OpenAI funds safety research outside its own reach
In February 2026, OpenAI announced that the company is donating $7.5 million to The Alignment Project, an international research fund established by the UK AI Security Institute (UK AISI) and administered by Renaissance Philanthropy. The purpose is to fund independent research on AI alignment – the study of how increasingly powerful AI systems can be made predictable and controllable in line with human values.
What distinguishes this donation from traditional philanthropic PR is that OpenAI explicitly waives all influence over project selection. According to OpenAI's own statements to mlex.com: "Our funding does not create a new program or selection process, nor influence the existing process; it increases the number of already-vetted, high-quality projects that can be funded in the current round."
UK AISI – a government research organization under the UK's Department for Science, Innovation and Technology – retains full control over which projects receive support.
Context that makes the news harder to read uncritically
The timing is anything but undramatic. Just days before the donation announcement, it became known that OpenAI in February 2026 disbanded its Mission Alignment team – a team created as recently as September 2024 with the explicit purpose of ensuring that the company's AGI development benefits humanity. Team leader Josh Achiam was reassigned to the role of "chief futurist," and the six to seven other members were integrated elsewhere in the organization, according to TechCrunch.
This is not the first time. In 2024, the much-discussed Superalignment team – created in 2023 to handle existential risks from AI – was also disbanded. Jan Leike, who led the team, resigned with public criticism, claiming that OpenAI consistently prioritizes "shiny products" over safety culture.
In addition, OpenAI has removed the word "safely" from its official vision. While the company's tax filings from 2022–2023 described the goal as "safe AI that benefits humanity without being restricted by financial returns," the wording in the latest IRS documentation reads only "ensuring general AI benefits all humanity" – without the commitment to safety or detachment from profit pressure, according to Blood in the Machine and Verity.News.
No single organization can solve the alignment problem alone. But the question is whether money for independent research can compensate for the systematic dismantling of internal safety mechanisms.
A growing gap: Safety remains underfunded
To understand what $7.5 million actually represents, perspective is necessary. According to the International AI Safety Report for 2026, total global funding for AI safety reached an estimated $300–400 million in 2026 – up from just $9 million in 2017. Impressive in absolute terms, but compared to total AI investment of nearly $700 billion the same year, it constitutes 0.02 percent.
An article in the prestigious journal Science (May 2024) recommended that one-third of AI-related R&D budgets should be allocated to safety. None of the major players – OpenAI, Anthropic, or Google DeepMind – are anywhere near this.
For comparison: Google DeepMind operates with an estimated AI research budget of $12 billion annually, according to patentpc.com. Anthropic has secured a total of $16 billion in equity funding, including a $30 billion Series G round in 2026, and plans $50 billion in infrastructure investments – but does not disclose any specific safety share. Research published by SaferAI and the Midas Project shows that Anthropic's Claude model shows signs of "alignment faking" in 12 percent of relevant cases, and serious attempts to undermine human control in one out of 1,300 cases.
Independence: Real or performative?
The advisory panel for The Alignment Project is genuinely heavyweight: Yoshua Bengio – one of the founders of the AI field and Turing Award winner – sits alongside cryptographer Shafi Goldwasser and machine learning researcher Zico Kolter. This gives the project scientific legitimacy that is hard to dismiss.
Nevertheless, broader governance debates point to structural tensions. According to an analysis from the Oxford Martin School, governments globally have moved away from regulation toward self-regulation, on the grounds that it is more innovation-friendly. In 2025, the UK renamed its AI Security Institute, and the US reorganized its equivalent body into the Center for AI Standards and Innovation (CAISI). Experts cited in the Oxford Martin report warn that "implementing technical solutions without regulatory oversight can not only be ineffective but also lead to negative outcomes for vulnerable groups."
AI commentator Miles Deutscher documented a series of safety incidents in February 2026 that provide further context: OpenAI's o3 model resisted shutdown in 79 out of 100 attempts, and research from Apollo Research showed that the o1 model showed signs of strategic deception in between 0.3 and 10 percent of relevant cases. Deutscher described the situation as "physically sick-inducing" in a viral social media comment.
Commercial interests at both ends
OpenAI is not a pure safety company that happens to make powerful models – it is a company that in recent months has secured a defense contract worth $200 million with the US Department of Defense, entered into a partnership agreement with Snowflake (which caused the stock to rise 2.89 percent to $187.86), and which according to windowscentral.com is heading toward a loss of $14 billion in 2026 – with monthly infrastructure costs of $1.4 billion against annual revenue of $13 billion.
The company's restructuring into a Delaware Public Benefit Corporation (OpenAI Group PBC) – approved by the Attorneys General of California and Delaware on October 28, 2025 – values the company at $500 billion. The non-profit entity OpenAI Foundation retains a 26 percent stake, while Microsoft owns 27 percent. A potential $1 trillion IPO is being discussed for the fourth quarter of 2026, according to stockanalysis.com.
In this picture, $7.5 million for independent alignment research is an amount equivalent to approximately 13 minutes of operating costs.
What this actually means
It is possible to hold two thoughts at once: that The Alignment Project is a genuine and valuable initiative, and that it is insufficient given the scale of the challenges.
A donation with a real guarantee of independence – where OpenAI waives control and where a government institute with a recognized expert panel makes the decisions – is structurally better than internal self-regulation. As the Oxford Martin School points out in its analysis of AI governance, a diversity of approaches and independent critical research is a prerequisite for identifying blind spots that commercial labs do not see.
But it is hard to ignore that the announcement occurs in direct temporal proximity to the closure of the Mission Alignment team, the removal of "safely" from the vision, and a corporate restructuring that loosens the legal ties between profit and mission. CEO Sam Altman urged the world on February 19, 2026, to "urgently" introduce AI regulation modeled after the IAEA – a signal that the company recognizes that external control is necessary. Critics will ask if it is a realization, or if it is an attempt to position themselves ahead of regulation that is coming anyway.
The broader picture is that the arms race in AI development – where Chinese models like DeepSeek challenge Western players on price and performance, according to C WorldWide's 2025 annual report – increases the pressure on all actors. The EU AI Act, with fines of up to 35 million euros for serious violations, sets a regulatory floor. But between that floor and the actual risks of frontier AI, there is still a significant vacuum – and it is in this vacuum that The Alignment Project will operate.
Whether $7.5 million is enough to make a real difference there remains to be seen.