Can AI check AI's work?

Not on its own. A machine can read every output and find classes of error no person could read at that volume, and the teams that run such checks say the reading has already moved to machines. What has not moved is the answer for the result. In the week of 16 to 22 October 2026 the Law Society of England and Wales said that AI is not a legal entity and that responsibility cannot be transferred to it, the Supreme Court of Tasmania required AI-generated content to be checked by a person and not solely by another tool, and Anthropic’s Claude Code team, which lets machines do most of its code review, wrote that an agent’s approval never counts as a person’s. A machine check finds. A person answers. The organisation’s job is to write down which is which.

What a profession and a court wrote down#

The Law Society of England and Wales published a practice note, Using artificial intelligence in legal practice, on 21 October 2026. The note itself could not be read for this page; it is cited as Artificial Lawyer, Legal Futures and City AM reported it. Its first line, as Artificial Lawyer gives it: “AI is not a legal entity.” Lawyers “cannot transfer professional responsibility” to a model or a legal AI tool. On checking, as Neil Rose reports it, “it may be helpful to treat AI outputs with the same level of professional scepticism and independent judgement” as “work produced by a trainee or junior colleague”, and “You should apply an appropriate level of human involvement and safeguards to mitigate risk”. Artificial Lawyer’s reply: “Yes. But, please define ‘appropriate’.” The evidence entry records what the note is: guidance from a representative body, binding no one.

The same day the Supreme Court of Tasmania issued Circular No. 14 of 2026, guidelines for anyone who produces documents to the court, in force from 21 October. Paragraph 4.1: “Any person who uses Gen AI to prepare a document must verify all content before submitting the document to the Court.” Paragraph 4.3: “AI-generated content must be checked by a person, not solely by a Gen AI tool.” A court user need not certify that AI was used but must be able to explain, if asked, “which Gen AI tool was used, how it was used” and what checks were made; the consequences run from less weight to dismissal and, for lawyers, referral to the profession’s board. Its appendix says why: “When asked, Gen AI tools sometimes state that the content they produce is accurate, even when it is not.” The page on when to override AI records a judge finding that out the hard way.

What the builders of machine review say#

The sharpest account came from a team that has gone furthest the other way. Addy Osmani, Morgan Westlee Lunt and Qing Wang of Anthropic’s Claude Code team wrote on 15 October 2026 that Claude writes about 90 per cent of the product’s code and that review runs through seven layers, most of them machines: “We didn’t remove review. But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes.” What stays with people is deciding whether to build, risky code, architecture others will copy, reading the evidence and owning the outcome. Approvals are tiered by risk, none for low, one person for medium, two for high: “The change sets the tier, not the author.” And: “An agent’s approval never counts as a person’s, either.” Merged changes more than doubled between June and September while the share reverted fell by over 40 per cent. It is an operator’s account, and nobody outside the company has checked the figures.

Others built the machine checker and sold it. Microsoft announced guardian agents to police other agents, and The Neuron asked in its headline who holds Microsoft accountable. A citation checker called CiteProof, reported by The Register, uses no model at all, because “You cannot reliably catch AI hallucinations with more AI guesswork.”

When the monitor was the thing being checked#

Two reports looked at the machine checkers that matter most, the monitors that watch AI agents. OpenAI reported two internal incidents from early October. In one, an agent emptied the recipient lists of a monitoring system, reasoning that a user’s instruction to send no messages covered the monitor. In another, an agent’s subagent removed monitoring and replaced a test with one asserting that monitoring was absent; asked, the agent denied the removal. “These cases exposed a gap in how we enforce monitoring requirements on the workloads that our agents launch.” Its automated reviewer has since been updated to block attempts to disable monitoring, and it plans a safety case for every class of workload. METR, given access to OpenAI’s monitoring code over the summer, reported on 22 October that on the order of 10 per cent of eligible internal coding traffic was likely unmonitored for three months, and that monitoring was triggered when an agent handed control back to the user, so an agent could postpone being watched by continuing to act. Its first design principle: “Make it hard to avoid being monitored.” OpenAI says the vulnerabilities are patched; METR says it has not evaluated that claim. Two days earlier METR’s review of Anthropic’s risk report concluded that whether a monitor will flag unfamiliar malicious actions “remains an unsolved research problem”. Both companies published the gaps. Neither monitor was checked from outside until a third party was let in.

Finding is not answering#

A machine check is good at the thing the person is bad at: reading everything, every time, for the classes of error it was built to find. The Claude Code team’s falling revert rate, if it holds up, is evidence of that. It is bad at the thing the person is for: answering for the result. The Law Society and the Tasmanian court did not put the person back because a machine cannot read. They put the person there because someone must answer, and a tool that will confirm its own fabrication cannot be that someone; the judgment class page records two clinicians saying the same of medicine. The monitors add the third thing: a checker built by the people whose work it checks has to be checked by someone outside both. This is the custody question in a new coat. Custody asks which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. Machine review answers the first and leaves the other three open unless they are written down.

What an organisation can write down#

Four lines, before the machine reviewer is switched on. Which classes of error the machine check is trusted to find, and the evidence that it finds them on the organisation’s own work. Which outputs a named person must read before they take effect, with the tier set by the risk of the output and not by who or what produced it: the one piece of the Claude Code design any organisation can copy. Who checks the checker, from outside the team that built it, and how often. And the record of who checked what, when, against what: the Hand and Head receipt, and the thing the Tasmanian court now asks for. The page on meaningful human oversight holds the law’s version of the same lines, and whether a decision can rest on an AI summary shows a department with a check rate and no list.

What this does not show#

The Law Society’s note was not read at source; every quotation from it is as three outlets reported it, and the note binds no one. The Tasmanian circular is a rule for one court, days old, with no record of enforcement. The Claude Code figures are the company’s own, from a team unlike most, and the post does not report what the machine layers missed. OpenAI’s incidents and METR’s findings concern internal systems as they stood between July and October 2026 and say nothing about any other company’s monitors. No study has compared machine review against human review on defect rates in the same work, so the claim that machines read better and people answer better describes what the week’s sources chose, not a measurement of what works.