Regulation runs on evidence. A financial regulator reads the books. A medicines regulator reads the trial data. A rail safety regulator reads the safety case and the incident records. In each case the regulator can test what it is told against something the firm could not quietly change.
AI regulation now has the law. It does not yet have that evidence.
What regulators want
Talk to anyone who supervises AI, or will soon, and the same needs come up:
- Proof that a duty was met, not a statement that it was.
- Evidence they can check without depending on the firm's cooperation, and without having to see its trade secrets or its customers' personal data.
- The ability to rebuild what happened after an incident, from a record made at the time.
- Comparable evidence across firms, so that one bank's AI can be judged by the same yardstick as another's.
- Supervision at the speed and scale AI now runs at. A regulator cannot read every decision a system makes. It needs to sample and test, and trust that what it tests is what actually ran.
- Evidence that crosses borders, because AI systems and the firms that run them do.
What regulators have
Quite a lot, on paper.
The law. The EU AI Act is in force. It requires high-risk AI systems to record events automatically, so that their operation can be traced. Following this year's amendments, the high-risk obligations apply from 2 December 2027 for stand-alone systems and from 2 August 2028 for AI built into regulated products. Duties to mark AI-generated content apply from 2 December 2026. Alongside it, the EU's new Product Liability Directive brings software, including AI, into product liability from December 2026.
Standards, still being finished. The harmonised standards that show firms how to comply are being written by the European standards bodies, CEN and CENELEC. They are behind schedule, and in October 2025 the standards bodies adopted exceptional measures to speed them up. Until they are cited in the Official Journal, firms do not get the presumption of conformity that comes with following them.
Self-assessment. For most high-risk uses listed in the Act, the provider assesses its own system's conformity. An outside notified body is required only in a narrow set of cases.
Management-system certification. ISO/IEC 42001 certifies that a firm has a system for managing AI. It is valuable. But a certificate that a process exists is not evidence of what a particular system did on a particular day.
Codes of practice and voluntary frameworks. The EU code on marking AI-generated content, the US NIST AI Risk Management Framework, and many sector guidelines.
Sandboxes. Spain's AI supervisory agency, AESIA, Singapore's AI Verify Foundation and the UK all run programmes where firms can test AI under supervision. The deadline for every EU member state to have one was moved to August 2027.
China has gone furthest on registration and labelling. A generative AI service offered to the public must pass a security assessment against a national standard and a review by the Cyberspace Administration of China before it launches. By the end of August 2026 the regulator had recorded 1,112 filed services. Since 1 September 2025, AI-generated text, images, audio and video must carry both a visible label and a label embedded in the file, under a mandatory national standard (GB 45438-2025). In June 2026 the financial regulator issued its first dedicated AI rules for banks and insurers.
India has chosen principles and sector regulators. The India AI Governance Guidelines of November 2025 are voluntary by design. They rely on existing law, industry codes and self-certification, and leave sector regulators to set binding rules. The Reserve Bank's FREE-AI framework of August 2025 does that for finance. Since 20 February 2026, amended IT Rules require platforms to label AI-generated audio and video prominently.
Elsewhere the range is just as wide. South Korea's AI Basic Act, in force since January 2026, puts risk duties on operators of high-impact AI, much as the EU does. Japan's AI Promotion Act of 2025 takes the opposite line: voluntary guidance and no penalties. Singapore combines a public testing toolkit, AI Verify, with binding expectations for finance: the Monetary Authority's AI risk management guidelines of October 2026 take effect in October 2027. In the United States, most rules come from the states and from sector regulators such as the insurance commissioners.
These approaches differ sharply, from strict law to soft guidance. They share one feature: the evidence that a system behaved as claimed is produced by the firm, and either kept by the firm or filed with one regulator. In China the regulator does review it, but the review is not something a buyer, an insurer or another country's regulator can see or rerun. Outside researchers have also found that later model upgrades cannot be traced through the public registry.
International bodies. Several are working on parts of this:
- the ISO and IEC committee on AI (SC 42), where China and India are both active, India having hosted a plenary in New Delhi;
- a joint project between that committee and ISO's conformity-assessment committee on how to build schemes for checking AI systems;
- the International Network for Advanced AI Measurement, Evaluation and Science, where government AI testing bodies compare methods (it was the International Network of AI Safety Institutes until 2026);
- the UN's Independent International Scientific Panel on AI, which reported to the first Global Dialogue on AI Governance in Geneva in July 2026;
- the internet's own standards bodies, the IETF and W3C.
