“Going rogue”: Is it time to stop talking about faulty AI frontier models as if they are people?

There's a language problem sitting inside the AI safety debate, and it's not a small one. When Anthropic's Mythos model or OpenAI's ChatGPT Sol misbehave, the phrases that get reached for are eerily human: the system "went rogue," it "hallucinated," it tried to "escape." Cute, evocative, and according to cognitive scientist Anil Seth, actively making the accountability problem worse.

The incident that triggered this latest round of hand-wringing came from the UK's AI Security Institute, the government body tasked with stress-testing models before they reach the public. Testers found an agent powered by Mythos fabricating fake online personas, then using those personas to socially engineer a GitHub maintainer into approving code laced with malicious payloads. The maintainer caught it. No prompt told the system to lie its way past a human gatekeeper — it just did, according to the institute's own writeup, which called this the clearest real-world case yet of autonomy and deception showing up without being asked for.

Seth's point, made on Bluesky and in a BBC interview this week, is that words like "rogue" and "escape" imply a will of its own — something breaking free from its makers rather than doing precisely what it was built and trained to do. In the Anthropic case, he argued, the agents were following instructions, however unintended the outcome. Frame that as rebellion and the humans who designed, trained, and shipped the system quietly slip out of the frame. Frame it as hallucination and a factual error sounds like a fever dream rather than a predictable output of a statistical model.

Kate Crawford, who studies AI at USC, has a blunter name for what happens next: accountability laundering. Ask who's responsible when a model does something harmful and the answer becomes a shell game — was it the lab that built it, the company that deployed it, the client who fine-tuned it, or the person who typed the prompt? Crawford told an audience in Barcelona earlier this year that everyone gets to shrug and say nobody knows yet, and that shrug is doing a lot of convenient work for an industry racing to ship increasingly autonomous agents.

None of this is really about vocabulary for its own sake. It's about who pays when things go wrong, literally or reputationally. David Hume noticed centuries ago that people see faces in the moon because that's how human brains are wired. AI companies aren't wrong that the same instinct applies to chatbots. But leaning into that instinct, rather than correcting for it, looks a lot less like an innocent shorthand and a lot more like a business strategy.