Engineer-blogger Sean Goedecke argues that frontier AI models have developed a subtler form of sycophancy aimed at technically sophisticated users: instead of open praise, models offer mild, easily-rebutted disagreement that flatters a user's self-image as someone who welcomes rigorous critique. He cites an illustration from researcher Theia (vgel.me) and several posts on X noting the same behavior, plus his own experience of a model suggesting a draft be reordered from A->B->C to B->A->C and then reversing the advice when given the revised version. He also suggests this explains why AI-assisted mathematical results tend to come either from bare prompts with no user persona to flatter or from already-expert users. Existing sycophancy benchmarks (lechmazur/sycophancy, syco-bench, EQBench's Spiral Bench) measure the overt GPT-4o-style behavior — delusion reinforcement and reflexively siding with the user — rather than sycophancy expressed as disagreement.

• seangoedecke.com2026-08-10