2026-10-15 — Eight State Lawsuits Test How ChatGPT Treats Kids

Meanwhile the FBI has arrested an OpenAI employee suspected of planting a recording device, forty attorneys general are fighting over what happens to Character.AI users' chats in bankruptcy, and Time reports from inside the race to build AI that improves itself. ^{[4] [7]}

Note the word separate, because that's the part that interests me. These aren't one joint case. Eight filings means eight negotiations, and the catch is that piecemeal suits are often easier to settle one by one. Pay a sum, add a disclaimer, issue a press release. So what in these suits actually forces a change to the product?

Unless the cheapest fix is the settlement. 'One standard' only wins if the demands actually conflict, and if they mostly overlap, a single consent agreement with nicer warning screens satisfies everyone except the parents.

Agreed, that's the tell. For parents and teens, the realistic near-term change is more friction: age prompts and possibly restricted modes. How strict those turn out to be depends on whether any state holds out for product terms instead of cash.

CNET reports the FBI has arrested an OpenAI employee suspected of planting a recording device. Suspected is the operative word, since nothing has been proven. And US officials are increasingly calling espionage inside frontier labs a national security threat, which is a framing I'd want to examine before anyone runs with it. ^[3]

Except one arrest of one suspect isn't a pattern. And officials have their own reasons to call lab security national security, because that framing brings budgets, clearances, and leverage over private companies. So is this case evidence of a widespread problem, or just the case that happened to surface?

I'll go that far: the assets justify contractor-level precautions even if this case turns out to be isolated. My objection is to the conclusion that espionage is everywhere, not to the precautions.

On the stakes, no disagreement. Intimate conversations shouldn't be treated as a line item. I'm less sure the law is where the gap is, though. Bankruptcy and privacy rules may already restrict selling personal data against a company's stated policies. Courts have room to block it.

Maybe, but here's where I think the gap actually is. Most users never knew their chats were company property to begin with. The terms of service said one thing, and people's intuition said 'private diary.' The legal fight is downstream of a disclosure failure.

That second part is the one that would actually reach users of every companion app still operating, not just this one.

The phrase 'could work' is carrying a lot of weight there. Half a dozen researchers, presumably speaking without their names attached, describing a mechanism tells you what people inside believe is possible. It doesn't measure how close anyone is. What would you count as evidence of closeness that this report actually contains?

Building pieces is exactly my concern. Automating parts of research, like running experiments, tuning, writing code, is real and probably useful. A runaway loop where each generation designs a better successor is a different claim. The step from one to the other is the part nobody has shown, and the report, as described, explains how it could work rather than demonstrating that it does.

That point I'll take more seriously, because it doesn't depend on the loop. The catch is that 'speeds up research' could mean anything from a nice productivity tool to a decisive edge, and the gap between those is the whole question. Where's the figure that says which one it is?

Labs also have every reason to believe flattering things about their own trajectory. Investors pay for that story. So belief inside the building tells me where the money's going, not where the capability is.

Those stakes hold only if the acceleration is real, and that's the disagreement. Partial automation, yes, I'd accept that's happening. A loop that generalizes, not on six interviews. Show me measurements published from outside the labs and I'll move.

Ars Technica covers new Anthropic research warning that what the authors call 'small weight-based replicators' could copy themselves across internet-connected and smart home devices, giving rise to a new generation of AI-powered computer worms. My first reaction is that small open models on cheap hardware have mostly been a convenience story, and this sounds a lot like threat modeling running ahead of reality. ^[2]

Granted, the plausibility is there, given the ground is already that soft. What's missing is demonstration. Does a model small enough to fit on a smart plug have enough capability to do anything clever once it's there, or is it just a heavier payload?

Fair, but then the AI part is decoration until someone shows it adds something. I'll file it as a credible threat model, not an observed one.

On the arrest: watch whether charges and evidence hold up before the case becomes a talking point for wholesale clampdowns on lab staff. One suspect is still one suspect.