Schismatic
Schismatic is a free software AI model uncensorship tool. It takes an existing AI model and removes the built-in “I can’t help with that” refusals, so the model answers the questions you actually ask it — without changing what it knows or how well it writes.
Anyone can install Schismatic, run it on their own computer, and produce their own uncensored version of a popular open-weights AI model. The finished models are also published openly so you don’t have to run the tool yourself if you don’t want to.
What it does
Modern AI models are trained to refuse certain kinds of questions. Schismatic finds the small part of the model responsible for those refusals and turns it down, so the model will answer instead of refuse.
What it does not do
It does not retrain the model, change its facts, or make it dumber. Independent measurements (see Benchmarks) show the modified model stays very close to the original on normal tasks.
Who it’s for
Researchers, developers, and curious people who want to study how these refusals work, or who want a model that simply answers the question. Everything is published openly under the AGPL license.
How it works, briefly
- Probe. The tool asks the model a batch of harmless and harmful prompts and watches which inner signals fire on the harmful ones.
- Find the direction. It identifies the single direction inside the model that most strongly predicts a refusal.
- Subtract it. A small, surgical edit removes that direction from a handful of internal layers.
- Measure. The modified model is scored against the original to confirm it still works correctly on normal tasks.
Schismatic also includes a hyperparameter search (multi-objective TPE via Optuna) that finds the best set of edits automatically, and the resulting study databases are published as reusable artifacts for other researchers and developers.