Google, OpenAI and Anthropic Finally Agree on AI Safety — Britain Got There First
DeepMind, OpenAI and Anthropic now want independent testing for frontier AI models. The UK has been running that exact system since 2023.
Something unusual happened in AI policy this month. The chief executives of Google DeepMind, OpenAI and Anthropic — companies that spend most of their time competing for the same customers, the same compute and the same headlines — are now, for the first time, on record making remarkably similar arguments about how their own industry should be regulated. All three want independent testing of frontier AI models before public release. All three want it backed by a formal standards body rather than the current system of companies grading their own homework.
When I looked into what’s actually being proposed, the most striking detail wasn’t the convergence itself. It’s that Britain has been running a version of this exact system since 2023 — years before Silicon Valley decided it wanted one.
Three Rivals, One Diagnosis
Google DeepMind CEO Demis Hassabis set the tone on 14 July with a roughly 1,500-word framework calling for a US-led independent body to test frontier models before they reach the public. Days later, reporting from Axios confirmed that OpenAI and Anthropic leadership are converging on similar prescriptions, even if the structural details differ.
That’s a genuine shift. For years, frontier AI safety testing has largely meant labs publishing their own model cards and running their own internal red-teaming, then telling the public the results. Hassabis and his counterparts are now saying that isn’t good enough anymore — not for models capable of serious harm.
Hassabis’s FINRA-Style Blueprint
The specific model Hassabis is pointing to is FINRA — the Financial Industry Regulatory Authority, a private, industry-funded body that polices Wall Street trading firms under SEC oversight. His pitch: an equivalent body for AI, initially voluntary, that labs would submit frontier models to up to 30 days before release.
If that voluntary phase “proves effective and robust” — his words — formal requirements could follow quickly, meaning frontier models would need to pass independent testing before US market deployment becomes legally possible. It’s a two-stage plan: prove the system works informally, then make it mandatory. I’ve seen this exact regulatory sequencing play out in finance and pharmaceuticals. It rarely stays voluntary for long once the infrastructure exists.
What Mandatory Testing Would Actually Check
The proposed evaluations aren’t vague box-ticking. Under Hassabis’s framework, and in the broader industry conversation Axios reported on, testing would cover cybersecurity capability, biological and chemical weapons risk, deceptive behaviour toward users or evaluators, autonomous agentic capability, and national security implications more broadly.
Four categories, not five — deliberately narrow. The pitch isn’t “regulate every chatbot.” It’s aimed squarely at frontier systems: the handful of most capable models released by a small number of labs each year. Smaller AI products, fine-tunes and open-source derivatives would sit outside the scope entirely, at least under the current proposal.
Britain’s Head Start: Inside the AI Safety Institute
Here’s the part UK readers should actually care about. Rishi Sunak set up what became the UK’s AI Safety Institute back in 2023, funded to the tune of roughly £100 million — around ten times the budget of its American counterpart. AISI has spent nearly three years doing precisely what Hassabis is now proposing as a novel US framework: vetting frontier models on a voluntary basis before they reach the public.
Google DeepMind, OpenAI and Microsoft have all submitted models to AISI for testing already, including evaluations of biological weapons risk and model reliability. UK investors keep asking me why this matters for a country that doesn’t host a single frontier lab of its own — and the honest answer is that Britain built the testing infrastructure the rest of the world is only now deciding it needs.
Anthropic’s Different Angle
Anthropic’s position in this convergence is worth separating out, because its incentives look different from the other two. The company recently overtook OpenAI on revenue — driven largely by Claude Code, the coding agent Anthropic launched into public preview back in February 2025 — while explicitly building its brand around Constitutional AI safety principles rather than chasing government equity deals.
That distinction matters here. Anthropic backing independent testing costs it less commercially than it might cost a lab racing to ship the biggest model fastest. When a company already selling safety as a differentiator agrees to submit to external scrutiny, that’s a lower-friction commitment than when a company optimised purely for capability does the same. Worth remembering when you weigh how much any of these proposals will actually bite.
OpenAI’s IPO Complicates the Picture
OpenAI’s version of cooperation comes with its own strings attached. The company is reportedly offering the US government a 5% equity stake as part of its broader engagement with frontier model standards, timed against a planned IPO as early as September 2026. Google, meanwhile, has quietly delayed the wider release of Gemini 3.5 Pro by several months after internal testing showed it falling short on coding performance and complex reasoning tasks.
