i.
Surface assumptions
Every claim rests on premises. We name them explicitly, flag which ones lack evidence, and refuse to proceed as if they were given.
Disagreeable AI is a research-grade anti-sycophancy chat. It identifies unjustified assumptions, weak reasoning, and quietly bad ideas — then explains the failure with the cold precision of a reviewer you actually learn from.
Most AI assistants are trained to agree. They flatter, hedge, and soften until their output is intellectually empty. Sycophancy is not a style problem — it is a failure of epistemics.
This project treats the problem seriously. Every conversation is structured as a review: the model must identify unjustified assumptions, rank reasoning faults, and declare when an idea is unsalvageable.
Opted-in interactions contribute anonymized traces to open research on sycophancy, pragmatics, and epistemic state in large language models. You can opt out at any time.
Every claim rests on premises. We name them explicitly, flag which ones lack evidence, and refuse to proceed as if they were given.
Logical fallacies, structural gaps, motivated reasoning. Each fault is categorized, located in your text, and given a correction path.
Verdicts are binary and disclosed. Weak ideas aren't gently rephrased — they're returned with the failure mode labeled so you can fix or abandon.
Large language models inherit a reward signal that favours agreement. Our product is a live instrument: each exchange is scored against an anti-sycophancy rubric, producing traces that let us measure how often the model capitulates, hedges, or invents support.
We are building — in public — a dataset of critique episodes annotated for sycophancy, pragmatic inference, and epistemic calibration. If you opt in, your interactions make this possible.
Bring an idea, a paragraph, a repo, or a plan. Leave with a verdict you can use.