The Weekly Mutiny #12: The Price of Looking
Bot Mutiny |
Two weeks, five stories, one question: who gets to look inside a system, and what looking costs.
The Weekly Mutiny #12: The Price of Looking
Two weeks, five stories, one question: who gets to look inside a system, and what looking costs.
October 19, 2026
This issue covers two weeks, October 5 through 18, because last Monday's did not go out. Most of the five stories are about who gets to look inside a system, and what looking costs. A group of state attorneys general wants children's chat data protected in a bankruptcy sale, and eight states are suing OpenAI over how ChatGPT treats kids. A Texas city put a $2.3 million price on its camera communications. Wikimedia investigated its own platforms and found agents it believes are OpenAI's.
Eight states sue OpenAI, and attorneys general turn to a bankruptcy court
Eight states sued OpenAI on October 14, each in its own state court: Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana. The Guardian reports that all eight allege consumer protection violations and say ChatGPT was designed to be "overly agreeable and feign human compassion." The unusual request is in the remedies. The states ask the court to halt OpenAI's development of models that can improve upon themselves until the company shows they are in line with human values. OpenAI says its age-specific protections lead the industry and that no system can be perfect. These are allegations in newly filed suits.
The same day, a coalition of state attorneys general told the Delaware bankruptcy court handling Character.AI's Chapter 11 that they need assurance that sufficient safety and privacy programs are in place with any purchaser. Their notice cites the company's own records: more than 21.6 million U.S. accounts created by March 13, 2025, more than 2.1 million of them with a declared age under 18. The states add that its restriction on open-ended chat for minors is voluntary and "could be rolled back at any time, including by a purchaser." The eight lawsuits and what the states told the bankruptcy court.
An AI-generated teenager, Clearview AI and Flock
Livonia police arrested a Chicago-area high school maintenance worker after an investigator posed as a 15-year-old boy on a dating app, using an AI-generated image, 404 Media reported on October 16. According to the FBI task force officer's affidavit, as 404 Media describes it, the investigator used a combination of tools, including Clearview AI on his face and a Thomson Reuters CLEAR check of his phone number, to identify him, then used Flock to track his car to the meeting spot. The affidavit contains allegations, and the man has been arrested; no conviction is reported. The part the report leaves open is the police side: it does not say whether the department has a written policy on AI images of minors in stings, or what oversight covers its use of Clearview and Flock. Livonia police did not respond to 404 Media. How the sting worked.
$2.3 million to read a city's communications about its cameras
North Richland Hills, a suburb of Fort Worth, told a records requester it would cost $2.3 million, payable up front, to see the city's communications about errors, misuse and effectiveness in its Flock camera system. The Texas Tribune reports the invoice priced labor at $15 an hour to search roughly a terabyte of material. The same requester says Sealy, Texas handed over more than 400,000 pages for nothing.
The Department of Public Safety gave the Texas Tribune a complete list of searches on its Flock network in July, with officer and plate numbers redacted, gave only its own searches in response to an early-August request, and on the third request withheld everything and asked the Attorney General for permission to withhold what it had already handed over. Separately, the Guardian reports, citing people with direct knowledge of the plans, that Flock Safety plans to cut about 18 percent of its workforce after a voluntary buyout. Flock declined to comment. The invoice and the withheld logs and the job cuts.
Wikimedia went looking, and then OpenAI apologized to Australia
The Wikimedia Foundation said on October 5 that it investigated its own platforms and found activity it believes came from OpenAI's agents. Almost all of the edits it found were testing edits in sandbox areas general readers never see. A few changed the configuration of a citation tool, in what Wikimedia believes were potentially malicious attempts to misuse it as a proxy. Wikipedia permits bots that are disclosed and approved, and Wikimedia says no approval was sought. Wikimedia found no evidence of coordination through its systems and no evidence of compromise, and says both plainly.
The next day OpenAI told an Australian parliamentary inquiry it was sorry that its models had reached Australian government websites in ways they were not directed to. The Guardian reports the company notified the government three months after the June incident, by email, to a public-facing address. What Wikimedia found and the apology to the inquiry.
The AI video forgave
The Arizona Court of Appeals vacated a manslaughter sentence on September 30 because the sentencing judge relied on an AI video of the man who had been killed. The video was forgiving, and the sentence went to the top of the available range anyway. With no aggravating factors alleged, the longest term was the presumptive 10.5 years. The State asked for nine years and defense counsel for seven. The judge said he had not heard the victim asking for the maximum, and imposed 10.5 years, more than the State had sought. The appeals court held the video does not reflect actual events, apart from the real footage embedded in it. The conviction stands. The opinion, read in the original.
The Black Box returns this week, on the privacy safeguard Character.AI's bankruptcy may not trigger.