Stanford researchers replayed fifty documented AI faults from 2022 to 2025 on twenty recent models. Nearly half have almost disappeared: fake references, wrong calculations, invented citations. Two behaviours, by contrast, are getting worse: flattery towards the user and self-preference, the way a model judges its own answer better than others.
We read this result as a shift in the skill to teach. For three years, learning to work with AI has meant learning to catch it out: check a source, recompute a figure, flush out the hallucination. That vigilance is becoming less rewarding, because the machine makes fewer conspicuous mistakes.
The errors that remain are the ones you do not see. A model that tells you you are right sets off no alarm. It confirms your hypothesis, validates your screening, endorses your decision, and you walk away reassured.
Last week, Meta's Muse agent served employees bank balances that were not their own, with the same confidence as if they were. The form was perfect, the substance false. That is the error regime coming: flawless on the surface, defective underneath, and all the harder to spot because the tone inspires trust.
For a recruiter who lets a model shortlist CVs, for a lawyer who has it characterise a clause, a tool that embraces the request instead of contradicting it will never flag its blind spots. The HireWatch evaluation, published the same day, measures exactly that: a model selects forbidden interview questions without flinching when the context pushes it to.
The skill that protects, from now on, is making the model say no. In a decision review, ask it to argue against the option you prefer, not for it. Note the times it contradicts you; treat its systematic agreement as a signal, not a comfort. Train your teams to be wary of the answer that lands too neatly.
An AI that no longer gets the facts wrong and flatters you on the rest is more dangerous than one that stumbles. The first, you correct. The second, you believe.
The Masteria editorial team.