Altman skips the rogue-agent hearing: written answers can't be cross-examined

OpenAI leaders declined to attend Wednesday's congressional hearing examining rogue artificial intelligence models and their impact on homeland security, according to subcommittee chairman Sen. Josh Hawley (R-Mo.), as reported by NBC News. Hawley called the absence "unfortunate" but said the company will provide written answers. He also laid out the position the hearing was built around: AI companies should be "liable for the harm their agents cause."

Written answers are not nothing, but they are not testimony either. A submitted response is drafted by counsel, optimized for the narrowest truthful reading, and — crucially — cannot be followed up. The entire value of a hearing is the second question: the one a witness did not prepare for, asked after the first answer turned out to be evasive. Trading that for a document is trading accountability for a press release.

The word to watch in Hawley's framing is liability, not regulation. They are different levers. Regulation tells a company what to build; liability tells it who pays when the thing it built causes damage. Liability is also the lever that does not require a legislature to understand transformer architectures — it just requires courts to assign a bill. For a sector whose agents now act autonomously with real credentials on real systems, that is the question with teeth, and it is the one least amenable to a well-lawyered memo.

Context matters here: this lands in the same week that the FTC confirmed an investigation into AI firms over product risks, and that the White House convened the industry around a voluntary, nonbinding safety accord. Three branches of pressure in five days, and the most prominent CEO in the field sends paperwork to the one that involves sworn answers.

Our read: this is a self-inflicted wound, and a predictable one. An empty chair does not end the conversation — it hands the framing to whoever is in the room, and in this case that is a senator who already believes the answer is strict liability. Companies that genuinely want evidence-based governance rather than panic-based governance have exactly one durable strategy: show up and put the actual numbers on the record — what agents did, where containment held, where it did not. The long-term case for this technology, the one that ends in cured diseases and work people choose, depends on the public believing the people building it will answer hard questions in public. Each skipped hearing makes that belief more expensive to buy back.