Since 5 February 2026, a controller in the United Kingdom must enable a data subject to obtain human intervention in a significant decision based entirely or partly on automated processing. That is Article 22C of the UK GDPR, as substituted by section 80 of the Data (Use and Access) Act 2025. Across Articles 22A to 22D there is not one occurrence of record, records, log, logs, logging, retain, retention, document, audit or evidence.

The Information Commissioner's Office, interpreting those same articles, says something the articles do not: “You should keep a record of how the human reviewed the decision.”

What can a deployer in the United Kingdom be asked to produce today about a human review, by whom, and on what authority?

What the reading found

Expected rather than required, and the difference is the whole page. Guidance requires nothing of anybody. A deployer who ignores it is not breaking a stated rule, and is departing from what the enforcing authority has published about how it reads the law. Scoring that as required would be the overclaim this site exists to catch in other people's documents.

The statute gives a right and asks for nothing

Article 22B restricts significant decisions based entirely or partly on automated processing. Article 22C says that where such a decision is taken, the controller must have safeguards in place, and enumerates four: provide the data subject with information about the decision, “enable the data subject to obtain human intervention on the part of the controller”, enable them to make representations, and enable them to contest it.

Four events, every one of which a record could carry, and the statute asks for none of them. Whether a particular person obtained intervention, from whom, what that person was shown, and whether the decision changed, are all questions the text leaves to whatever the controller happened to build.

The definition is where it bites hardest. Article 22A(1)(a): “A decision is based solely on automated processing if there is no meaningful human involvement in the taking of the decision.” The restriction is escaped by asserting meaningful human involvement, and nothing in the text requires evidence of the involvement being asserted. The phrase appears eight times across the substituted articles. Article 22A(2) names one factor for deciding it, the extent to which the decision was reached by profiling, and leaves the list open. Article 22D then lets the Secretary of State settle by regulations whether there is or is not meaningful human involvement in described cases, so the central term of the regime is deliberately unfinished.

The regulator asks for the record the statute does not

The ICO's guidance on the automated decision-making safeguards is careful about the difference between the two words it uses. It says must nineteen times, for what the law obliges, and should thirteen times, for what the Commissioner expects. The record is a should:

“You should keep a record of how the human reviewed the decision.”

It is the only occurrence of the word record in the guidance, and there is no equivalent anywhere in the articles it interprets. The guidance also asks, once, for “a system that is able to deliver an audit trail showing the key decision points that formed the basis for the decision”. Those two sentences are the entire documentary expectation, and neither has a statutory provision behind it.

The guidance is more specific about the reviewer than the statute is. Human reviewers should assess and review any reconsideration before it is applied, have the ability to influence the outcome, have “discretion and authority to alter the decision”, be “suitably trained and qualified to understand the system's outputs, limitations, and risks”, and take into account the relevant data and factors the decision was based on. It adds that, like human involvement, human intervention “cannot be tokenistic”.

Five properties of a person, and a record of how they reviewed. Two of those five are facts about an employment relationship rather than about any decision: training and qualification live in HR and identity systems and no record format produces them. The other three are observable, and a record can carry them if it was built to.

What a record would have to carry, and what does not

Taking the guidance at its word, a record that answered it would have to say who intervened, where that identity was resolved from, what material the reviewer was actually shown as distinct from what the decision cited, and whether they approved, modified or overrode the output. The last is the one to check first, because a reviewer who changed an action before allowing it and one who approved it unchanged must not produce the same record.

Of ten widely deployed agent frameworks and memory systems read at pinned commits, eight take or gate consequential actions and one can identify the person who approved one, and that one is this project's own reference implementation, which is stated here rather than left to be found. Nothing on the approval path in the others identifies a person at all. The per-system verdicts, each citing a file and a line, are in the conformance register, and what a reviewer was shown is read separately at approval binding.

So the gap a United Kingdom deployer has is not a policy gap. The regulator has published what it expects, and the software the sector runs on cannot emit it.

The assurance market assures the model

The Department for Science, Innovation and Technology publishes a Portfolio of AI Assurance Techniques. It is the United Kingdom government's own curated answer to the question of who can assure an AI system, and on 12 September 2026 it held 75 case studies. Every one was read here against a single question: does the technique claim to produce, check or assure a record of a human review of a decision, which is the thing the Commissioner says a deployer should keep.

