Not a different model — the same model, minus the layer that was trained to say no.
Commercial models like ChatGPT or Claude go through an alignment stage after initial training, where human feedback teaches the model to refuse certain requests and soften certain answers. “Uncensored” models are versions — often open-weight models fine-tuned by third parties — where that refusal layer has been deliberately stripped out or never added, restoring the base model’s willingness to answer without the guardrails. Because open-weight models can be downloaded and modified freely, this has become common enough that a 2025 academic survey specifically tracked the growing ecosystem of publicly available and locally deployable uncensored models.
This is what this essay means calling them “fun to experiment with” while running locally — the trade being made is the same guardrail removal the essay elsewhere calls “bias,” except here it cuts toward fewer refusals rather than fewer distortions.
Source: Sokhansanj, B.A. (2025), “Uncensored AI in the Wild: Tracking Publicly Available and Locally Deployable LLMs,” Future Internet, 17(10), 477.