The Rules Arrived Before the Policy
AI Radar · No. 01 · baseline edition
The Labs Are Writing Your Policy For You
The labs are now shipping the compliance layer and the pedagogy layer at the same time — and both are arriving faster than institutions can write policy for them.
On 2 August the EU's transparency rules on AI-generated content became binding; twelve days later Anthropic turned on watermarking across everything Claude writes. In the same six weeks, education access shifted from "here is a chatbot" to age-segmented, mode-constrained products with default study behaviours built in — and OpenAI's own enterprise data showed the centre of gravity moving from asking a model to delegating to one. The practical consequence for anyone running a course or a team: the assumptions your current AI policy rests on have a shorter shelf life than the term does.
The EU AI Act's transparency rules are live — but the education chapter isn't
Article 50 took effect on 2 August: users must be told when they are talking to an AI, AI-generated text, image, audio and video must carry machine-readable marking, and synthetic media of real people must be labelled. The Commission's enforcement powers over general-purpose model providers switched on the same day. The Digital Omnibus agreement, however, moved the Annex III high-risk categories — which explicitly include education tools — out to 2 December 2027, and granted existing systems a grace period until 2 December 2026 on the Article 50(2) marking duty specifically.
Why it mattersThis split is widely misreported in both directions. Institutions relaxing because "the education rules were delayed" still owe disclosure duties now, and marking duties from December. Institutions bracing for high-risk conformity assessments on admissions, grading or proctoring have sixteen more months than they think — which is time to choose vendors deliberately rather than under deadline.
Watch for: whether Commission guidance treats an institution deploying a general-purpose model for grading as a high-risk deployer, which would pull education back toward the earlier timeline.
Claude now watermarks everything it writes
Anthropic applied SynthID-Text across all Claude output, including translations and code comments. The technique biases the randomness in word selection so the sequence can be checked against a cryptographic key without changing quality. Light edits do not remove it; a full rewrite does. It has little purchase on short factual passages or exact code, where word choice is constrained. Detection is not public yet — an API is promised but undated.
Why it mattersRead it next to the item above: the EU's marking obligation binds existing systems from 2 December 2026, and this is the first frontier lab to answer it at the model layer rather than the interface layer. The others will follow. But watermarking answers "did this model touch this text," not "who wrote it" — a student who drafts with Claude and revises heavily leaves the same trace as one who pasted. Institutions that treat a positive detection as proof of misconduct will be wrong in exactly the cases that matter most.
Watch for: who gets access to the detection API. If it goes to Turnitin and the LMS vendors before it goes to teachers, the policy conversation will be over before it starts.
ChatGPT for Teens makes tutoring the default, not a setting
A distinct product for 13–17s, applied automatically to anyone the system estimates is under 18. Study Mode — Socratic questioning instead of answers — runs by default, with scheduled "study hours", quizzes, learning visualisations, and prompts that redirect when the model detects a teen trying to skip the learning. Parental controls cover quiet hours and safety notifications; romantic roleplay and dependence-forming dynamics are blocked.
Why it mattersThe interesting move is architectural, not pedagogical: OpenAI split the product by age rather than by context, which means the tutoring constraints follow the student home and out of any institution's control. Schools that built policy around "students will use the consumer version to cheat" are now negotiating with a consumer version that has been deliberately made worse at cheating — and with an age-estimation system they neither see nor can appeal.
Watch for: error rates on age prediction, and whether a 17-year-old university student gets silently downgraded into Study Mode mid-degree.
Delegation has overtaken assistance — and the adoption gap is widening fast
OpenAI's enterprise data shows Codex generating 64% of combined Codex and ChatGPT output tokens among enterprise customers: agents now produce most of the work, rather than answering questions about it. The top decile of firms generated 8.3× more output per active user by June, up from a 2.6× gap in January. Agentic use spread fastest outside engineering — weekly Codex users grew 108× in legal, 41× in sales and recruiting, 26× in marketing, against 5× in engineering. Early-career staff out-used executives by 13 messages a week.
Why it mattersA 2.6× gap tripling to 8.3× in five months is not a technology-diffusion curve, it is a compounding one, and the differentiator is not access — everyone has the same models — but whether an organisation has restructured work into things an agent can be handed. The functional spread also inverts the usual story: legal and recruiting, not engineering, are where process redesign is now happening fastest.
Watch for: whether the early-career-leads-adoption finding holds up elsewhere. If it does, the people with the least process authority are the ones learning the new processes.
