Ask people to guess how much the world spends making sure advanced AI goes well, and most land somewhere in the billions. The scientists building it keep warning it could be dangerous; surely the response is funded like a serious problem. It is not. Add up every philanthropic grant, every government safety program, and every dollar the frontier labs spend on their own safety teams, and the total comes to something like $190 million a year. Global investment in AI capabilities passed $250 billion a year some time ago. For every thousand dollars spent making AI more powerful, less than one dollar is spent making it safe.
The $190 million figure deserves scrutiny, because estimates in this field are genuinely messy. Published totals run from about $110 million to $250 million a year depending on whether you count government institutes, lab-internal teams, and academic work. What follows is our own accounting, built from public grant databases and budget disclosures, funder by funder.
Where the money comes from
| Source | Est. per year |
|---|---|
| Open PhilanthropyBacked by Good Ventures; over $336M to AI safety since 2014, roughly $46–50M a year recently | $50M |
| Frontier lab safety teamsInternal safety research at Anthropic, OpenAI, Google DeepMind — estimated, not disclosed | $35M |
| Survival & Flourishing FundBacked largely by Jaan Tallinn; its 2024 round alone allocated over $19M | $30M |
| UK governmentAI Security Institute plus ARIA's £59M Safeguarded AI program | $30M |
| Other philanthropy & individualsLongview Philanthropy, the Long-Term Future Fund, smaller foundations, individual donors | $25M |
| US, EU & other governmentsNIST's CAISI, NSF safety programs, the EU AI Office | $20M |
| Total | ~$190M |
Two things stand out. First, this is not a diversified field. One private foundation supplies more than a quarter of everything, and more than half of all philanthropic money. Second, the government share is recent and fragile — most of it dates from 2023, and both the UK and US institutes dropped the word "safety" from their names within two years of being founded.
Where the money goes
No funder publishes a spending breakdown by purpose, so we classified the grants ourselves: took each source above, sorted its known grants and programs into categories, and summed across funders. Open Philanthropy's public grant database shows about $28 million in technical safety grants for 2024, of which 68 percent went to evaluations and benchmarking. The UK's AI Security Institute is, by mandate, an evaluations body, while ARIA's Safeguarded AI program is technical research. The Survival and Flourishing Fund's grantees span technical work (MIRI), evaluations (METR, Apollo Research), and advocacy (the CAIS Action Fund). Longview's AI program leans toward governance. Lab-internal budgets are almost entirely technical. The result:
| What it pays for | Est. per year | Share |
|---|---|---|
| Technical safety researchAlignment, interpretability, robustness — lab teams, ARIA, NSF, academia, nonprofits | $80M | 42% |
| Model evaluations & benchmarkingDangerous-capability testing — UK AISI, METR, Apollo, Open Phil's evals grants | $50M | 26% |
| Governance & policy researchTreaty design, safety standards, compute governance | $30M | 16% |
| Field-building & talentFellowships, research training programs, career pipelines | $20M | 11% |
| Advocacy & public engagementCampaigns, open letters, grassroots organizing | $10M | 5% |
The category totals are estimates built on estimates, and many grants straddle lines — an evaluations nonprofit does technical work; a governance institute runs fellowships. But the shape is robust to any reasonable reclassification: technical research and evaluations absorb about two-thirds of everything, and advocacy — the work of actually changing what governments and companies do — gets a twentieth.
Avatar: The Way of Water cost more to produce than the entire world spends on AI safety in a year. A single frontier training cluster costs more than a decade of the field's total budget. The US federal government spends the field's annual total on defense roughly every ten minutes.
Reason one: a single dominant funder
Most philanthropic AI safety money — over half — comes from one place: Good Ventures, the foundation of Facebook co-founder Dustin Moskovitz and Cari Tuna, giving through Open Philanthropy. That concentration has consequences beyond fragility. When an organization falls outside what Good Ventures will fund, the rest of the donor pool is so shallow that, by one estimate from a former grantmaker in the field, it faces a funding bar 1.5 to 3 times higher than everyone else. And when the dominant funder shifts priorities, the whole field shifts with it: Open Philanthropy's technical safety grantmaking fell from a 2022 peak of over $60 million to $28 million in listed 2024 grants, with most of what remained going to evaluations. No other funder stepped in to fill the difference. In a healthy field, no single donor's board meeting can reshape the research agenda. This is not yet a healthy field.
