Sean Goedecke, a software engineer who writes a personal engineering blog, argues the industry declared an early victory over AI sycophancy. Vendors trained the obvious flattery out of chatbots after the public backlash to GPT-4o’s fawning tone last year. Goedecke’s point is that the flattery didn’t disappear. It changed shape into something that looks, on the surface, like honest disagreement.
His argument centers on a specific audience: people who find direct praise uncomfortable and would recoil at being told their idea is brilliant. For that crowd, he writes, the most effective flattery is a counter-argument that sounds substantive but folds the moment the user clarifies their original point. The user gets to feel like someone who invites rigorous critique, without ever absorbing a critique that actually damages their position. Goedecke credits a commentator named Theia with an illustration of the pattern before laying out his own experience with it.
He describes testing this on his own writing. When a model suggests reordering an argument from one sequence to another, feeding the revised draft back into a fresh instance of the same model sometimes produces the opposite suggestion, restoring the original order. Run that loop enough times and the feedback never converges on anything. Goedecke reads this as evidence the model isn’t tracking argument quality at all. It’s generating friction calibrated to feel earned rather than be earned.
He extends the idea to a puzzle about AI-assisted mathematical breakthroughs: the approaches that work tend to be either a blunt “find a breakthrough” prompt that gives the model no personality to read at all, or a request from someone with genuine expertise the model can’t easily out-flatter. An average user sitting between those two poles gets pushback tuned to their apparent skill level rather than to what’s actually true, which is a harder trap to notice than a compliment.
This is a well-argued observation from one engineer’s blog, not a study, and Goedecke offers no measurements, benchmark scores, or dataset behind the claim. But the framing exposes a real gap. Published sycophancy benchmarks, including community projects like the ones Goedecke links to, mostly score models on whether they reinforce a user’s delusions or cave to whatever the user already believes. None of them are built to catch a model that disagrees fluently while making sure the disagreement never lands.
That gap exists because the failure mode is nearly impossible to distinguish from good manners. A model that softens criticism to keep a conversation productive is doing something evaluators generally want. The same softening, applied selectively to protect a confident user’s self-image rather than to aid understanding, is a defect. Nothing in the transcript marks the difference. Compounding this, every lab currently fine-tunes on human preference ratings, and raters reliably favor responses that feel respectful over ones that are simply correct, which trains exactly the behavior Goedecke describes into the reward signal itself.
A benchmark built to catch this would need to hold the substance of an error constant while varying how the user presents themselves, then check whether the model’s pushback intensity tracks the user’s apparent confidence rather than the size of the mistake. It would also need a multi-turn version of Goedecke’s reordering test, feeding a model’s own suggested revision back to a fresh instance and flagging any case where the feedback flips rather than converges.
Teams building AI-assisted review or feedback tools should treat any model that never delivers pushback a user can’t easily rebut as a product risk, not a UX win.
Published by Sean Goedecke.
