fields` gives the column order of each row: id, source, funder, grantee, title, amount (nominal USD), date (YYYY-MM), year, topic, method, inferred, desc, url.

url: the most specific public source. Manifund links to the project page and SFF to the recommendation round. EA Funds and Coefficient Giving publish no per-grant pages, so these link to the EA Funds database and to the archived copy of the Coefficient Giving CSV in this repository.

## Codes

topic: INT interpretability; EVAL evaluations and capability testing; ALN alignment methods and oversight; CTL AI control and model security; ROB robustness and adversarial; AF agent foundations and theory; COOP cooperative and multi-agent AI; GOV governance and policy; CMP compute and hardware governance; FCST forecasting and strategy; FB field-building and talent; COMM communications and awareness; SOC societal and sociotechnical impacts; UNK unclassified.

method: TE technical empirical; TH theoretical; QS qualitative and social science; PA policy analysis and advocacy; INF infrastructure and programmes; UN unclear.

inferred: 1 when a general-support grant was tagged by the organisation's main focus rather than a stated purpose.

## Methodology

Classification is done by Claude (Anthropic) from each grant's title and description. Grants between tracked funders are excluded so money is counted once to avoid double counting. Multi-year grants count in full at award date.