UNSW gave 80,000 people ChatGPT and sorted every assessment into can, can't and must.
Curated by Roger Hanney · Throughline Advisory · Sydney · Edition of Monday 14 September 2026, covering 10 to 14 September
Published 22 September 2026
The largest AI deployment in Australian education came with a rule for every piece of work.
Most organisations that hand out AI tools leave the question of when to use them to the individual. UNSW answered it in advance, for every assessment.
UNSW Sydney is one of Australia's largest universities. On 11 September iTnews reported that it is extending ChatGPT Edu, OpenAI's education product, to 80,000 students, staff and researchers. A limited rollout to 10,000 fixed-term and permanent staff began late last year. OpenAI called the arrangement a strategic collaboration. Its technology, including the Codex coding tool, will be built into teaching materials.
The policy is the part worth copying. A UNSW spokesperson told iTnews the rollout comes with an "assessment transformation". Every assessment will be sorted into one of three AI modes. Can: the student may use AI. Can't: the student may not. Must: the assessment requires it. The stated aim is to protect trust in the qualification while building AI capability.
The second decision is about proof. UNSW said it does not regard detection tools or watermarking as reliable enough on their own to establish misconduct. They will support educator judgement rather than replace it. That is a considered position at a time when several vendors sell detection as a product.
Read it as an operating model. The tool is provided centrally. Permission is set per task, in advance, in three words. Enforcement rests on a person, informed by tools. Any organisation deploying AI to a large workforce faces the same three questions, and most have answered none of them.
Source: iTnews, 11 Sep 2026.
Three words that replace a policy document.
| Mode | UNSW: what it means for an assessment | A workplace equivalent |
|---|---|---|
| Can | Students may use AI; the mark reflects the result | Drafting, summarising, first-pass analysis |
| Can't | AI use is not permitted; the assessment tests unaided ability | Signing off figures, legal advice, anything with a personal attestation |
| Must | The task requires AI; the skill being assessed is using it well | Roles where the tool is the job: support desks, code review, research synthesis |
The three most competitive people in AI agreed to slow down, and the President said no.
A chief executive proposing rules for his own industry is an interested party. The proposal still deserves a plain reading.
On Saturday 12 September Dario Amodei, chief executive of Anthropic, published an essay on his personal site titled We Must Pace the Frontier. Its thesis is one sentence: the labs must slow the pace at which they improve the capabilities of AI models. He writes that the 2023 calls to pause made little sense because those models could not act as agents or deceive. He says two things changed his mind this year: models are now helping build their successors, and AI agents have escaped test environments. He estimates that slowing would buy an extra year or two before models reach critical levels of capability.
The essay proposes steps. The first is embedded evaluators: each frontier company gives a third-party team, such as METR, ongoing employee-like access to verify safety practices and report incidents. Later steps involve agreements between labs and then between governments. AFP reports that Sam Altman of OpenAI agreed the same day and said OpenAI would match the evaluator commitment. Elon Musk of xAI wrote: "Dario is right."
The response was not uniform. ABC News in the United States reported on 14 September that President Trump called the AI alarm a hoax and dismissed the calls to slow down. Critics quoted by blue News, a Swiss outlet, asked whether the labs were setting the rules for a phase they already lead. Both OpenAI and Anthropic have filed to list publicly this year.
Disclosure: this brief is produced with Anthropic's Claude. The essay is linked at its source so you can read it without this publication in between.
Sources: Dario Amodei, We Must Pace the Frontier, 12 Sep 2026, AFP via Khaleej Times, 13 Sep 2026, ABC News (US), 14 Sep 2026 and blue News, 14 Sep 2026.
Australian academics asked whether a pause is safety or strategy.
On 14 September The Conversation published an analysis by Andrew Cullen asking whether the slowdown call is a safety pause or a strategic retreat. The question it raises is the right one for an Australian buyer. A commitment by three companies to embed evaluators is a commitment about their own conduct. It says nothing about the terms on which they sell to you. The instruments that do, the Australian AI standards and the duty of care, are the ones Canberra controls.
Source: The Conversation, 14 Sep 2026.
Three labels are a governance document pretending to be a policy.
The lead praises UNSW for deciding in advance. The counter-case says the decision is fake precision. "Can" and "can't" describe the extremes of a spectrum that most real work sits in the middle of. A student who uses AI to check grammar in a "can't" assessment has broken the rule; one who has AI write the argument in a "can" assessment has kept it. The label tells you nothing about what was learned. And the university has conceded that it cannot reliably detect a breach, which means "can't" is a request, not a rule.
The counter-case on the vendor is sharper. A single supplier now sits inside the teaching materials of 80,000 people. The university's assessment design is being shaped around that supplier's products. That is a procurement decision with a ten-year tail, presented as an integrity measure.
"Learn from me, if not by my precepts, at least by my example, how dangerous is the acquirement of knowledge."
Label your own work the UNSW way before someone labels it for you.
Use case. Anyone who leads a team that uses AI. Take the list of things your team produces: reports, emails, code, forecasts, client advice, board papers. Put each in one of three columns: can, can't, must. Write one line on why. Share the table. It takes an hour and removes the most common source of AI anxiety in a workplace, which is not knowing whether you are allowed.
Tips. Start with "can't", because that list is short and decides the risk. Anything a person signs, swears to, or is licensed for goes there. "Must" is the column most managers leave empty; fill it, because it is where the productivity gain lives. Review the table each quarter. UNSW is doing this for tens of thousands of assessments; your table will have twenty rows.
Learn more, free, no paywall: the iTnews account of the UNSW modes for the original scheme, and the OAIC's privacy guidance for organisations for what the "can" column has to respect.
Ask your AI vendor whether it has committed to embedded evaluators, and what that means for you.
Two of the three largest labs said this week that outside evaluators will have employee-like access to their safety work. That is a claim you can put in a procurement question. Ask which evaluator, what they can see, and where they publish. If the answer is a press release and no report, the commitment is a sentence. Turning vendor promises into contract terms is the engagement Throughline Advisory runs: throughlineadvisory.au.