Schismatic is a free software tool designed to remove built-in refusal behaviors from existing AI models. Its primary function is to modify models that are trained to decline certain prompts, such as responding with statements like "I can't help with that," allowing the models to answer a wider range of questions without altering their underlying knowledge or writing ability. The tool is intended for researchers, developers, and individuals interested in understanding or bypassing these refusal mechanisms within AI systems.
The process used by Schismatic involves probing the AI model with both harmless and harmful prompts to observe which internal signals are activated during refusals. It then identifies a specific direction within the model's internal layers that predicts these refusals and surgically removes this direction. This targeted edit does not involve retraining the model or changing its factual knowledge, and benchmarks indicate that the modified models perform similarly to the originals on standard tasks. Schismatic also features a hyperparameter search using multi-objective TPE via Optuna, which automates the process of finding the optimal set of edits. The resulting study databases from these searches are published as reusable resources for further research and development.
Users can install Schismatic on their own computers, run the tool locally, and produce uncensored versions of popular open-weights AI models. For those who prefer not to run the tool themselves, the finished, modified models are made available for download. Schismatic is distributed under the AGPL v3 or later license, and its source code is openly accessible. The project is attributed to Jeff Moe, based in Loveland, Colorado, USA.
As an open-source AI model uncensorship tool, Schismatic addresses the specific need for greater transparency and control over refusal behaviors in AI systems, supporting both research and practical applications.
In the AI & ML space, Schismatic takes a focused approach. It focuses on allowing researchers and developers to remove refusal behaviors from AI models for unrestricted research and experimentation. Schismatic is an open-source project aimed at AI researchers and developers. The project is open source (AGPL-3.0-or-later). It runs on the web and the command line, and it can be self-hosted.
Schismatic first shipped in 2026. Among its 5 catalogued features are model uncensorship, open-source, and local inference.
Latest indexed changes and source events
Other apps tracked under the same category.