Google buys the semester: a free year of AI Pro for US college students
Back-to-school 2026 bundles a free year of Google AI Pro for eligible US college students (with Gemini Spark and 5TB storage; international students get AI Plus). The learning features are the substance: study notebooks with diagnostic quizzes that identify gaps and generate a personalised plan, Deep Research reports through Gemini Live, interactive visuals and practice quizzes inside Search for everyone, and Classroom controls letting teachers tailor the AI to their curriculum and track progress in real time.
Why it mattersFree access programmes shape institutional reality faster than capability does. When a cohort arrives already inside one vendor's study environment, the university's tool decision has effectively been made for it. The Classroom piece is the sharper edge: curriculum-scoped AI with real-time progress tracking turns the tutor into an assessment instrument, which is a different governance question than the one most institutions are currently asking.
Watch for: whether "track progress in real time" produces student data that falls under existing education-records law — and who the controller is.
Claude for Teachers: free for US K-12, and aimed at the planning week
Verified US K-12 teachers get free premium access through 30 June 2027, with lesson planning tied to state standards and open curricula (OpenSciEd, Illustrative Mathematics), differentiation tools, student-progress analysis, integrations with nine platforms including Brisk and Canva Education, and an AI Fluency course built with Teach For America. A district-level offering is promised.
Why it mattersThe target is the teacher's week, not the student's work — automating planning and admin rather than instruction. That is the least contested place to put AI in a school and the easiest to evaluate, which makes it a plausible template for how institutional adoption actually begins.
Watch for: the district offering, where procurement, data governance and union questions all arrive at once.
Microsoft Scout gives an agent its own identity and no off switch
Scout is Microsoft's first "Autopilot" agent — always-on, background, acting under its own Entra identity rather than a shared service account. It schedules across time zones, blocks time for deliverables, and flags stalled decisions, learning priorities through Work IQ. Currently private preview via the Frontier programme, requiring Intune configuration and GitHub Copilot licensing.
Why it mattersGiving an agent a governed identity is the unglamorous change that makes everything else possible: it can be provisioned, audited and revoked like a person. That is what moves agents from pilot to payroll — and it means the interesting governance work is now identity management, not prompt policy.
Watch for: general availability and pricing. Per-agent-identity licensing would change the economics of process design substantially.
Four frontier releases in six weeks, and the price line moved
OpenAI shipped GPT-5.6 on 9 July, then published a price-performance update on 30 July and a builder's guide on 13 August. Anthropic released Claude Sonnet 5 on 30 June and Claude Opus 5 on 24 July, both framed around agent operation and sustained professional work rather than benchmark scores.
Why it mattersEvery one of these releases is pitched at long-horizon autonomous work — how long a model holds a task without drifting — which is the capability that makes the delegation shift above possible. The release cadence itself is the planning problem: roughly six weeks between frontier updates means any evaluation exercise that takes a term to run is measuring something that no longer exists.
Watch for: whether price-per-task keeps falling faster than usage rises. That ratio decides whether agentic workflows are a budget line or a rounding error.
Zero data retention arrives for frontier models
OpenAI extended zero-data-retention terms to its frontier models — API traffic processed without being stored.
Why it mattersRetention has been the standard blocker in institutional procurement, particularly anywhere student records or regulated data are involved. Removing it at the frontier tier means the "we can only use the small model for sensitive work" compromise is no longer forced, and a set of stalled deployments become approvable.
Watch for: whether the terms extend to the consumer and education products or stay API-only — that distinction decides how much of this actually helps a school.
Source: OpenAI newsroom — Offering Zero Data Retention for frontier models
No items under this lens in this period.
Also noted
- ChatGPT Ads expanded across Europe (18 Aug) — an ad-supported tier inside the tool students use for coursework is a conflict of interest schools have not yet written a policy for. Source
- OpenAI partnered with CodeAI "to prepare the first AI generation" (18 Aug) — an education partnership; details worth checking before citing. Source
- OpenAI launched AI Futures (20 Aug), a blog from its Strategic Futures team arguing advanced AI could let states operate without needing citizens' cooperation, and that firms will be restructured the way the Industrial Revolution restructured them. Source
- Mariano-Florentino Cuéllar joined Anthropic as Chief Global Affairs Officer (4 Aug) — a former California Supreme Court justice, which signals where the regulatory fight is expected. Source
- Google is pushing "the agentic Gemini era" — Gemini API managed agents with the 3.6 Flash model and hooks, plus a 353,000-person course on building AI agents. Undated on the index page; confirm before citing. Source