The incapacity defense

Thursday was a split screen for OpenAI. In one window: GPT-5.6, three new models, and a rebranded ChatGPT Work — an agent that runs jobs in the cloud for hours and, with permission, reaches into your local files and desktop apps. In the other: a sanctions motion from The New York Times and its co-plaintiffs alleging that OpenAI spent two years telling a federal court it lacked the technical ability to search its own chat logs.

The two windows are one story. The motion isn't really about redactions and meet-and-confers; it's about whether an AI company's statements about what its systems can do mean anything once capability becomes inconvenient. OpenAI's answer, per the filing, was to treat capability as a claim priced by who's asking.

The specifics came out of an April deposition of OpenAI privacy engineer Vincent Monaco. Per the motion, OpenAI held two large de-identified log samples — 10 million and 78 million conversations — and disclosed neither for two years while pleading that searches of this kind were infeasible. It had already run them. OpenAI searched those logs for Times content while building its own filter to block regurgitation of copyrighted text — capability it found readily enough when the output was a product feature instead of an exhibit.

Meanwhile the plaintiffs got the burdened version. Eight months in a sandbox with a 20-million-log sample — a sixth of what they asked for — to which OpenAI applied 19 billion AI-generated redactions, so many the court declared the sample unusable. And when a preservation order required OpenAI to retain user chats, Monaco testified the company concluded compliance would be hard, and thus took no steps to do so. Not appealed and lost. Decided, internally, that hard meant optional.

Take OpenAI's defense at full strength, because it isn't fake. The company cast this fight as protecting users — its spokesperson calls the motion "blatantly false allegations" and an effort to "invade the privacy of people who have nothing to do with this case." In principle that objection is real: a court order preserving every deleted ChatGPT conversation is genuinely invasive, and resisting it isn't paranoia. But per the motion, the argument wasn't OpenAI's to make. It had already de-identified 78 million conversations and mined them for its own engineering. A privacy shield raised only over data you've finished using isn't a principle. It's an inventory strategy.

A privacy shield raised only over data you've finished using isn't a principle. It's an inventory strategy.

The plaintiffs want the 20-million sample barred, an adverse finding that the withheld logs contained substantial regurgitation, and a jury instruction that OpenAI deleted billions of conversations. Whether the judge goes that far, the deposition already did its damage: it fixed a price on OpenAI's self-descriptions. The market prices what OpenAI says its models can do. Discovery prices what it says they can't. Thursday produced a reading of both — and only one of them was under oath.