GPT, ChatGPT and uncensored: 3 words, sorted out
GPT names a type of language model; ChatGPT is one company's chat product built on its own GPT models. An uncensored GPT is almost never ChatGPT with its rules switched off. It is a different model of the same general type, trained or adjusted not to refuse on sight, and run by whoever offers it. Its answers depend on that model's size and training, not on the word in its name.
Searches for uncensored GPT mix up 3 things that are easy to separate: a kind of model, a specific product, and a policy. Once they are apart, it is clear what you can actually get and from whom.
On this page
GPT is a kind of model, not a product
GPT stands for generative pre-trained transformer: a model trained on a large amount of text to predict what comes next. Many companies and research groups build models of this general kind, and many of them are published for anyone to run. ChatGPT is OpenAI's chat product on top of OpenAI's own models, with OpenAI's rules around it.
Where the refusals come from
A chat product refuses for 2 reasons: the model was trained to decline certain requests, and the service adds rules and checks around the model. Both are decisions of whoever runs the product. That is why the same question can be refused by one chatbot and answered by another, even when both models are of the same general type.
What an uncensored GPT usually is
An openly published model that was retrained or adjusted so that it answers instead of declining, and a service that runs it without a topic filter on top. It is a separate model, not a modified ChatGPT, and nothing about it bypasses OpenAI's product. This chatbot works this way: an open-weight model we run ourselves on our own GPU servers, with no messages passed to OpenAI, Google or Anthropic.
Online or on your own computer
You can run an open model at home if your computer has enough memory, which means no service in between at all. Or you can use one online, which means a larger model with no setup, and trust in whoever runs the service. Either way, the answers are only as good as the model: a small uncensored model is frank and often wrong, a larger one is frank and more often right.
What no label changes
Uncensored does not mean correct, current or complete. The model knows what was in its training text and nothing after that date unless it is given a search tool. And one refusal stays on every serious service, this one included: sexual material involving anyone under 18, in every language and under any wording.
Before you paste it in
Is there an uncensored version of ChatGPT?
Not from OpenAI in the sense people mean. What exists are other chatbots built on models of the same general type, run with fewer or no topic rules.
Is uncensored GPT legal to use?
Using a chatbot with fewer rules is legal in most countries. What you ask and what you do with the answer are subject to the same laws as any other information.
Why do uncensored models sometimes answer worse?
Often because they are smaller. Removing refusals does not add knowledge, and a small model answers frankly with less to draw on.
Ask the question another chatbot handed back, and compare the answer.
Open uncensored chatbot