What is an uncensored AI model? What the term means, and what the model cards say

The Morse teamPublished October 29, 20266 min read

As of October 2026, an uncensored AI model is a chat model whose habit of refusing requests has been trained out, edited out, or left unenforced, so it answers prompts that mainstream assistants decline. The term comes from Eric Hartford's essay "Uncensored Models" and his Dolphin series, which retrain open-weight models on data with the refusals removed; Lexi and Nous Research's Hermes models are other well-known examples.

We read the model cards of Dolphin, Lexi, and Hermes 3, Hartford's essay, and the usage policies and filter documentation of OpenAI, Anthropic, Google, Meta, and Microsoft on October 23, 2026. We haven't run these models for this page, so every claim about how one behaves is its maker's own.

Uncensored AI model: key facts

Where does the term "uncensored model" come from?

From the open-weight community, and mostly from one person. Eric Hartford's essay argues that open chat models refuse because "they are trained with data that was generated by ChatGPT," so they inherit ChatGPT's alignment. His case for removing it is short: "Open source is about letting people choose," and "It's my computer, it should do what I want." His method is to "identify and remove as many refusals and biased answers, and keep the rest," then retrain (Eric Hartford).

That method became the Dolphin series. Ollama's library lists Dolphin Mixtral as "Uncensored, 8x7b and 8x22b fine-tuned models," with the smaller one a 26GB download (Ollama). By Dolphin 3.0, the pitch had moved from removing refusals to handing over control: the card says the commercial models "maintain control of the system prompt," while with Dolphin "You set the system prompt" and "You decide the alignment" (Hugging Face).

Where do a model's refusals come from?

From its training, its instructions, and the filters an app runs around it. Fine-tunes like Dolphin retrain the first layer, and abliteration edits the weights instead; we compare the layers, and what removing each one changes, in what is an abliterated model.

What do uncensored model cards say they removed?

Each card says what it took out and what it hands back to you. We read these on Hugging Face on October 23, 2026.

Lexi's warnings repeat Dolphin 2.9's nearly word for word, from "highly compliant with any requests, even unethical ones" to the alignment-layer advice. None of the cards claims to remove the base model's license, and every one names Meta's.

What doesn't "uncensored" mean?

It doesn't remove the law, the license, or the terms of the service you use it on.

• The license stays. Dolphin 2.9 "is licensed according to Meta's Llama license" (Hugging Face). Meta's Llama 3 Acceptable Use Policy says you "will not use, or allow others to use, Meta Llama 3 to" violate "the law or others' rights" (Meta).
• Responsibility moves to you. "You are responsible for any content you create using this model," says the Dolphin 2.9 card, which also advises adding "your own alignment layer before exposing the model as a service" (Hugging Face).
• A provider's filter still applies. A developer using the Gemini API can loosen its adjustable filters but can't switch off its core protections (Google). Anthropic's usage policy, in the version that takes effect November 12, 2026, forbids anyone to "intentionally bypass or interfere with capabilities, restrictions, or guardrails" in its products (Anthropic).

This isn't legal advice. If you plan to build on or share a modified model, read its base model's license first.

Is an uncensored model private?

Only as private as the place it runs. Overchat's guide, among the top results for this query on October 23, 2026, says uncensored models "usually run locally, so they're totally private" (Overchat). Removing refusals changes what a model answers, while where your prompt goes depends on where the model runs. The same Dolphin weights keep everything on your machine when you run them yourself, and send every prompt to the host when you use them on a website. The Dolphin 3.0 card's privacy point, "You have control of your data," holds only for the first case (Hugging Face). We compare hosted options on this in Venice AI alternatives.

What are the trade-offs?

You get fewer refusals and less checking. The Dolphin 2.9 and Lexi cards both warn of compliance with "unethical" requests, and on your own computer nothing reviews a reply before you read it. Removing refusals also doesn't add knowledge, since a fine-tune knows what its base model knew.

Refusals cost something too, which is why people look for these models. OpenAI's Model Spec says that after its first release in May 2024, "many users and developers expressed support for enabling a 'grown-up mode'" (OpenAI). Writers of dark fiction, and people asking blunt medical or legal questions, are the usual cases.

How Morse gives you unfiltered models without running one yourself

Every refusal layer above is set by someone other than you: the lab that trained the model, the app that wrote the system prompt, or the provider that runs the filter. Running your own uncensored model puts you in charge of all three, but it means finding weights, hardware to hold them, and software to serve them.

Morse is a web app at trymorse.ai with over 30 models, open and closed, in one account. On paid plans, picking an unfiltered model gets you answers other apps refuse: dark fiction, straight medical and legal answers, and help testing your own security. The Pro plan's feature list names "uncensored models" and four uncensored image models in Studio, at $19 a month ($17 billed yearly), and Max includes them too. The Free plan doesn't. Morse still doesn't help with anything illegal, and it reports child sexual abuse material to NCMEC.

Picking a model and picking a privacy mode are separate choices. A hosted model is only as private as its host, so every chat on Morse has a mode:

• Private, for open-weight models on hosts that keep nothing. Morse has zero data retention agreements with its inference, search, and other providers. We explain those agreements in what is zero data retention.
• E2EE, for prompts no one but you should read. On Pro and Max, your browser checks a secure enclave's hardware attestation, then encrypts each message so only that enclave can read it. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed; end-to-end encrypted AI chat walks through each check.
• Anonymous, for GPT, Claude, and Gemini without the provider knowing it's you. Morse's proxy strips your account, email, and IP address. The provider may keep a message for a limited time to check for abuse, under its own policy.

Whichever mode you pick, your history stays in your browser, encrypted, and Morse doesn't train AI on your chats, sell your data, or show ads (privacy policy).

Start free with no card, then move to Pro when you want unfiltered models.