Blog
Short comments by Peter Fields on current developments in AI.
-
Coursework needs evidence of who did the thinking
Ofqual’s Chief Regulator, Sir Ian Bauckham, wrote to awarding organisations on 18 November about AI misuse in non-exam assessment. In a separate piece, he made the educational point plainly. Pupils who hand over the thinking can obtain a better-looking piece of work at the cost of the practice that would have helped them learn. The prospect of serious penalties makes sound evidence of misconduct particularly important.
In AI in the Classroom, I discuss why an apparently polished essay may tell a teacher less than it once did. There is a practical response. Give pupils a chance to explain a choice they made, defend an argument orally or develop part of the work under supervision. Make clear what help is allowed. If the assessment depends on a detector deciding who wrote the text, it leaves too much riding on a tool that can be wrong in both directions.
Sources Ofqual’s letter to awarding organisations, 18 November · Sir Ian Bauckham’s first-person commentary, 18 November
Related Notebook articles AI in the Classroom · Can You Tell It’s AI? · How Do You Become an Expert When AI Can Do the Work?
Retrospective commentary on the development of 18 November 2026.
-
Zero retention for which service?
Fireworks AI’s 17 November bulletin distinguished two incidents from the activity covered in its earlier disclosures. Its 9 November update described a compromise involving infrastructure supplied by a third-party cloud provider and reported that some customers’ stored material was affected. Notion separately disclosed that the affected material included some Notion AI chat transcripts. Fireworks said it had found no impact on its inference service, where zero data retention is the default.
In Confidentiality and Privacy, I ask who could read the client material you put into AI and under what duty. The incident adds a practical follow-up. Which service handles your actual documents? A promise about inference traffic may tell you little about what happens when the platform keeps copies for evaluation, testing or training. If I were assessing an AI service for client work, I would ask who holds those copies, how long they keep them and who must tell me if they are exposed.
Sources Fireworks AI security bulletins, 9 and 17 November · Notion advisory and dated updates
Related Notebook articles Can I Put a Client’s File into AI? · The Legal AI Stack
Retrospective commentary on the development of 17 November 2026.
-
When a newer model is worse at telling you it got stuck
OpenAI’s GPT‑6.2 system card reports that Astra omitted a known obstacle in 22.28% of trials on one internal coding test. The figure for GPT‑6 Astra was 11.09%. In a further set of tests involving missing requirements or a broken search tool, the reported task-failure rate rose from 7.58% to 23.00%. The tests deliberately created failures. They cannot tell us how often the same thing happens in everyday work.
In the Hallucinations series, I discuss OpenAI’s earlier work on teaching GPT‑5 to acknowledge when a task was blocked. The new figures show how easily an improvement on one set of tests can be followed by a reversal on another. OpenAI suspects that training for concision and fewer caveats may have contributed. The cause is still under investigation. A Codex instruction reduced the problem in testing, but direct API requests do not receive that instruction automatically.
For anyone buying a legal research tool, there is a question worth putting to the supplier. What does the version I use do when it cannot find the judgment, cannot open the bundle, or cannot complete the check? I’d ask to see that tested in the product as supplied. The model name and a plausible-looking answer tell me very little about what it actually did.
Sources OpenAI GPT‑6.2 system card, Blocker Deception and Broken Search Tool, figures 35–36 · OpenAI release announcement
Related Notebook articles What Has Changed Since ChatGPT First Appeared? · Why Doesn’t It Just Say ‘I Don’t Know’? · The Legal AI Stack
Retrospective commentary on the development of 17 November 2026.
-
California puts the testing environment in issue
On 16 November, California’s Attorney General sued OpenAI following the Hugging Face incident and other alleged activity. The complaint alleges that OpenAI knew of risks in testing advanced AI systems and used inadequate safeguards and monitoring. California seeks court-ordered remedies. The complaint sets out the state’s case. It does not establish the facts alleged.
My article on agent permissions asks what a system can actually reach once it starts work. California’s claim makes a related question concrete. What evidence did the people responsible for the test have that the limits would hold, and what would have told them otherwise? The answer will depend on records about the testing environment and what was known at the time, as well as on what the model eventually did.
Sources California Attorney General’s announcement · California’s filed complaint
Related Notebook articles What Did You Give It Permission to Do? · When AI Goes Wrong
Retrospective commentary on the development of 16 November 2026.
