AI without restrictions: what it means, and the limits every chatbot still has
Is there an AI without restrictions? What 'no restrictions' can mean for mainstream, uncensored and local models, and the limits every chatbot still has.
By Unrestricted team · Editorial standards
No AI has zero restrictions. Every chatbot is shaped by how its model was trained, what the company running it adds on top, the rules that apply to it (terms, license and the law), and usage limits. In practice, "AI without restrictions" means a model that refuses fewer legal questions, not one that refuses nothing.
The same goes for an AI you can ask anything: you can type any question into any chatbot, and what differs is how many legal questions get a straight answer. Below: the four layers, how three kinds of AI compare on each, what an uncensored model still won't do, and what to ask before you pick one.
The four layers that restrict any chatbot
Four things decide what any chatbot will say: how the model was trained, what the host adds on top (a system prompt, classifiers, safety settings), the rules that apply (terms, license and the law), and usage limits on how much you can use it and how fast. How each one turns into a refusal is covered in how AI content filters work.
An uncensored model strips out much of the first layer. A service that adds no filter of its own keeps the second thin, though whoever serves the model can still refuse some prompts. Nothing removes the third. The fourth depends on where the model runs.
Mainstream, uncensored hosted and local models compared
"AI without restrictions" gets used for three different things: a mainstream assistant with looser settings, a hosted chat built on an uncensored model, and an open-weights model you run on your own computer. Here is where each layer sits for each. The sources are the companies' own documents, the model licenses and the research on uncensored models; the labeled example comes from our Terms and our own product facts.
| Layer | Mainstream assistants (ChatGPT, Claude, Gemini) | Uncensored hosted chat (example: Unrestricted) | Open-weights model you run yourself |
|---|---|---|---|
| Training | Safety training in the weights. OpenAI says it is training its models to align to its Model Spec (while noting production models don't yet fully reflect it), whose red-line principles rule out facilitating critical and high-severity harms [1]. Anthropic says it trains Claude to refuse queries involving the production of biological or chemical weapons [2]. Google's Gemini API docs refer to "the model's inherent safety." [3] | One uncensored open-weights model based on Dolphin-Mistral-24B. | Whatever you download. An official instruct model keeps its safety training. Community uncensored versions are retrained on data with the refusals filtered out [4], or edited instead: researchers found that, across 13 open-source chat models, erasing one internal "refusal direction" prevents the model from refusing harmful instructions. [5] |
| Added by the host | Rules and classifiers on top. OpenAI's Model Spec sets system-level rules that developers and users can't override [1]. When it launched Claude Opus 4 in May 2025, Anthropic added real-time classifiers that monitor inputs and outputs and block a narrow class of harmful chemical, biological, radiological and nuclear (CBRN) weapons information, initially focused exclusively on biological weapons [6]. Anthropic's current Usage Policy says its products "include real-time safeguards that may block or limit certain outputs" [7]. Google's Gemini API has adjustable safety filters, off by default for Gemini 2.5 and later models, plus built-in protections, such as child safety, that are always on and can't be adjusted. [3] | No filter or classifier of our own. The model provider can still refuse some prompts. | Nothing, unless you add it. You are the host. |
| Rules (policy, license, law) | A usage policy. OpenAI pairs its Model Spec with usage policies for ChatGPT and the API [1]. Anthropic's bans weapons development and sexually explicit content for all users [7]. Google's bans content that facilitates hate speech, violence or the incitement of violence, and sexually explicit content (for example, content made for pornography or sexual gratification). It says it may make exceptions for educational, documentary, scientific or artistic reasons, or where substantial benefits to the public outweigh the harm. [8] | Our Terms ban planning or carrying out violence against an identified person, sexual content involving minors or sexualizing real people without consent, malware aimed at a specific target without authorization, impersonating a real person for fraud or harassment, illegal use and attacks on the service. [9] | The model's license. Mistral Small 3 is released under Apache 2.0 [10], which grants broad rights and lists no banned uses [11]. Llama 3's license requires you to comply with applicable laws and follow Meta's Acceptable Use Policy [12], which bans uses such as violating the law or helping plan activities that present a risk of death or bodily harm. [13] |
| Who runs it | The company that makes it (OpenAI, Anthropic or Google), on its own service and under its own usage policy. [1] [7] [8] | A third-party inference provider serves the model. Unrestricted is the chat app, on the web and iPhone (iOS 18 or later). | You, on your own hardware. Mistral says Mistral Small 3, a 24B model, runs quantized on a single RTX 4090 or a MacBook with 32GB of RAM. [10] |
| Usage limits | Varies by service. Google's Gemini Apps, for example, use compute-based limits that refresh every 5 hours up to a weekly limit, are higher on paid plans, and may change without notice. [14] | A free allowance to try it, then Pro. On the web, the free allowance is one-time and doesn't reset, and Pro removes the message quota; per-minute rate limits still apply. iPhone limits are set separately. | No message cap. Your hardware sets the speed. |
Mainstream with looser settings.Some mainstream services let developers change the defaults. OpenAI's Model Spec gives developers broad latitude inside rules they can't override [1], and on the Gemini API the adjustable safety filters are off by default for Gemini 2.5 and later models (developers can turn them on), while built-in protections against core harms such as child safety are always on and can't be adjusted [3]. Changing these settings moves the host layer; the training and the rules stay.
Uncensored hosted chat.Most of the model's trained refusals have been removed [4], and the service decides what, if anything, it adds on top. You still depend on whoever serves the model and on the service's terms.
Run it yourself. You control the host layer, and nothing caps how many messages you send. You pay for that in hardware and setup, and the license still applies.
