Macquarie's Virtual Peer, Anthropic watermarks and Melbourne summit close

Digital Attitudes, 4 September 2026

Friday of Term 3 Week 8, and the AI conversation that walked off the job at Camperdown on Wednesday has walked into a psychology tutorial at another Sydney university, at the same time the assessment integrity conversation gets a new global signal and the National Education Summit closes at MCEC.

Macquarie replaces psychology tutorials with Virtual Peer, and the NTEU has its first live test case

The Guardian reported on Monday 1 September 2026 that Macquarie University has replaced the weekly in-class tutorial in two mandatory psychology units with a chatbot activity, called Virtual Peer, plus online quizzes and optional Zoom drop-in sessions, and the university updated its position on Wednesday to describe the tool as "an optional supplementary tool" rather than a replacement for teaching (ACS Information Age on Macquarie denying the chatbot swap, The Australia Today on students saying they feel "short-changed"). Both units carried face to face classes in Semester 1 and moved to an online only mode for Semester 2, with the university signalling that in-person classes return in 2027 on alternating semesters so students can pick the format. Virtual Peer, built with Microsoft and trained on educator uploaded material, answered close to 80,000 student questions across the whole of 2025 and roughly the same volume in the first half of 2026 alone, mostly administrative. One anonymous student on the record: "Instead of a group discussion or a tutor teaching you, it's a class activity done with a robot." NTEU national president and Macquarie academic Dr Alison Barnes, in the same coverage: "University staff are being hammered by a workload crisis, but simply papering over those cracks with piecemeal AI rollouts could actually create even more damaging issues."

Why this matters: this is the first Australian university deployment that turns Wednesday's USyd bargaining ask (enforceable AI clauses in the enterprise agreement, not a policy document) into a live test case with a syllabus and a student cohort behind it. Two operational reads. First, for a pro vice-chancellor academic drafting a Term 4 online delivery policy, the Macquarie sequence is the one to study: introduce the tool as a supplement, let usage curve up (80,000 to 80,000 in half the time), remove the paid tutor slot, defend the position when the Guardian rings the press office. Every step is defensible on its own. The composite is what the union just filed against, and what any faculty board considering a similar move now needs to have an answer for before, not after, the phone call. Second, for a head of learning design or a dean of teaching whose 2027 assessment framework depends on tutor led interaction, the Macquarie case sharpens the vendor conversation. Any AI tutor product now needs to answer the same question in the sales meeting that Barnes just put on the record, which is whether the deployment reduces workload for staff or reduces staff for workload. The Term 4 procurement paper cannot leave that question inside the vendor's own slide.

Anthropic starts watermarking Claude output worldwide, and the case against AI detection tools has a new twist

Anthropic confirmed on Tuesday 11 August 2026 that Claude models released on or after Saturday 2 August will weave an imperceptible SynthID watermark into their text output, the change is global rather than Europe only, and Anthropic is working to backfill support into models released before that date (ACS Information Age on the Anthropic Claude text watermarking rollout, CSU academic on what the watermark means for cheating students, via The Conversation). The mark travels when the text is copied and pasted, does not change meaning, quality or readability, and is triggered by Europe's AI Act, which took effect on 2 August and requires providers to make computer generated text detectable. OpenAI, Google, Meta and Microsoft have signed on to the same regime. On the record from RMIT's Dr T J Thomson in the ACS piece, watermarking is "much more precise and reliable" for detection than the general purpose AI detectors that Australian schools and universities have been using or, in some cases, disabling. The Conversation piece from Charles Sturt University reads the same signal the opposite way: watermarks are a first party fingerprint for one model, they do not cover the stack the average student actually uses in a Term 4 assessment, and abandoning task design in favour of vendor detection is still the wrong lesson to learn.

Why this matters: this landed three weeks before NESA told NSW schools, in the same 31 August notice that limited take-home HSC assessments to one at 15 per cent, that "schools should not rely on AI detection tools as the main safeguard against malpractice" (Tuesday's brief on the finalised take-home rules). The regulator's line and the vendor's rollout now sit on the same desk, and they are asking the same question from opposite ends. Two operational reads. First, for a director of digital learning or a head of academic integrity, the practical use of Claude's watermark is a narrow one, it tells you when a piece of text came from a recent Claude model. It does not tell you when a piece of text came from ChatGPT, Gemini, a Chinese model, an open weights model running on a school laptop, or a rewrite of any of the above. If the assessment framework still relies on knowing "was AI used", the watermark buys nothing over what the last six months of detector abandonments (Curtin, Adelaide, Turnitin's own move at UWA, and now NESA) have said. Second, for a school council or academic board writing a Term 4 AI position paper, the watermark is a useful anchor in the parent night conversation about what a "responsible AI vendor" now looks like, because Anthropic's decision to apply the mark worldwide rather than just in Europe is the first commercial demonstration that a compliance ceiling written in Brussels sets the floor for the Australian classroom. That is the practical answer to the "we will wait for local guidance" objection.

National Education Summit closes at MCEC today, and the readout is what the Victorian sector takes into Term 4

The National Education Summit Melbourne closes today, Friday 4 September 2026, after two days at the Melbourne Convention and Exhibition Centre, and the AI in the Classroom stream, Classroom of the Future, Wellbeing for Future Focused Educators and Capacity Building School Libraries wrap this afternoon before the free expo shuts at 3pm (EducationHQ on the Summit returning to Melbourne with a broadened program, the National Education Summit's own Melbourne page). The two-day pass was priced at $790, the one-day pass at $450, and CPD hours mapped to the Australian Professional Standards for Teachers. The Teacher Award winners were announced at the Summit Social on Thursday night. More than 150 exhibitors sat on the free expo floor across both days, and the AI in the Classroom program covered the Australian Framework for Generative AI in Schools, generative AI for lesson planning and differentiation, intelligent tutoring, AI for accessibility, and embedding AI across the curriculum.

Why this matters: for a Victorian principal, director of digital learning, deputy or curriculum leader walking out of MCEC this afternoon with a bag of vendor lanyards and a term to go before the state's assessment year closes, the practical read is that the conversation in the AI room this week ran a full step ahead of last year's version. The frame was no longer "should we let students use AI on this task", it was "what does an assessment program look like that assumes the student had every tool at every step, and marks the depth of their own thinking anyway". That is the frame NESA just wrote into an official notice (Monday's brief on the take-home cap), the frame Macquarie's Virtual Peer defence is being tested against (today's first story), and the frame Anthropic's watermarking rollout does not, on its own, get around (today's second story). Two operational reads. First, the vendor decks that survived Melbourne this week are the pitches that will land in November board packs, and the ones that did not survive Melbourne are unlikely to see 2027. Any Term 4 procurement paper coming out of a Victorian government or Catholic system this month should treat the vendor list a leadership team met at MCEC as a shortlist, not a longlist, and cut anything that could not answer the workload question NTEU put on the record about Virtual Peer. Second, for the Victorian teaching workforce, this was the last set piece before the Grand Final long weekend and Term 4, and the CPD hours mapped to APST are now on the professional register. The teaching plan for the first fortnight of Term 4 is where the CPD from this week has to show up. Anything else, and the room the sector paid to sit in this Thursday and Friday was a workshop that did not leave the loading dock.

Three moves this week, one connecting thread. The AI protection debate walked off the picket line at USyd and into a Macquarie psychology tutorial, the assessment integrity debate walked out of NESA's take-home notice and into an Anthropic product decision, and the Victorian teaching workforce walked out of MCEC this afternoon with the Term 4 assessment paper still to write. More on how those three pieces fit together at digitalattitudes.com.au.