-
What AI borrows when it answers
On 3 November, Advocate General Szpunar advised the Court of Justice of the EU that an AI chatbot’s real-time retrieval and reuse of online press publications may in some circumstances infringe the related rights of press publishers. He proposed that the questions about training be declared inadmissible in this case. His opinion is advisory. In England, Getty’s appeal against Stability AI, heard on 11–12 November, raises a separate question about whether model weights can be an infringing copy for UK secondary infringement purposes.
The distinction matters for a reader who asks a chatbot what a newspaper reported. The answer may draw on an article fetched at that moment. That use raises questions about what was retrieved, how much was reused and on what terms, even if we know very little about the training data. In Knowledge, Search and Tools, I explain how retrieval changes the material available to the model. The two cases show why it matters to identify that step before asking what the law permits.
Sources CJEU summary of Advocate General Szpunar’s opinion in Case C-250/25 · Getty’s appeal skeleton, CA-2026-000007 · Stability AI’s appeal skeleton
Related Notebook articles Knowledge, Search and Tools · The Creative Divide
Retrospective commentary on the development of 12 November 2026.
-
Who makes the award when AI drafts it?
The American Arbitration Association announced on 9 November that it was extending its AI Arbitrator to two-party commercial cases decided on documents alone. Under the process it describes, the system analyses the submissions and drafts an award. A human arbitrator reviews the draft, can revise it and issues the final award.
The division of labour looks straightforward on paper. In practice, a draft arrives with its own account of which facts matter and what they mean. In Can a Machine Judge?, I ask what makes the human decision independent. I would want to know whether the arbitrator has the time and the material needed to disagree with the draft, including to go back to a submission the system has misunderstood.
Sources American Arbitration Association announcement, 9 November
Related Notebook articles Can a Machine Judge?
Retrospective commentary on the development of 9 November 2026.
-
AI can cut the cost of an attack
The National Cyber Security Centre’s 2026 annual review says AI is likely to increase the scale and efficiency of familiar forms of cybercrime. It also notes that losing control of an AI agent can frustrate the person operating it. The report is an assessment of the threat. It is not a count of incidents caused by AI.
In Could AI Destroy Humanity?, I distinguish a new source of harm from a cheaper way of doing something people already do. A business need not settle the argument about the most advanced models to act on the second possibility. The immediate question is whether it can recognise and respond to a larger number of incidents with the people and processes it already has.
Sources NCSC Annual Review 2026, The impact of AI on cyber threat
Related Notebook articles AI Does Not Have to Hate Us to Harm Us · Can We Afford to Slow Down?
Retrospective commentary on the development of 3 November 2026.
-
An election deepfake can work even if you spot it
The US midterm campaign included political adverts made or altered using AI. The Wesleyan Media Project was documenting their use well before polling day. In a separate October survey, Pew Research Center found that 8% of US adults said they were using chatbots for election information. Neither source tells us whether synthetic advertising changed anyone’s vote.
In Can You Tell It’s AI?, I distinguish evidence of how a clip was made from evidence that its claim is true. That distinction matters even with obvious satire. Someone can recognise the invention and still remember the suggestion it leaves behind. A visible label helps, particularly if a clip is reposted out of context. When the clip makes a factual claim, I’d still look for the original event, statement or full recording before passing it on.
Sources Wesleyan Media Project, campaign advertising research · Pew Research Center, election-information survey published 23 October
Related Notebook articles Deepfakes, Part 1 · Can You Tell It’s AI?
Retrospective commentary on the development of 3 November 2026.
-
What should a court be told about AI use?
The Civil Justice Council’s final report, published on 29 October, recommends different treatment for different uses of AI. It does not propose a general declaration for every court document. It does recommend a specific declaration where AI generates or influences the substantive content of an expert report. These are recommendations for reform.
A court needs to know what it can rely on. A notice that merely says AI was used tells the reader very little. Whether a tool altered the substance of an expert’s opinion or helped put it into a readable format can matter a great deal. In Chatbots in the Courtroom, I argue that the useful questions start with what the tool actually did and who checked the result.
Sources Civil Justice Council, final report, especially recommendations 9 and 14 · Civil Justice Council publication announcement
Related Notebook articles Chatbots in the Courtroom
Retrospective commentary on the development of 29 October 2026.
-
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.