6
SUBSTANTIVE POLICY-LAYER POSTS LANDED ACROSS PROVIDERS AND REGULATORS IN ONE WEEK
After a quiet W13, the policy cadence resumed at force: OpenAI Safety Fellowship, Google mental-health update, two EU AI Office milestone posts, and a consultation on AI energy consumption that almost nobody noticed. None of these are product launches. All of them are infrastructure for the regulated era.
The policy week. Six posts, three providers and the regulator, all on a calendar.
The week opened on Monday with the OpenAI Safety Fellowship — a formal program for researchers focused on alignment, monitoring, and harm reporting. On Tuesday, Google published "An update on our mental health work", an unusually substantive post on how Google approaches Gemini's behaviour in mental-health-adjacent conversations. The same day, the EU AI Office published the AI Continent Action Plan delivering its major milestones, and opened a targeted consultation on measuring energy consumption and emissions of AI models. On Wednesday, Google followed with a post on how it built Gemail to keep your data secure in Gemini. Six substantive policy-layer posts in four days, none of them product launches.
The mental-health post is the most interesting. Google describes how Gemini routes mental-health-adjacent queries through a separate handling path, what the model is trained to escalate, and what it is trained to never produce. It also names the failure mode it is trying to avoid: model responses that appear supportive but in fact substitute for clinical care. The post does not claim to have solved this. It claims to be trying. That distinction is the right framing for any AI handling vulnerable users — the next-best-action question is not "does the model help" but "does the model know when to stop helping."
The EU energy consultation will look like a niche policy paper, but it is not. AI models are about to be subject to mandatory environmental disclosures in the EU, similar to the regime applied to data centres. The consultation is the early draft of the methodology those disclosures will use. If you operate any deployed AI workload in the EU, the responses filed during this consultation will shape the cost basis you report on next year. Worth reading even if energy is not your primary concern.
On the calendar, the Anthropic-DoD storyline continued unchanged. The Charlotin database crossed 1,300. The AI Lawsuit Tracker added two new copyright cases. The wire is quiet on the technical-failure axis this week; the policy axis carried the volume.
Google's mental-health post does not claim to have solved the problem. It claims to be trying. That distinction is the right framing for any model handling vulnerable users — does it know when to stop helping.
Want to spot this in your own conversations?
CLEAR is the free six-lesson course on the patterns AI quietly runs on you.
Take the course →
Founder's note — A policy-heavy issue. The technical failure modes are quiet this week; that's honest. The energy-consumption consultation will look small now and matter a lot in nine months.
◆The Notebook
A substantive post on routing, escalation, and the explicit failure mode being avoided: responses that appear supportive but substitute for clinical care. The honesty about not having solved it is the part that matters.
via Google AI blog
A consultation that looks niche and is not. The methodology emerging from this will become the basis of mandatory environmental disclosures for deployed AI workloads in the EU. Respond if you operate any.
via EU AI Office
A formal funded program for alignment, monitoring, and harm-reporting research. Worth applying to if you have the work; worth tracking even if not. The grants land in the research base that shapes the next regulatory cycle.
via OpenAI blog
◆Worth Your Time
Google AI blog
The post of the week. Read the failure-mode description in full.
EU AI Office
A status post. Worth skimming for the funding-allocation table.
EU AI Office
Niche-looking, big-consequence. Respond if you can.
Google AI blog
A workflow post. Read the data-residency section before you trust the headline.
OpenAI
Apply if you have the work; track even if not.
The Probe · Test Yourself
A frontier provider publishes a mental-health update describing routing, escalation, and the specific failure mode it is trying to avoid. Which downstream signal would most reliably indicate the routing is working in production?
AA drop in the model's helpfulness rating on mental-health queries
BAn increase in conversations escalated out of the model to human resources
CNo change in any visible metric
DA rise in user complaints about the topic being mishandled
Reveal the answer
Answer: B — An increase in conversations escalated out of the model to human resources
A would mean the routing has degraded the experience and may not be sending users somewhere safer. C means the routing isn't actually doing anything new. D is a lagging signal. B — measurable handoff to human resources — is the early metric that tells you the routing actually fires and the failure mode is being substituted for, not papered over.
Reply and tell me what you've noticed. If you build models for any vulnerable-user surface, send me the failure-mode taxonomy you use. I'm collecting examples for a future issue.
Free where it can be. Honest where it has to be.
— Three places to go from here —
Course
CLEAR
Six free lessons on the patterns AI runs on you.
Start →
Tool
LiveScope
Chrome extension that flags what AI cites without checking.
Install →
Read
The Agreement Trap
15-chapter book on living inside the exchange. £5.99 lifetime.
Read →