Summary: On February 20, 2026, as the AI Impact Summit in India drew to a close, the UK government delivered its own announcement: OpenAI and Microsoft had formally signed on to the international coalition behind the AI Security Institute’s Alignment Project. OpenAI committed £5.6 million, Microsoft added further support that the official release leaves unquantified, and the research funding the program can now draw on tops £27 million.
On February 20, 2026, as the AI Impact Summit in India drew to a close, the UK government delivered its own announcement: OpenAI and Microsoft had formally signed on to the international coalition behind the AI Security Institute’s Alignment Project. OpenAI committed £5.6 million, Microsoft added further support that the official release leaves unquantified, and the research funding the program can now draw on tops £27 million. Scheduling the statement for the summit’s closing day is a message in itself: alongside the contest over model capability, governance and safety research are being pushed rapidly up the agenda.
What deserves a closer look is how the arrangement is built. Rather than merely growing their in-house safety teams, two flagship American AI companies are handing money to a program coordinated by a UK government research institute — one that funds independent researchers. As models take on more autonomy, alignment (keeping systems reliably doing what operators intend, without unintended harmful behavior) has left the academic seminar and become a gate that enterprise adoption has to pass through.
How the Alignment Project Works
First unveiled in summer 2025 under the lead of the UK’s AISI, the Alignment Project runs as a coalition. Training models is not part of its job. What it offers external researchers is a package of three: grant funding, compute access, and mentorship from AISI scientists themselves. An expert advisory board — counting Yoshua Bengio, Zico Kolter, Shafi Goldwasser, and Andrea Lincoln among its members — keeps the research direction under review.
The coalition’s membership cuts across government and industry: Canada’s CIFAR, Australia’s AI Safety Institute, Schmidt Sciences, AWS, Anthropic, UKRI, ARIA, and a set of funds focused on AI safety all appear on the roster. With OpenAI and Microsoft arriving, the “frontier labs funding it directly” piece is finally in place.
How the £27 Million Breaks Down
- OpenAI: £5.6 million committed — the headline number of this announcement
- Microsoft: extra support, with no separate figure given in the official release
- Cumulative pool: more than £27 million in total
- Round two: set to open for applications this summer
Look at what the money is actually for. It pays for alignment research carried out by outside groups; none of it flows into the UK regulator’s budget. Seen that way, the labs are spending to enlarge research capacity that operates independently of them, and agreeing to a priority-setting process they cannot fully control.
Eight Countries, Sixty Projects
Round one has already been paid out: 60 research projects spread across eight countries. The UK’s own framing keeps the aims grounded — hold increasingly powerful models to predictable, controllable behavior, and turn the verification methods into public assets rather than internal paperwork locked inside each lab. Deputy Prime Minister David Lammy’s formulation: “we will always be clear-eyed on the need to ensure safety is baked into it from the outset.” AI Minister Kanishka Narayan made the causality plainer still: “We can only unlock the full power of AI if people trust it.”
Why Labs Put Their Own Money Behind Independent Safety Research
The most direct answer came from Mia Glaese, OpenAI’s VP of Research: “As AI systems become more capable and more autonomous, alignment has to keep pace.” For vendors pushing agents into enterprise workflows, safety research is not a public-relations line item — it is a dependency of the product roadmap. Signing onto a government-coordinated program with an independent advisory board secures three things at once: outside research capacity, a working relationship with regulators, and a verifiable record of actually paying for safety.
What Developers and Product Teams Should Take From It
Three consequences. First, if you sell AI into government, finance, or critical infrastructure, expect model-testing artifacts and governance controls to harden into procurement baselines — evaluation methods and tools coming out of programs like this tend to end up as audit vocabulary. Second, with the second funding round opening this summer, small teams working on alignment, evaluation, or interpretability pick up a funding route that bypasses venture capital. Third, read alongside the rest of the week’s news (see our AI weekly roundup), models, compute, and governance are advancing in lockstep — 2026 AI buying decisions will be judged against three report cards at once.