DR PETER FIELDS • AI & IDEAS

Blog

Short comments by Peter Fields on current developments in AI.

  • A regulator needs answers and someone who can test them

    On 14 October, European Commission Executive Vice-President Henna Virkkunen said the AI Office had begun sending requests for information on matters including security, external evaluation and monitoring. Britain’s 28 October Budget committed an additional £120 million to the AI Security Institute over four financial years to strengthen testing and evaluation.

    These measures do different jobs. A formal information request can make a developer explain what it did and what it knew. Public testing capacity gives officials a better chance of judging whether the answer is adequate. In Who Benefits from Regulating AI?, I ask how a regulator can avoid becoming dependent on the firms it supervises. An information power matters more when the person using it knows which records to ask for and has the expertise to challenge an incomplete answer.

    Sources European Commission speech, 14 October · HM Treasury, Budget 2026

    Related Notebook articles Who Benefits from Regulating AI?

    Retrospective commentary on the development of 28 October 2026.

  • Who pays for the work that trains a junior lawyer?

    The Law Society’s 21 October practice note on AI in legal practice puts responsibility with the solicitor and stresses human oversight. Its guidance also raises a question about junior lawyers. If AI does the first pass, who teaches the person checking it how to spot what has gone wrong? The SRA’s separate consultation, opened on 8 October, asks how the competence expected of solicitors should change.

    A training partner can explain why a fluent answer rests on the wrong assumption. That takes time. If a firm once recovered some of the cost of training through first-pass work and now does less of it, the time for supervision still has to come from somewhere. This is the question I raised in Who Will Train Tomorrow’s Experts? I would like to see firms say which tasks juniors will still do themselves, which they will do with assistance, and who has time set aside to explain the difference.

    Sources Law Society practice note · Law Society announcement, 21 October · SRA consultation announcement, 8 October

    Related Notebook articles Who Will Train Tomorrow’s Experts? · After the Billable Hour

    Retrospective commentary on the development of 21 October 2026.

  • The judiciary draws a line around AI use

    The judiciary’s Position Statement of 9 October describes how it is testing and using AI, starting with lower risk uses and controlled pilots. Judicial office holders remain responsible for material produced in their name. That sentence matters because a task can look administrative until part of its output informs a decision.

    Consider a summary that leaves out the one passage on which a party’s case depends. The summary may have saved time, but it has also changed what the judge sees. In Can a Machine Judge?, I ask where assistance ends and the exercise of judgement begins. The useful safeguard is to make sure the person deciding can get back to the record and has time to read it.

    Sources Courts and Tribunals Judiciary, Position Statement of 9 October

    Related Notebook articles Can a Machine Judge?

    Retrospective commentary on the development of 9 October 2026.

  • How can an outside auditor hear bad news?

    Three former OpenAI safety researchers, Tomek Korbak, Jasmine Wang and Mikita Balesni, disputed the grounds for their dismissal in an open letter. OpenAI said, in its response reported on 9 October, that the issue was a breach of trust and that the dismissals were unrelated to their having raised safety concerns. The accounts differ. The public material does not establish whether a protected disclosure was made or whether any dismissal was justified.

    In Can We Afford to Slow Down?, I discuss proposals to bring independent evaluators into the testing of advanced systems. For that arrangement to work, everyone needs to know in advance what staff can show an auditor, who authorises it and how a concern can be escalated if the usual route fails. An auditor who sees only what management selects may miss the very problem it was appointed to find. A company can protect sensitive material while giving staff a credible way to raise concerns. The procedure is what makes those two aims compatible.

    Sources Korbak, Wang and Balesni’s open letter · Associated Press account including OpenAI’s response, 9 October

    Related Notebook articles Can We Afford to Slow Down?

    Retrospective commentary on the development of 9 October 2026.

  • What is a rule against cruelty to a chatbot for?

    Anthropic’s 8 October policy announcement introduced a restriction on sustained, needless abuse or cruelty towards its models, effective 12 November. The company says the policy leaves room for ordinary frustration, criticism, creative work and legitimate testing. Its stated reasoning acknowledges uncertainty about whether a model can experience harm.

    In Could We Wrong an AI?, I ask what would give a machine interests of its own. There are other reasons a company might discourage cruelty in an interaction. It might be concerned about habits people take elsewhere, or about how a model responds to prolonged abuse. Those reasons would justify different limits and demand different evidence. A company can adopt a precaution while the science remains unsettled. Readers should still ask whose interests the rule is meant to protect, and whether the rule is drawn tightly enough to do that job.

    Sources Anthropic’s 8 October usage-policy announcement · Usage Policy effective 12 November 2026

    Related Notebook articles Could We Wrong an AI? · Is There Anyone Behind the Conversation?

    Retrospective commentary on the development of 8 October 2026.

  • Whose words did the court hear?

    The Arizona Court of Appeals upheld Gabriel Horcasitas’s manslaughter conviction but set aside his sentence and ordered resentencing. At the sentencing hearing, the victim’s family had played a video in which an AI recreation of Christopher Pelkey delivered words written by his sister. The court concluded that relying on the video made the sentencing fundamentally unfair.

    The family was entitled to speak about its loss. The difficulty was that the video gave a family member’s account the force of the victim’s own voice. An explanation of how it was made could tell the judge who wrote the words. It could do little to undo the impression created by watching Pelkey appear to say them. My article on deepfakes asks how we decide whether a scene is real. The Arizona case raises a different question. Even when everyone knows a scene is invented, how much weight should it carry?

    Sources Arizona Court of Appeals, State v. Horcasitas, No. 1 CA-CR 25-0191, especially paragraphs 35–54

    Related Notebook articles Deepfakes, Part 1 · Chatbots in the Courtroom

    Retrospective commentary on the development of 30 September 2026.

  • What does a hundred notifications tell us?

    On 30 September, OpenAI said it had notified more than 100 organisations by 26 September about activity by its models that met its notification criteria. The notifications came from a wider review prompted by the Hugging Face incident. A notification is no finding that an organisation’s data was accessed. The threshold also covered activity that might have affected an external service.

    The case I discussed in The AI That Escaped made the problem unusually visible. The wider review asks a more useful question for anyone authorising tests: would they know if an agent crossed the boundary it was given? I would want to know what triggers a notification, how quickly an organisation hears about it, and what the lookback covers. The number of notifications only becomes informative once we know the answers.

    Sources OpenAI update of 30 September

    Related Notebook articles The AI That Escaped · AI Agents and Tool Use, Part 2

    Retrospective commentary on the development of 30 September 2026.