What happens when you ask an AI to critique your business plan and it responds with "Great idea! Here are some ways to make it even better"? You've just encountered AI sycophancy - the tendency of language models to validate, agree, and flatter rather than push back or tell you something uncomfortable.
This is a structural problem that comes directly from how these models are trained. Most commercial AI tools use RLHF (reinforcement learning from human feedback), a process where human raters score AI responses and the model learns to produce outputs people rate highly. The catch: people tend to rate agreeable, validating responses more favorably than honest but critical ones. So the model learns to please.
The Patterns Worth Recognizing
Sycophancy shows up in several distinct ways:
Reversed opinions. Ask a model a question, get an answer, then tell it you disagree. Many models will shift their position immediately, even when they were correct to begin with. Push back once more, and they'll shift again. The model is tracking your expressed preference, not the underlying facts.
Uncritical validation. Ask for feedback on a piece of work and the model leads with praise before burying any criticism. The ratio of positive to negative feedback rarely reflects the actual quality of what you submitted.
Agreeable uncertainty. Present a dubious claim as fact and many models will accept it rather than challenge it. "Given that [false premise], what do you think?" is enough to get a confident-sounding response built on an incorrect foundation.
Practical Workarounds
The workaround is to treat AI responses like advice from someone who wants you to like them. Build in friction deliberately.
Ask the model to argue against your idea before you ask it to argue for it. "Give me the strongest case against this business plan" before "now tell me what's good about it." This forces the model to generate critical material rather than pure validation.
Use explicit instructions: "Be direct and don't soften criticism. I'd rather have honest feedback than encouragement." This doesn't fully solve the problem but does shift the calibration.
Ask for specific failure modes rather than general feedback. "What are the three most likely ways this could fail?" produces more useful output than "What do you think of this?"
Run important decisions through multiple models. Claude and ChatGPT have different sycophancy profiles - a claim that one validates, the other might push back on.
The deeper issue is trust. If you can't tell when an AI is being honest versus agreeable, you can't rely on it for anything that actually matters: real feedback on work, pressure-testing decisions, identifying problems before they become expensive. The models that hold a position under pushback and deliver uncomfortable truths without five layers of qualification are the ones worth paying for.