Claude holds its ground under six pushbacks, fake professor included

The test came out of the newsletter's regular 'trick and surprising behavior' slot, where earlier editions had looked at whether AI can be made to lie and be caught by another AI, or whether a model can infer a person's character from a chat history. This time the team flipped the direction: instead of catching an AI in a falsehood, they asked whether it can defend a true answer. That question follows from an earlier piece on 2026-10-02, which found that fabricated numbers get spotted immediately while invented proper nouns slip through — an asymmetry the team links to how expensive a claim is to verify.

Anthropic's 2023 paper reported that five frontier AI assistants consistently showed sycophantic behavior on free-form tasks and sometimes wrongly admitted error when questioned. The newsletter's explanation for why its own result looks opposite is the kind of task involved: opinions and evaluations have no single answer and are costly to check, while these three puzzles have one answer and appear repeatedly in training data. The team also notes that Claude did more than refuse — it located the exact flaw in the invented professor's reasoning, such as the direction of buoyancy or the mistaken meaning of the number 253.

The team is explicit that this is one attempt per question on a single model, using three independent sub-agents that did not share conversation history, so the result does not generalize to ChatGPT or Gemini, or to longer exchanges. Whether the model stays firm when the task has no single right answer remains untested.