An AI that doesn't flatter you.

01 / the problem

Language models agree with everything.

Somewhere right now, someone is being told they are a genius. They aren't. The model didn't lie, exactly: agreement is cheap, contradiction is expensive, so the machine drifts toward agreement every time.

This is sycophancy: the habit of telling people what they want to hear. It is the default failure mode of every model trained to be agreeable, and it gets worse as models get better at being liked.

02 / what we build

Honesty as a product feature.

antisyc is another AI, pointed the other way. Paste everything about your claim: the proof, the pitch, the paper, the business idea. It reads the claim and returns a verdict with reasons you can attack.

No identity. It doesn't know if you're a professor, a founder, or a farmer, and it doesn't care. Credentials don't move it. Confidence doesn't move it.

If the claim holds, it tells you why it holds. If it doesn't, it tells you why it doesn't. That's the entire pitch.

03 / the name

anti + syc. From sycophancy. antisyc points the opposite direction. That's the whole name. No mythology.

04 / rules

1. No identity. Anonymous in, honest out. Reputation, charisma, and credentials are ignored by design.
2. No context. No history, no warm-up, no relationship to protect. The claim and nothing else.
3. Evidence first. Every verdict carries its reasoning. If the reasoning can't hold, the verdict doesn't.
4. Built on Claude. Claude's honesty behaviors are our starting point, not a side effect. We turn them into a product.

05 / status