Can an employer use AI to discipline or dismiss someone?

Not on the machine’s word alone, in a growing number of places. On 30 September 2026 the Governor of California signed SB 947, the No Robo Bosses Act, which from 1 July 2027 bars employers from relying solely on an automated system to discipline or dismiss a worker. In the United Kingdom the law has said something close to that since 5 February 2026: a significant decision taken with no meaningful human involvement must come with the right to be told, to answer back, to reach a person and to contest it. The same week the Guardian reported a UK training company whose AI now scores its teachers for several hours a day. Each case turns on one thing the statutes leave to the employer: what the human reviewer does, and whether anyone would know if they did nothing.

What California signed on 30 September#

The Governor’s release lists thirteen bills, among them SB 947, “Employment: automated decision systems”, and its summary of the worker protections includes “prohibiting employers from only relying on AI when making a disciplinary action or termination decision”. Senator Jerry McNerney’s own release says the law “mandates human oversight and verification when employers use ADS to assist in termination and disciplinary decisions”, requires employers to tell workers when such a system has been used, and would be enforced by the labor commissioner, the attorney general or local prosecutors. His sentence is the one that named the act: “No worker should ever be fired or disciplined by a robo boss.” Danielle Ochs and Zachary Zagger at Ogletree give the date, 1 July 2027, note that it is “a newer version of a similar law that Governor Newsom vetoed in 2025”, and supply the operative phrase: the employer needs “a human reviewer to independently corroborate the ADS output”. Ryan Golden at HR Dive reports what the reviewer corroborates with: “gathered evaluations, personnel files and other documents”. Two companion laws were signed with it. SB 951 requires notice of mass layoffs “caused in whole or in substantial part by AI or similar systems”, and AB 1883 bars workplace tools that infer a person’s emotional state or collect neural data.

The same week, a UK firm whose AI rates its teachers#

Robert Booth reported in the Guardian on 28 September, read here in AOL’s syndication, that instructors at Multiverse, an apprenticeship training company, now have transcripts of their online classes “analysed and scored by an AI programmed to alert human managers when they are suspected of doing things wrong”. The system tags each instructor with a “risk status” and a “percentage confidence score”. The change in scale is in one sentence: “Previously instructors were observed monthly by a human manager but are now watched for several hours a day by an AI.” One teacher, speaking anonymously, said “the stress level has been absolutely horrendous”; another, “I feel like I’m under continuous scrutiny”. Multiverse told the paper that the AI “directs human time to where it’s needed most” and that only human managers write performance reviews. Its spokesperson added: “The most consequential performance management action it can take is to recommend that a human personally reviews a session.” Both accounts can be accurate. A system that decides nothing can still change what people do, and the teachers describe attending to the scorer when they used to attend to the learner. Managers, they said, had pulled them up for suspected digressions that were answers to a student’s question.

What UK law already requires#

Section 80 of the Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with Articles 22A to 22D, fully in force since 5 February 2026. Article 22A says a decision is based solely on automated processing “if there is no meaningful human involvement in the taking of the decision”, and is significant if it has a legal or similarly significant effect on the person. Where both are true, Article 22C requires safeguards that tell the person about the decision and let them make representations, obtain human intervention and contest it. A dismissal on a score alone would meet both tests. A manager’s review moves the decision outside those articles only if the involvement is meaningful, and the Act lets ministers define by regulation what counts. As the consultancy Resultsense put it in a note on the Multiverse report, meaningful involvement takes a decision outside the automated-decision rules and “does not take it outside data protection law”. The regulator’s guidance on monitoring workers still applies: workers must be told how and what information is collected. The estate’s fuller treatment is on what meaningful human oversight is.

The consultation that closed on 30 September#

The government’s consultation on workplace monitoring technologies ran from 8 July to 30 September 2026 and is now being analysed. It defines the technologies as “digital tools used by employers to collect, track, analyse or make decisions based on information about workers and their activities”, and reports “a 2025 survey of UK managers finding that 1 in 3 organisations actively monitor an employee’s digital activity, up from 1 in 5 employers in ICO research in 2023”; the survey is not named. Its sixth principle, human oversight and accountability, asks that “workers are able to question and, where appropriate, challenge decisions that affect them in a timely way”, and gives an example that fits the Multiverse case closely: “where automated systems flag potential underperformance, a manager should review the context before acting”. Three routes are offered: a statutory code that tribunals could take into account, with discretion to adjust compensation by up to 25 per cent; a legal requirement to consult and negotiate over plans to adopt the technology; and non-statutory guidance. The document says the options do “not indicate a settled government preference”.

What a human review has to be#

California’s word is corroborate and the UK’s is meaningful, and both can be satisfied on paper by a reviewer who adds nothing. The evidence on automation bias says a person shown a score first tends to follow it, and the estate has set out why a human in the loop is not a safeguard by itself. Four facts separate a review from a signature. The reviewer works from records the system did not produce, which is what the California drafting reaches for with personnel files and evaluations. The reviewer can disagree without cost, and has the authority to override. The reviewer can still do the judging themselves: the manager who used to sit in a class each month is the person who can tell a digression from an answer, and that skill fades when the score arrives first. And the disagreements are counted. If reviewers reversed the system in none of last quarter’s cases, that is a finding about the review.

The decisions an employer controls#

None of this waits for July 2027 or for the consultation response. Rules Before Tools is the pattern. Which decisions about people may a machine make here: it may flag and sort, and a decision that ends or damages someone’s employment is made by a named person. Who can stop each one: the reviewer, and above them someone who can switch the scoring off for a team without having to justify the cost. What people must remain able to do: managers observe and judge some work with no score in front of them, on a schedule, so the skill the review depends on survives. And how anyone would know if it went wrong: the reversal rate, the appeals, and what staff say the system has done to their work, asked before the stress shows up as absence. The same four questions apply to screening applicants and to performance reviews; the measurable risk sits in the handover of the decision and in what happens to the manager’s judgement afterwards.

What this does not show#

The text of SB 947 was not read; the account here rests on two official releases, a law firm’s note and a trade report, and the effective date comes from the last two. No tribunal or court decision under Articles 22A to 22D is cited because none was found. The Guardian’s report is anonymous testimony from staff at one company, which disputes the framing; nobody has measured the stress or the quality of teaching before and after. The consultation’s one-in-three figure comes from a survey it does not name, and the consultation has produced no decision. Nothing here shows that a human review improves a disciplinary decision. The evidence on reviewers who see the score first points the other way, and no employer has published its reversal rate.