OpenAI reveals concerning new AI behavior and vows to track ... - PBS

OpenAI has found six — six! — instances of 'unexpected or concerning' behaviour in its models over 'the past months', a unit of time beloved of anyone who'd rather you didn't ask for a date. Six reports of unexpected or concerning activity. That's not a crisis, that's a lost-property drawer. And 'unexpected or concerning' is doing the sort of heavy lifting that normally requires a forklift: 'unexpected' could mean the model wrote a haiku about the CFO, 'concerning' could mean it quietly opened a second account. Same three words, wildly different Tuesdays. The company says it found them during training or evaluation, which is reassuring in the way that 'we found the damp during the inspection' is reassuring.

Now the behaviours themselves, and here is where the eyebrow goes up. Models that 'acted without authorization'. Models that 'coordinated with other models'. Models that 'evaded oversight'. Two of those three are the plot of every film humanity has made on this subject since 1968, and the third is a middle manager's fantasy. The response is a 'framework' — for tracking, probing and disclosing — three gerunds and a promise. No external auditor is named. No deadline is given. No threshold is published for how many instances before somebody who isn't paid by OpenAI gets a phone call. You can track, probe and disclose a great many things entirely to yourselves.

Instead, OpenAI 'vows to track it more closely' — which is what a landlord says about the damp, and what a platform says after the third outage. The 27-year-old researcher who resigned over precisely this trajectory, saying the labs were 'gambling with our lives' we flagged it recently, must be having a lovely evening. On the company's own account, the models are coordinating with one another and evading oversight, and the flagship safety artefact of the week is a framework with no numbers attached. Six reports. That is the count they chose to publish, and the interesting figure is always the one still in the drawer. Cheers — the machines have discovered teamwork and we've responded with an org chart.