How uncensored models are made, by retraining or by editing the model, is the subject of what uncensored LLMs are. If you searched for "derestricted AI", the editing route is explained in abliterated models.
A jailbreak is something else again. It is a prompt designed to get a model past its safety guardrails [2], so the trained refusals, the host's classifiers and the terms are all still there, which is why jailbreaks keep breaking.
What an AI without restrictions still won't do
Taking refusals out of a model changes one layer. Three things stay.
It can still refuse.A hosted uncensored chat runs on someone's infrastructure, under someone's policy. The model may have had its trained refusals removed, but the company serving it can still turn some prompts away. The model itself may not be fully cleaned either. Eric Hartford, whose 2023 essay on uncensored models describes the retraining method, sums up the data step as "Identify and remove as many refusals and biased answers, and keep the rest," and describes the resulting WizardLM-Uncensored as "much more compliant." [4] Removing as many refusals as you can find is not the same as removing all of them.
The rules still apply.Removing refusals from a model doesn't remove the service's terms, the model's license or the law. Of the usage policies, terms and licenses cited here that list banned uses, all ban illegal use: Anthropic's [7], Google's [8], Meta's Llama 3 terms [12] [13] and our own Terms [9]. Apache 2.0 lists no banned uses [11], but the law still applies. As Hartford puts it: "You are responsible for whatever you do with the output of these models..." [4]
It can still be wrong.Removing refusals is meant to change what a model is willing to say, not what it knows. The researchers who found the refusal direction report that erasing it disables refusal "with minimal effect on other capabilities," though in their tests accuracy on TruthfulQA consistently dropped for the edited models [5]. Language models also sometimes guess when they are unsure and produce plausible but incorrect statements. [15] Check anything that matters against a primary source.
On Unrestricted specifically, we add no filter or classifier of our own, but the model provider can still refuse some prompts, and our Terms ban some uses. The Unrestricted FAQ covers what it still refuses, accounts and how the free allowance and Pro work.
Usage limits are a different thing
"AI with no limits" can also mean usage limits, the caps on how much you can use a service and how fast, which don't change what a model will answer. Our guide to AI limits, message caps and rate limits compares them across ChatGPT, Claude and Gemini.
Three questions to ask before you pick one
- 01
Who runs the model?
A mainstream assistant runs its own models. A hosted uncensored chat may send your prompt to another company that serves the model. A local model runs on your own machine.
- 02
What does it still refuse?
Read the usage policy or terms, and check whether the service adds a filter of its own. A plain list of what is banned is more useful than a promise that nothing is.
- 03
What does it store?
Find out whether chats are saved on the service's servers, whether they are used for training, and what any incognito mode actually skips.
Our own answers: a third-party inference provider serves the model, and what still gets refused is covered above. Saved chats are stored on our servers, and we don't train models on your chats [9]. If you want the background, here is why we built Unrestricted.
If cost is what decides it, the free uncensored AIguide compares the free routes, from a local model to a hosted chat's free allowance, and the catch with each.
Frequently asked
Is there an AI with no restrictions at all?
No. Every chatbot is shaped by its training, whatever the company running it adds on top, the rules that apply to it (terms, license and the law), and usage limits. Uncensored and self-hosted models can remove some of those layers, mainly trained refusals and added filters, but the rules and the law still apply.
Can an AI without restrictions still refuse?
Yes. An uncensored model has had most of its trained refusals removed, but a hosted service can still refuse some prompts, and its terms ban some uses. Read the service's terms for a plain list of what is banned.
What AI doesn't say no?
None says yes to everything, so judge by criteria rather than a ranking. Look for a model with its refusals trained or edited out, no extra filter added by the host, terms that name what is banned instead of promising that nothing is, and clear answers on who runs the model and what is stored. A model you run yourself removes the host layer entirely, but its license still applies.
Is there an AI with no limits?
Not in either sense. On content, a model comes with a license or terms, the law still applies, and hosted services can still refuse some prompts. On usage, hosted chats can cap how much you can use them or how fast, and even a plan that drops the message quota can keep rate limits.
Sources
Checked November 2026.
- Model Spec (October 29, 2026), OpenAI
- Constitutional Classifiers: Defending against universal jailbreaks, Anthropic
- Safety settings (Gemini API), Google
- Uncensored Models, Eric Hartford
- Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., arXiv 2406.11717), arXiv
- Activating AI Safety Level 3 protections (May 22, 2025), Anthropic
- Usage Policy (version effective November 12, 2026), Anthropic
- Generative AI Prohibited Use Policy, Google
- Terms of Service, sections 4 (Acceptable use) and 5 (Your content), Unrestricted
- Mistral Small 3, Mistral AI
- Apache License, Version 2.0, Apache Software Foundation
- Meta Llama 3 Community License Agreement, Meta
- Meta Llama 3 Acceptable Use Policy, Meta
- Gemini Apps limits and feature availability, Google
- Why Language Models Hallucinate (Kalai et al., arXiv 2509.04664), arXiv
Keep reading
AI limits: message caps and rate limits in ChatGPT, Claude and Gemini
How AI usage limits work: message caps, rate limits and context windows in ChatGPT, Claude and Gemini, what 'unlimited' means, and how to tell which you hit.
ChatGPT Temporary Chat: what it saves, what it skips, and how long OpenAI keeps it
ChatGPT Temporary Chat skips your history, memory and training, but OpenAI can keep a copy for a while. How to start one, and how Claude and Gemini compare.
Dolphin Mistral: what the Dolphin models are and how they're trained
What Dolphin Mistral is, who makes the Dolphin models, how they are fine-tuned to follow the system prompt instead of refusing, and their licenses.