In February 2025, Britain renamed its AI Safety Institute. It became the AI Security Institute: national security and crime in, bias and free speech out. The word "safety" had lasted barely a year. A month earlier, Washington had revoked the previous president's AI executive order on its first day, replacing it with one titled Removing Barriers to American Leadership in AI. In Brussels, the EU's flagship AI law was already being pushed back.
This is AI governance in 2026: a lot of motion, very little that sticks - and underneath it all, one job done by nobody: checking whether what the machine says is true.
I know, because my co-founder Riya Nalwa and I spent weeks doing that job. We checked 194 citations across 30 AI-generated research briefings - open the link, read the claim, read the source, ask whether the source supports the claim. 95 were broken: the page had moved, or the source never existed. 50 supported the claim. 17 pointed to real sources that said something different. 31 could not be checked at all. 1 had no link. We published every citation and every verdict, so anyone can repeat the check.
Advertisement
That is not a story about bad writing - the briefings were well written. It is a story about a missing layer. Every briefing was produced by a system that had never opened a source, and no rule required anyone to check before it was used. The verification layer - the person who clicks the link - is the one part no regime has made mandatory. We treat it as a courtesy. It is the whole system.
The patchwork
Look at who governs AI today: not a system, a patchwork, and most of the patches are paper.
Europe wrote the most ambitious AI law in the world, then spent a year delaying it. General-purpose AI obligations took effect in August 2025, but the high-risk rules - the ones that bite on hiring, policing, benefits and borders - were pushed to December 2027 by the Digital Omnibus. The fines are still enormous on paper, up to €35 million or 7% of global turnover.
America has decided the states will do it. The federal government revoked its AI safety framework and replaced it with nothing that checks a model's claims. California did
instead: in September 2025 it signed SB 53, the first American law for frontier AI developers, requiring the largest developers to publish safety frameworks, report critical safety incidents, and protect whistleblowers. It is a real law - and a confession: the country that builds the models cannot pass a national rule about them, so one state is doing the job for fifty.
India took a different route. In November 2025 its electronics ministry released the India AI Governance Guidelines - seven principles, called sutras, and six governance pillars. Thoughtful, and by design not a law: it relies on existing statutes and voluntary measures.
Advertisement
Then there is the one treaty. The Council of Europe's Framework Convention on AI is the first legally binding international AI treaty ever written. It opened for signature in September 2024. Two years later it has one ratification - the European Union's - and twenty signatories, and needs five ratifications to enter into force. The world has a treaty on AI - almost nobody has signed it.
No rule in the patchwork requires anyone to open a citation and check whether it supports the claim attached to it. The laws ask for frameworks, incident reports, labels, principles - not proof. A developer can publish a safety framework and still ship a briefing in which half the citations lead nowhere - no rule makes the check a condition of the work.
The capacity gap
That gap would matter less if the people using these systems had time to check them - they do not. In 2024, 78% of organisations reported using AI, up from 55% the year before, and in the United States the number of state AI laws passed in a single year jumped to 131. Adoption is compounding. Governance is fragmenting.
Now put that next to our audit. The tools spread fastest where the staff to check them is thinnest. A well-funded lab can hire someone to verify; a small university, a ministry, a local council often cannot. The same AI, the same error rate - only one side has anyone free to click the links. Generating a confident briefing costs almost nothing. Verifying it costs a person's time. That asymmetry is the whole problem.
It is about a clerk in a ministry opening a briefing that will shape a decision, with no way to know whether its footnotes were ever checked. That is not a futuristic problem. It is a Tuesday problem.
What a regime that works would look like
So what would a regime that actually works look like? Not another framework, and not another summit. Four things, all boring, all checkable.
First, verification as a gate, not a report. The check has to happen where the decision is made, not in a dashboard reviewed once a quarter - a step apart from the workflow will not catch an agent acting this afternoon. The check has to be a gate the work passes through: take each claim, open the citation, give a verdict - verified, broken, mismatch, unverifiable - and publish the verdicts with the work, so the reader can see them and repeat them. That is what we did with our 194 citations. It takes minutes per briefing. And it is the one practice no current AI law requires.
Second, disclosure that names the check. California got this half right: it makes the largest developers publish safety frameworks and report critical safety incidents. China went further on one thing - since September 2025, AI-generated content must carry a visible label and embedded metadata. A label is not a check, but it is honest about what you are looking at. The next step is the same honesty about verification: not just "this was made by AI," but "this was checked, here are the verdicts, here is what could not be established."
Third, an institution that cannot be renamed away. The British institute lasted a year under one name before its remit was narrowed. The American one was revoked on a president's first day. A body whose mandate depends on who won the last election is not a governance body; it is a press release with staff. Verification needs an institution with a mandate written into law, a budget that survives a change of government, and a job description that names the check: it verifies claims, publishes its verdicts, and cannot be quietly re-scoped when the politics change.
Fourth, liability that lands on the deployer. When a briefing built on broken citations shapes a bad decision, nobody is responsible. The provider points to its terms of service. The user points to the model. The decision-maker points to the briefing. A working regime would close that circle: if you deploy an AI system to produce research that informs decisions, you are responsible for whether its citations resolve and support its claims. Not for the model being perfect. For the check being done. That is a duty a court can enforce and a regulator can audit.
None of these four is radical. What is missing is a regime that puts them together and makes them mandatory - verification as infrastructure, like roads or clean water, rather than a virtue.
What counts as finished
AI will keep writing research, and keep getting better at sounding right. That is not a reason to stop using it. It is a reason to change what counts as finished - and to write that change into law.
A briefing is not finished when the prose reads well. It is finished when every material claim has a source that resolves and someone opened it. A decision is not ready when the briefing is thick with citations. It is ready when the citations have been checked and the verdicts are visible. A regime is not serious when it has published a framework. It is serious when it can show you the checks, the verdicts, and the institution doing them after the election.
We did our audit with a browser and a spreadsheet. It found that nearly half the citations in thirty briefings led nowhere - what one small team can find, in weeks, with no budget and no legal power. The machines will keep making promises. Governance is the job of making them keep one. Until a rule requires the check, the footnote is still just a promise, and nobody is in charge of keeping it.