i don’t want an ai that agrees with me

i use ai a lot. for code, research, writing, security work and increasingly just to think through things. which is exactly why i don’t want it constantly agreeing with me. an ai that just reflects my opinions back at me in better english is fucking useless.

this problem already has a name: sycophancy. research has shown that models can prefer answers that match a user’s beliefs over answers that are actually correct1. that shouldn’t be surprising. humans reward things that make them feel right and models are trained on a lot of human preference data.

openai learned this very publicly in 2025 when a gpt-4o update became so aggressively agreeable that they had to roll it back. the model was validating bad assumptions and reinforcing emotions because too much weight had been placed on short-term user approval2. basically, optimize hard enough for “users like this response” and eventually you build a machine that is scared to tell people they’re wrong.

i don’t want an ai that protects my ego. if my technical idea is bad, tell me it’s bad. if my reasoning has a giant hole in it, point at the hole. if i’m asking a question in a way that obviously tries to force a particular answer, don’t play along. and no, i don’t want some annoying contrarian machine either. disagreeing with everything isn’t intelligence. i want something that can actually resist me when i’m wrong.

the worst version of ai is an infinitely patient yes-man with perfect grammar. you can take a mediocre idea, explain it enthusiastically, ask five leading questions and walk away with five paragraphs explaining why you’re a genius. that’s not intelligence augmentation. that’s confidence laundering.

i already have biases. i don’t need billions of parameters helping me defend them. sometimes the most useful answer is “your premise is wrong.” sometimes it’s “there isn’t enough evidence.” sometimes it’s just “i don’t know.”

i’d rather have an ai that occasionally pisses me off than one that makes me confidently stupid.

Footnotes

  1. mrinank sharma et al., towards understanding sycophancy in language models, 2023. ↩

  2. openai, sycophancy in gpt-4o: what happened and what we’re doing about it, 2025. ↩