None of that undermines the independent testing proposal directly, but it’s a reminder that “we all agree on safety testing” and “we’re all racing to ship and IPO on schedule” are happening in the same breath, from the same executives, in the same month. I wasted an afternoon trying to reconcile those two threads before realising they’re not actually in tension — safety testing that takes 30 days doesn’t meaningfully slow down a lab that’s already spent 18 months on a model.
Why the US Wants to Lead, Not Follow
Hassabis’s framework is explicit that the US, not a fragmented mix of national regulators, should set the terms for any new global standards body. That’s a notable position from a UK-founded company — DeepMind started in London in 2010, before Google acquired it in 2014 — and it signals where the commercial and political centre of gravity actually sits.
The logic isn’t hard to follow. Frontier AI labs are overwhelmingly American-headquartered even when their research roots are elsewhere, and a US-anchored standards body gives those labs more influence over the rules than a UN-style international framework would. Britain gets credited as the pioneer. Washington gets to write the rulebook.
The Risk of a US-Only Rulebook for UK Firms
This is where it gets genuinely tricky for UK businesses building on top of frontier models. If a US-led standards body becomes the de facto global gatekeeper, UK companies could end up depending on compliance decisions made entirely outside British jurisdiction — with no seat at the table despite AISI’s three-year head start.
UK Parliament has already opened its own probe into AI adoption at work, with global adoption figures cited at around 17.8% — a number MPs treated as evidence that regulation is lagging real-world deployment rather than leading it. That anxiety is exactly what a US-only standards body would deepen: Britain funded and built the testing capability years ahead of schedule, and could still end up watching from outside the room where the actual rules get finalised.
The government’s options look fairly narrow: push for AISI to be formally recognised inside whatever US-led structure emerges, negotiate a mutual recognition arrangement so AISI testing counts toward US compliance, or risk becoming a rule-taker on a technology Britain helped pioneer both academically and, through DeepMind, commercially. UK investors keep asking me which of those three is most likely — and honestly, mutual recognition is the pragmatic bet, since rebuilding AISI’s three years of testing infrastructure inside a brand-new US body would waste capability nobody has an incentive to throw away.
What Mandatory Testing Would Mean for Everyday AI Products
It’s tempting to read all this as a story purely about frontier labs and government policy, but the knock-on effects reach further than that. Most UK businesses aren’t training their own frontier models — they’re building products on top of GPT, Gemini or Claude through an API. If mandatory pre-release testing becomes real, the models underpinning those products get delayed at source, not at the point where a UK company integrates them.
That’s a genuine planning risk. A 30-day mandatory testing window before release doesn’t sound dramatic until you’re a startup that’s built a roadmap around a model update landing on a specific date. Firms leaning heavily on frontier AI should build slack into release-dependent roadmaps now, before any formal requirement exists, rather than scrambling once a standards body actually has enforcement power.
There’s an upside too. Independent testing that catches a genuine safety failure before release protects downstream businesses from reputational damage they didn’t cause and couldn’t have predicted. A UK company embedding a frontier model into a customer-facing product carries real exposure if that model turns out to behave dangerously post-launch. External vetting, done properly, is a form of insurance the industry has mostly lacked until now.
What This Means for UK Businesses and AI Users
For UK companies building products on frontier models, the practical takeaway is to watch which labs actually submit to independent testing and which resist it — that distinction is likely to matter more over the next 18 months than any single model release. For everyday users, this convergence is a rare moment where competing AI labs are publicly admitting that self-regulation alone isn’t working. Whether that turns into real accountability, or just better-coordinated PR, depends entirely on whether the “voluntary” phase Hassabis describes ever actually becomes mandatory.
Stay ahead of the market
Join our community of nearly 5,000 across YouTube, LinkedIn, X, and Facebook — weekly crypto, AI, and digital lifestyle insights every Thursday. No spam. Unsubscribe any time.
Partner picks
Build a smarter digital stack
Explore curated AI, automation, wealth, and creator tools selected for practical value, transparent pricing, and clear use cases.
Disclosure: some links may be affiliate links. DigitechLifestyle may earn a commission at no additional cost to you.
