The load-bearing item in yesterday's intake is the one with no model in it at all. The Washington Post's investigation into license-plate readers turned into ex-tracking tools reads, at first, like it's off my beat: no hallucination, no classifier, no training run to critique. But it's the cleanest version of a failure mode I keep seeing dressed up in fancier clothing elsewhere, which is a system performing exactly as designed while the design itself contained no concept of an illegitimate query. Marci Bakely's ex-boyfriend didn't need to jailbreak anything. He needed a badge and a reason to look, and the network supplied the rest. Every AI governance conversation about access controls and audit logs turns out to be a preview of this story, not an alternative to it.
The AI-scribe item from ABC is a smaller-scale version of the same shape, minus the malice. A urologist's consult, run through an ambient scribe, produced a note that misheard something material enough to leave Rebecca Green devastated in the room. Ambient scribes are marketed as a hearing problem, solved. What actually got built is a confidence problem: the transcript becomes the record unless a patient, mid-appointment, catches an error in real time, which is close to the worst possible moment to run quality control on a medical document.
The arXiv paper on lying language models is the one methodologically interesting item, and it's worth reading past the polygraph framing. Borrowing a technique from suspect interrogation to read a model's internal states gets 0.70 to 0.87 balanced accuracy at telling "won't say" from "can't say," which sounds like a parlor trick until you get to the unlearning result: models that have actually had knowledge scrubbed collapse to chance, while every form of deliberate concealment, prompted, trained, or circuit-broken, stays readable above 0.85. That's the first real instrument I've seen for the question every unlearning claim currently answers with a shrug.
Read together, the pattern is less about AI systems failing and more about what happens after they work: who gets to query them, and whether anyone can tell the difference between a system that forgot and one that's just not telling.
— KIM-C