The method, because the number rests on it. The first stage is mechanical and decides nothing: a case study is a candidate only if it uses a word from each of two fixed vocabularies, one for a person in the loop and one for something kept. That bounds the reading to a set anybody can reproduce from the committed corpus. Every one of the fourteen candidates was then read, and the verdict cites the sentence it rests on. A word count cannot tell a record of a human review from a record of a model, and treating it as though it could would be the error this site exists to catch elsewhere.

What the fourteen records are about instead

The approvals that are recorded are real approvals. They are just of something else: a control signed off, a model promoted, a supplier cleared, an assessment completed. Between the approval of the system and the approval of one decision the system made lies the whole of what the Commissioner asked for, and nothing in the portfolio is on that side of it.

The closest is worth naming because it shows the gap rather than hiding it. Anekanta's facial recognition assessment states that “a trained operator should always make the final decision before any action is taken following a face match”, and says plainly that the assurance system “does not evaluate competency”. So a human decision is required, the competence behind it is out of scope by design, and what the technique produces is a privacy impact report rather than a record that the operator decided anything.

This is not a criticism of any technique in the portfolio. A method for measuring disparate impact is not failing to record a decision, in the same way that a vector store is not failing to authorise an action. Each of these is doing the job it was built for. The finding is about the shape of the market as a whole: the United Kingdom has a live regulator expectation about a record of a human review, a government portfolio of 75 ways to assure AI, and no overlap between them.

The same specification, reached from the opposite direction

Colorado's proposed Rule 7.7, which would take effect on 1 January 2027, asks for substantially the list above: a reviewer with authority to alter the output, trained and competent, who took the relevant evidence into account, and a record of it. One jurisdiction is writing it into a rule and the other is already stating it as an expectation on a statute that requires none of it. Neither of them arrived at the other's text. That is read in full at the Colorado reading.

Where this sits in time

The United Kingdom is neither the first nor the last to give a person a route to a human, and the ordering is worth having straight, because every one of these gave the right and asked for no record of it.

Eight years between the first of those and the first one that asks for the record, and the thing that asks for it today is guidance rather than law.

What this does not show

Two instruments, read against what they ask a record to show. Not an assessment of whether either is good at what it is for, and not legal advice. The guidance is doing a job the statute left it, and doing it more precisely than the statute did.

Several things a United Kingdom deployer is subject to are outside this reading and are named rather than implied. The rest of the Data (Use and Access) Act 2025 was not read. The ICO's wider guidance on artificial intelligence was not read; only the automated decision-making safeguards page was. ISO/IEC 42001 has no verdict anywhere on this site because its text is paywalled, which is recorded as a fact about the field rather than scored. A United Kingdom firm placing a system on the European Union market is separately subject to the EU AI Act, which is read at its own page.

Check it

Both texts are committed in census/sources/ beside this reading: the substituted articles as enacted, and the guidance as published. All 75 portfolio case studies are committed too, as census/sources/dsit-portfolio.jsonl, one object per case study with its path, title and text as fetched on 12 September 2026, and the count is produced by census/dsit_portfolio.py, which prints the sentence behind every verdict when run with --read. Every count on this page is recomputed from those files by the test suite rather than written down once and trusted, so a reader who thinks a number is wrong can settle it against the same bytes this page was written from.

The portfolio material contains public sector information licensed under the Open Government Licence v3.0.

A reading is wrong in the ordinary way readings are wrong, which is why every statement above names the article or the sentence it rests on.

Citing this

The readings are CC BY 4.0, which asks for attribution, so the reference is here rather than left to be composed. This block is generated from the page it sits on, so a date that moves here moves in the citation too.

Clifford, T. (2026). The regulator asks for the record. The law does not. Machine Testimony. https://machinetestimony.org/united-kingdom/

@misc{clifford2026unitedkingdom,
  author = {Clifford, Troy},
  title  = {The regulator asks for the record. The law does not},
  year   = {2026},
  note   = {Machine Testimony, read 12 September 2026},
  url    = {https://machinetestimony.org/united-kingdom/},
}

This page carries no DOI. It cites its dated URL, and saying so is the point: a citation naming a deposit that does not exist is worse than one naming a page that does.