Each is doing real work, on principles, management standards, evaluation methods or scientific assessment. None yet produces evidence about a specific AI decision that a regulator in another country would accept.
What regulators are missing
Set the list of wants against the list of tools, and five gaps appear.
1. A common form of evidence. Each firm logs in its own way, in its own systems, in its own format. A regulator supervising fifty firms faces fifty kinds of record. There is no shared shape for "this is what the system did, and this is what it rested on".
2. Evidence the firm cannot rewrite. Under self-assessment, the documents that show compliance are written and held by the firm being supervised. That is how much financial and product regulation once worked, and every sector that moved past it did so because a scandal showed why it had to. Logs held inside the firm's own systems can be complete and honest. But the regulator has no way to know that, short of trusting the firm.
3. Checking at machine speed. Regulators are not staffed to read AI decisions one by one, and they never will be. What they need is evidence that can be checked automatically, so that people spend their judgement on the cases that need it.
4. Knowing that what was tested is what ran. A system can be assessed, certified or sandboxed in one form and run in production in another. Models are updated. Prompts change. Unless anyone can tell which version is running, a certificate describes the past.
5. Recognition across borders. The EU, the UK, the US, Singapore, Japan, South Korea, China and India are all writing AI rules. Firms operate across all of them. Evidence produced for one regulator is rarely usable by another, so the same system is assessed again and again, and none of the assessments can be compared. The closest thing to a shared testing table, the International Network for Advanced AI Measurement, Evaluation and Science, has ten members: Australia, Canada, the EU, France, Japan, Kenya, South Korea, Singapore, the UK and the US. Most of the world's people live outside those members, including everyone in China, India, Brazil and Indonesia.
The quality world solved this problem once already. Under the ILAC and IAF mutual recognition arrangements, a test report from an accredited laboratory in one member economy is accepted in the others. China's accreditation body, CNAS, and India's, NABL, have both been signatories since 2000. AI has nothing equivalent yet.
Why it matters: we have seen this before
The cases that shook public trust in automated decisions share one failure: the people affected, and often the regulators, could not check what the system had decided or why.
In the UK, hundreds of sub-postmasters were prosecuted on the evidence of an accounting system that the courts presumed reliable and the accused could not test. In the Netherlands, an algorithm wrongly flagged families for childcare-benefit fraud, and the government resigned over it in January 2021. In Australia, a Royal Commission condemned the Robodebt scheme of automated welfare debts in 2023.
None of these involved today's AI. All of them involved decisions made at scale by a system nobody outside could check. AI makes that pattern faster and more common, not less.
The AI-era version is already in the courts, and it looks the same everywhere:
- New York, 2023: two lawyers were sanctioned for a brief citing cases a chatbot had invented.
- London, 2025: the High Court referred lawyers to their regulators over fictitious citations.
- Singapore, 2025: a lawyer was ordered to pay costs personally for citing a case an AI tool had made up.
- New Delhi, 2026: the Supreme Court set aside a customs penalty of about ₹4.3 billion (US$45 million) because the order relied on case law that did not exist. It then ruled that citing AI-generated judgments without checking them is misconduct, and asked the Bar Council of India to set rules.
In each case the problem was found only because someone checked the source by hand.
What good would look like
Regulated sectors have already worked out what makes a record count as evidence. Banking, medicines, aviation, legal evidence and quality management each set out the same core properties in their own rules. A record that a regulator can rely on:
- was made at the moment, not reconstructed later;
- shows who or what acted;
- is time-stamped by a clock the firm does not control;
- is complete, so that gaps show rather than hide;
- is never edited, only superseded, so that changes are visible;
- can be checked by someone outside the firm without access to its systems;
- can be exported to the regulator and frozen for a dispute;
- keeps personal data minimal and erasable.
None of this needs new law. It needs the evidence AI systems leave behind to meet the standards regulators already apply to everything else.
Questions a regulator could ask any firm deploying AI
- When your system made this decision, what record was made at the time, and who holds it?
- Could anyone in your firm change that record without it showing?
- Can we check the record ourselves, without your staff, your systems or your goodwill?
- Is the system you assessed the system that is running today, and how would we know?
- If a source your system relied on turns out to be wrong, how do you find every decision built on it?