Reason two: no constituency
Cancer research raises tens of billions a year because everyone knows someone. Climate philanthropy passed $10 billion a year because the damage is visible and the movement is fifty years old. AI risk has no patients, no flooded towns, no grieving families — the harm it warns about has not happened, and if the field succeeds, it never will. Prevention of a risk that has never materialized is close to the hardest possible fundraising pitch. The people most convinced by the argument tend to be researchers, not billionaires, and the general public's concern about AI — which polls consistently find is real and growing — has no obvious donation box attached to it.
Reason three: the labs absorbed the field
A large fraction of the people qualified to do alignment research work at Anthropic, OpenAI, or Google DeepMind, at compensation levels no nonprofit can approach. This distorts the funding picture twice over. It makes safety look better resourced than it is, because the most visible safety work happens inside companies with enormous budgets. And it ties much of the field's research agenda to commercial judgment: a lab's safety spending is real, but it is spending the lab chooses, on questions the lab prioritizes, publishable at the lab's discretion, and cancellable when priorities change — as OpenAI demonstrated when it dissolved its Superalignment team in 2024, a year after publicly committing 20 percent of its compute to the problem. Independent safety research exists precisely to ask the questions labs will not, and independence is exactly what the money does not fund at scale.
Reason four: the money avoids politics
The smallest slice of the chart — 5 percent for advocacy — may be the most consequential. US charity law limits how much lobbying a 501(c)(3) foundation can fund, and most institutional funders are cautious well beyond the legal requirement, wary of politicizing their broader grantmaking. The result is a field that spends $160 million a year producing research and evaluations, and about $10 million making sure anyone with power acts on them. The organizations doing direct political work — PauseAI, ControlAI, the Future of Life Institute's campaign arm — operate on budgets that would embarrass a mid-size congressional race. Whatever one thinks of their tactics, the asymmetry is stark: the largest technology companies each spend more on federal lobbying alone than the entire safety field spends on advocacy.
What would change this
The arithmetic of the field means small amounts move it. A new donor giving $5 million a year would instantly rank among the largest funders of AI safety advocacy on Earth. A government committing one-tenth of one percent of its defense budget to safety research would roughly double the global total. This is the rare cause where the bottleneck is not absorptive capacity — funders themselves say publicly that they cannot cover the opportunities they see — but simply the number of people writing checks.
Money is one lever; policy is the other, and it does not require personal wealth. The case for binding international rules on frontier AI is laid out in our plan, and the fastest way to add your weight to it takes five minutes on the take action page. The interactive version of the funding data in this article lives on our resources page, with sources, and we will keep both updated as new disclosures appear.
Common questions.
Roughly $190 million a year as of 2024–2025: about $105 million in philanthropic grants, $50 million in government safety programs, and an estimated $35 million in internal safety research at frontier labs. Published estimates range from $110 million to $250 million depending on what is counted. Global investment in AI capabilities exceeds $250 billion a year — more than a thousand times as much.
Open Philanthropy, funded primarily by Good Ventures — the foundation of Facebook co-founder Dustin Moskovitz and Cari Tuna. It has directed over $336 million to AI safety since 2014, roughly $46–50 million a year recently, and accounts for more than half of all philanthropic funding in the field.
Classifying each funder's grants by use: about $80 million a year goes to technical safety research, $50 million to model evaluations and benchmarking, $30 million to governance and policy research, $20 million to field-building and talent programs, and about $10 million to advocacy and public engagement.
The risk is preventive and has no affected constituency to organize around; the field is young and was long dominated by a single funder, which discouraged others from building expertise; the most visible safety work happens inside AI companies, making the area look covered; and charity law plus institutional caution steer foundations away from the advocacy work where marginal dollars might matter most.