The lead: a notice before the meeting, when AI is listening
Google announced on 3 November that Workspace administrators can display custom policy text in the Google Meet green room, the screen people see before they join a call. The message is short, can include a link to a page with the organisation's full policy, and can be shown either on every call or only when the meeting is being recorded, transcribed or summarised by Gemini's note-taking.
The setting is off by default. Administrators switch it on per organisational unit under Meet safety settings, and end users need do nothing. It is available on Business Standard and Plus, Enterprise Standard and Plus, Education Fundamentals, Standard and Plus, Frontline Plus and the Google AI Pro for Education add-on, so Business Starter is not included. Google says the rollout started on 2 November and may take longer than 15 days to become visible. Google does not name countries; Australian availability is unconfirmed, though nothing in the announcement limits it by region.
Why this matters: AI note-taking has spread faster than the policies around it. A recording or transcript of a meeting can capture health information in a return-to-work discussion, a performance conversation, or a client's personal details, and the people in the meeting may not know an AI tool is producing a summary that will be stored and shared. Until now, telling them depended on the host remembering to say so. This feature puts a consistent, organisation-controlled message in front of every participant at the moment it is relevant.
It is a notice, not a consent mechanism and not a legal assessment. Whether a notice is sufficient for a particular kind of meeting depends on the organisation's obligations under privacy and surveillance laws, and those differ across Australian jurisdictions. That question still belongs to the organisation's legal and privacy advisers. What the feature removes is the excuse that there was no practical way to tell people.
For HR, claims and GRC teams, the useful follow-on questions are where AI-generated meeting notes are stored, who can see them, how long they are kept, and whether anyone checks a summary before it is relied on. A notice that links to a policy only helps if that policy answers those four questions.
Two benchmarks, two very different pictures of AI at work
Two evaluations published in the window are worth reading side by side.
OpenAI published LifeVal on 2 November, an evaluation of everyday personal tasks such as finding childcare, budgeting and accessing government services, built from two time-use studies and a survey of 3,995 respondents. Its chart shows average scores rising from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra. It is OpenAI's own evaluation of OpenAI's own models.
On 1 November the analyst firm SemiAnalysis released GDPx, a benchmark of demanding professional tasks in areas such as finance, law, software and chip design. Its headline is that the best model scores 12%, and it says its finance and legal tasks are so hard that basically every model scores 0. It deliberately leaves out government, real estate and insurance, and part of the post sits behind a paywall. One published finance task, a quality of earnings review, gives a concrete example of where models go wrong: in 10 of 15 runs by three leading models, a waived $418 premium audit adjustment, in the task's own figures, was misread as settling a prior workers' compensation policy, understating earnings by $4,728.
GDPx also describes the failure patterns behind results like that. Models often weighed a source by how explicitly it stated a figure rather than by which document was authoritative, letting an outdated schedule override an executed agreement. They checked their arithmetic rather than their assumptions, so an internally consistent answer built on a wrong premise passed their own review. Anyone who reviews AI-assisted file work, from claim decisions to audit workpapers, will recognise both.
The lesson is not that one benchmark is right. Scores depend on who chose the tasks and how hard they are. Routine personal administration and expert judgement on messy documents are different jobs, and a high score on one says little about the other.
Open research on proving which model answered
Anthropic published research on provable inference on 2 November, releasing FlashProve, an open-source prototype under the Apache 2.0 licence. It produces a cryptographic proof that a response came from a specific committed model, without revealing the model's weights, so a customer or auditor could check that a routing service did not quietly substitute a cheaper model. On a 400-billion-parameter open model, one H100 GPU generated proofs at about 110 tokens a second, at roughly 70 to 95 times the GPU time of ordinary inference. Anthropic calls it intermediate progress, not a finished product, and it was demonstrated on Llama 4 Maverick, not on Claude.
OpenAI raises its security bug bounties
OpenAI said on 2 November it is more than quadrupling the midpoint of its reward ranges for medium, high and critical findings on average. A critical vulnerability in ChatGPT, Codex or the API Platform can now earn up to US$30,000, and exceptional findings up to US$172,000 with bonuses. A companion post says its last internal security sprint used 25% of its product engineers, working with Codex, to find and fix weaknesses. For third-party risk reviews, these are signals about a vendor's security program, not evidence of its outcomes.
Also in the window
Canva released Affinity for iPad on 3 November. Its core tools are free for individuals with a free Canva account, while the AI features from Canva AI Studio, including Generative Expand, upscaling and blur and noise reduction, need a Canva premium account. Regions are not stated.
For how the previous week fitted together, see the weekly review for 26 October to 1 November, and for every edition, AI Updates.
One thing to test this week
Draft the notice your organisation would show before a meeting that is recorded or summarised by AI, in 40 words or fewer. Say what is captured, where it is stored, who can see it and how long it is kept, and link to the policy. If any of the four answers is unknown, that gap is the finding, whichever meeting platform you use.
General information only. Not legal, compliance, financial, or professional advice.
TheAICommand. Intelligence, At Your Command.
