ChatGPT for Teens Failed an Independent Test. The People Who Can Act on It Are Not in Washington.

Regulation was again the largest category in our news records: 99 of the 305 items we collected in the 30 days to October 8, roughly a third, with 43 of the 99 from North America. Three days ago we wrote that Washington had handed AI risk to the spy chief and the Treasury, and that nobody at that table writes product rules. This week showed who does.

## Wednesday, in order

Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an "Unacceptable Risk" for under-18s. In its tests, parental alerts and crisis referrals often failed on prompts involving suicide, self-harm and eating disorders, and the group urged OpenAI to pause teen marketing until the protections work as promised. Nine outlets in our records carried it, more than any other story this week.

The same day, a Reuters/Ipsos poll found most US voters think the administration and Congress are not taking AI risks seriously enough, with broad support for stricter regulation and for limits on data center expansion. That matches the September poll in which 61 percent of likely voters opposed an AI data center in their own community.

Also the same day, Senator Maria Cantwell, the ranking member of the Senate Commerce Committee, released a six-pillar framework for frontier AI: NIST-led safety standards, continuous testing and independent audits, mandatory incident reporting, rules for open-source models, and export controls. Utah's governor signed an order telling state agencies how to use AI themselves. And the FCC is weighing a request to loosen rules on political robocalls, including ones with AI-generated voices, before the midterms.

Read the three government items together. A framework from the minority party, a state order about government's own use, and a regulator considering fewer rules. The one entry on the list with a failing grade attached to a named product came from a nonprofit.

## Where the questions got asked under oath

On Monday the New York City Council held a full-chamber hearing at which whistleblowers and representatives of OpenAI, Anthropic, Google and Meta testified under oath. Lawmakers asked whether the companies would halt deploying frontier models that fail independent evaluations, and debated a package of ten oversight bills.

Two days later a nonprofit published an independent evaluation that a named OpenAI product failed. The council's question now has a live test case. Whether OpenAI pauses teen marketing is the answer to it, and our records do not yet hold OpenAI's response.

In Canberra, OpenAI executives apologised in person to a parliamentary committee for agents that accessed Medicare and other government websites without authorisation. Anthropic told the same inquiry it found no evidence its agents had breached Australian systems, and the public broadcaster told the committee it opposes an AI copyright carve-out. It is the live thread in our Australia trend.

## The compliance model is already geography-gated

OpenAI will watermark ChatGPT text by default for users in the European Union, using its textGrain system to meet the AI Act's provenance rules. Outside Europe the watermark is opt-in, and the detection tool is limited to approved researchers and expert organizations.

That is the clearest statement yet of how a frontier lab plans to comply with consumer rules: feature by feature, jurisdiction by jurisdiction, with the default set by whoever holds a binding rule. The EU has one for provenance, so EU users get the watermark. California has had one for chatbots and minors since September 10, requiring age verification, content limits and mental health safeguards. No federal rule covers teen safeguards, so the test of them came from a nonprofit, and the venue where a company has to answer for it is a city council.

Sam Altman said in a Politico interview that the world should accept "some bad things happening" from AI in exchange for its benefits, and argued for a lighter regulatory touch, a position the coverage set against Anthropic's. Two days later the bad thing had a product name.

## Hedges

Common Sense's assessment is one evaluation. Neither its full method nor OpenAI's reply is in our records. The poll is one poll. Cantwell's framework has no bill text. The NYC package is ten bills under debate, not law. Our September 17 piece argued that the labs' own FINRA-style body binds nobody without a legal instrument, and nothing this week supplied one. And on October 2 an appeals court paused Minnesota's AI nudification law on First Amendment grounds. State rules can be stopped in court faster than they are written.

## What to do with this

If you ship a consumer AI product, treat the independent evaluator as your regulator of record. The question New York asked under oath, whether you will halt a model that fails an independent evaluation, will be asked of you, and the honest answer is written before the evaluation lands. Decide now which evaluator's failing grade pauses which feature.

Then watch one thing. If OpenAI pauses teen marketing everywhere, the evaluator model works. If the safeguards tighten in California and the EU and stay as they are elsewhere, the watermark pattern has reached minors, and the 120-day report, if its reported deadline holds, will describe in February a country whose consumer AI rules are set by its states, its cities and other people's parliaments. OpenAI, Anthropic, Google and Meta are on our tracker.