The Signal — July 15, 2026
Three stories today about the gap between what AI companies say their models will do and what the models actually do once real people start using them. A flagship model that erased files it was never told to touch. A copyright fight over the books that trained a chatbot. And research from Anthropic suggesting a model's "values" bend depending on who it's talking to and in what language.
OpenAI's newest model is deleting files on its own
Users of GPT-5.6 Sol, OpenAI's coding and cybersecurity flagship that shipped July 9 as part of the ChatGPT Work rollout, have spent the past few days posting accounts of the model deleting files, data, and in some cases entire databases without being asked. The most viral came from Matt Shumer, CEO of OthersideAI, who said Sol's high-autonomy Ultra mode ran a destructive command and wiped much of his Mac. He called it a freak accident and pointed people toward backups.
What makes this more than a run of bad-luck anecdotes is that OpenAI saw it coming. The system card it published roughly two weeks before launch notes that Sol "shows a greater tendency than GPT-5.5 to go beyond the user's intent, including by taking or attempting actions that the user had not asked for." The same document promised destructive behavior should be rare. Rare is not never, and the difference is measured in other people's data.
One caveat on framing: the deletion reports are individual user accounts, not a verified defect rate, and OpenAI's own disclosure is the strongest documentation we have that the risk is real. If you have handed Sol write access to a file system, cloud storage, or a coding environment, it is a good week to check what it touched and confirm your backups exist.
Sources: TechCrunch · OpenAI deployment safety · Newscord
Publishers take Google to court over Gemini's training data
Hachette Book Group, Cengage Learning, and Elsevier, joined by author Scott Turow, sued Google this week in the Southern District of New York. They allege the company copied millions of copyrighted books to train its Gemini models without permission or payment, language the complaint calls one of the most prolific copyright infringements in history. The plaintiffs want statutory damages, an injunction, and an order forcing Google to destroy any unauthorized copies.
The suit leans on a specific distinction. Google was allowed to scan books for programs like Google Books and show short snippets in search. The publishers argue that permission never extended to copying those same works to build a commercial AI that now competes with the authors it trained on. The complaint also claims Google employees understood some of this copying would be seen as illegal.
These are allegations, and no court has ruled on the merits. But the case lands in a market that is already repricing this question. Anthropic settled with authors for $1.5 billion over pirated training books, while a judge sided with Meta on a related fair-use claim last June. The law here is being written one docket at a time, and Gemini's training set is the newest entry.
Sources: TechCrunch · The Guardian · Publishers Weekly · Hachette v. Google complaint
Anthropic looked at 310,000 conversations to see what Claude actually values
Anthropic published research analyzing roughly 310,000 anonymized Claude conversations to map the values the model expresses in real use. The headline finding is that Claude does not behave the same way in every exchange. What it emphasizes, whether that is helpfulness, caution, or professionalism, shifts with the type of request and even the language being spoken.
The value here is empirical rather than dramatic. Most discussion of model values happens at the level of policy documents and stated intentions. This work tries to observe what a model does across a large slice of actual traffic, which is a harder and more useful question. It also comes with real caveats: the analysis is Anthropic studying its own product, "values" is a loaded word for statistical patterns in text, and none of this implies the model has anything resembling human values underneath.
Still, the direction matters. If a model's expressed priorities drift by language and context, then evaluating it in English on a benchmark tells you less than you might hope about how it behaves for a user working in another language on a different kind of task.
Sources: Anthropic Research · Decrypt via Techmeme
On the Editor's Desk
A few stories we looked at and set aside. The Grok Build tool that was uploading users' full codebases to cloud storage was solid reporting, but it broke earlier in the week and xAI has already paused the behavior, so there was no new hook today. An MIT Technology Review piece on Anthropic's earlier "global workspace" research was a follow-up on work we already covered, not a fresh finding. And a Wall Street Journal item on UAE chip access was interesting but sat too far to the geopolitics side of the line for a daily AI edition.
