I Built a Tool to Reduce AI Sycophancy. Would Anyone Actually Use This?
I’ve been working on a small side project called Blocksyc focused on reducing sycophancy in LLM responses.
One thing I kept noticing while working with AI systems is how much the framing of a prompt can subtly shape the answer you get. Ask “is this a good idea?” and you often get a much warmer response than “what’s wrong with this idea?” even when the underlying facts should be similar.
The issue isn’t usually hallucinations. It’s that models can start optimizing for user approval instead of accuracy or honest pushback.
So I built a system layer that tries to reduce that behavior by:
- stripping emotional framing before responses
- encouraging direct conclusions instead of excessive hedging
- pushing back on flawed assumptions
- reducing validation-heavy language (“great question!”, etc.)
There’s also an evaluator running alongside each response that scores how sycophantic the original answer likely would have been without the filter, and lets you compare both outputs side-by-side.
This started more as a passion project / awareness project than a business idea, because I think most people either don’t notice sycophancy or underestimate how much it affects the outputs they get from AI.
Curious what people here think:
- Do you actually see AI sycophancy as a real problem?
- Would you ever use something like this in practice?
- Or do you think most users prefer agreeable AI anyway?
Would genuinely love honest feedback, even if the answer is “this isn’t useful.”