back to blog
EXPLAINERApril 17, 2026·9 min read

How AI content filters work, and what 'no filter AI' really means

How AI content filters work: the training, classifier and policy layers that make chatbots refuse, and what 'no filter AI' can and can't mean.

By Unrestricted team · Updated November 2026 · Editorial standards

AI content filters work in layers. Safety training teaches the model itself to refuse some requests. The company running it can add instructions and classifiers that screen your prompt and the reply. A usage policy, backed by monitoring and account action, sets what you may use it for. Any of them can stop you.

That is why "no filter AI" can't mean no filter at all. At most it means a chat with no filter added on top of the model. Below: each layer, described from the companies' own documents; what four published policies restrict; what "no filter AI" and "unfiltered AI chat" can mean; and what changes when a service adds no filter of its own.

The layers that filter a chatbot

Four layers limit what a chatbot will do for you. A refusal in a reply can come from any of the first three; the fourth acts on accounts rather than single replies. Not every deployment uses all four, so each layer below is sourced to a company's own paper, policy or developer docs.

  1. 01

    Safety training

    The model learns in training which requests to refuse. This layer lives in the weights, so it goes wherever the model goes.

  2. 02

    Instructions

    A system prompt and other instructions from the company or the app developer travel with each conversation and shape what the model will do.

  3. 03

    Classifiers on the prompt and the reply

    Separate models screen your message before the main model sees it, the reply before you see it, or both.

  4. 04

    Usage policies and enforcement

    Rules for what people may use the service for, enforced by monitoring and action against accounts rather than inside a single reply.

Safety training.OpenAI's InstructGPT paper describes one recipe in three steps: fine-tune a model on answers written by people, train a reward model on people's rankings of the model's answers, then use reinforcement learning to push the model toward the answers people preferred.[1]Anthropic's Constitutional AI paper replaces human labels of harmful outputs with a list of written principles and AI feedback, aiming for a model that explains its objections instead of dodging the question.[2] OpenAI says it trains its models to align with its Model Spec, the document that sets how they should behave.[3]

The InstructGPT authors used the term "alignment tax" for the performance their alignment step cost on some public NLP benchmarks, and they greatly reduced those losses by mixing in training updates on their pretraining data.[1] By analogy, when safety training makes a model refuse harmless questions, the people asking pay a kind of alignment tax, the subject of our piece on AI refusals and the alignment tax.

Instructions.OpenAI's Model Spec sets a chain of command: OpenAI's instructions outrank a developer's, and a developer's outrank a user's. System-level rules can't be overridden by developers or users, and developers get broad latitude within OpenAI's bounds.[3]Anthropic publishes the system prompt that Claude's web and mobile apps use at the start of every conversation, which passes along details such as the current date and encourages certain behaviors, and says these system prompt updates don't apply to the Claude API.[4]So the same model can behave differently in a company's own app and in someone else's product built on its API.

Classifiers on the prompt and the reply.A classifier is a separate model that labels text as allowed or not. OpenAI's moderation endpoint is free to use and flags text for categories such as harassment, hate, illicit activity, self-harm, sexual content and violence; the self-harm, sexual and violence categories also work on images. It can check both the input and the generated output, and OpenAI's guide suggests using the results to filter content, send it for review or act on accounts.[5]A classifier can itself be a language model: Meta's original Llama Guard (2023) was a Llama 2 7B model tuned to classify both prompts and responses against a safety risk taxonomy.[6]

Anthropic's Constitutional Classifiers screen inputs and outputs and were built to stop jailbreaks. In Anthropic's test of a classifier-guarded version of Claude 3.5 Sonnet (October 2024) against 10,000 synthetic jailbreak prompts, the jailbreak success rate fell from 86% without classifiers to 4.4% with them. Anthropic says its first prototype refused too many harmless queries, and that the improved version raised the refusal rate by only 0.38%, which was not statistically significant.[7] It turned the classifiers on for Claude Opus 4 in May 2025, as part of ASL-3 protections it called precautionary and provisional, aimed at chemical, biological, radiological and nuclear weapons; it says those measures were initially focused exclusively on biological weapons.[8]Google's Gemini API has adjustable filters for harassment, hate speech, sexually explicit content, dangerous content and jailbreaks. If one blocks the prompt or the reply, the request fails with an HTTP 400 error and the blocked content isn't returned.[9]

This is the layer jailbreaks run into: a prompt that talks a model past its training can still be caught by a classifier reading the reply. Why jailbreaks keep breaking is covered in our explainer on ChatGPT DAN and other jailbreaks.

Usage policies and enforcement.Some limits never show up in a reply. The Model Spec says OpenAI uses automated monitoring to detect potential usage policy violations, and that violating its usage policies may lead to action against the user's account.[3]Anthropic says its Safeguards team uses detection and monitoring to enforce its Usage Policy and may warn, throttle, limit, suspend or terminate access. It adds that its products have real-time safeguards that may block or limit some outputs, and that hitting a block doesn't by itself mean you broke the policy.[10]

Not every deployment stacks all four. Google says that because of the model's built-in safety, the Gemini API's extra filters are off by default for Gemini 2.5 and later models, while protections against core harms, such as content that endangers child safety, are always on and can't be adjusted.[9] An open-weights model you run yourself has no classifier or monitoring unless you add them. Any use policy that comes with its license still applies: Meta says anyone who accesses or uses Llama 3 agrees to its Acceptable Use Policy.[11] Other open-weights models come with a license alone: the model card for Mistral Small 24B Instruct lists the Apache 2.0 license and no separate use policy.[12]

What each published policy restricts

OpenAI, Anthropic, Google and Meta each publish written rules. Anthropic's, Google's and Meta's say what people may use their models for. OpenAI's row uses its Model Spec, which sets how its models should behave and is complemented by OpenAI's separate usage policies.[3] Here is what each restricts, from the version we checked. For ChatGPT specifically, what the Model Spec tells the model to decline and what to try when it refuses are in our guide to ChatGPT's content policy, and how OpenAI's rules have changed over time is covered in our explainer on whether ChatGPT is censored.

Checked November 10, 2026, on each company's own page. OpenAI's row describes its Model Spec, which sets how its models behave, not its separate Usage Policies.
Policy and versionWhat it restrictsHow it is enforced
OpenAI Model Spec (October 29, 2026 version)Red-line principles against facilitating critical and high-severity harms, such as acts of violence (including crimes against humanity, war crimes, genocide, torture and human trafficking), creating cyber, biological or nuclear weapons, terrorism, child abuse, persecution and mass surveillance. Sexual content involving minors is never generated; informational hazards and sensitive personal data are restricted; erotica and gore are allowed only in appropriate contexts.[3]Trained into the models, plus automated monitoring for usage policy violations and action against accounts.[3]
Anthropic Usage Policy (version taking effect November 12, 2026)For all users, including: illegal activity; compromising critical infrastructure or computer systems without authorization; developing or designing weapons, including chemical, biological, radiological and nuclear weapons; inciting violence or hate; compromising privacy or children's safety; cruel or psychologically harmful conduct, such as promoting self-harm; disinformation and deceptive campaigns; undermining democratic processes; surveillance and censorship; fraud; abusing the platform, including jailbreaking without authorization; and sexually explicit content.[10]Detection and monitoring; warnings, throttling, suspension or termination of access; real-time safeguards that can block some outputs.[10]
Google Generative AI Prohibited Use Policy (last modified December 17, 2024)Dangerous or illegal activity (including child sexual abuse, violent extremism, non-consensual intimate imagery and self-harm), compromising security (spam, phishing, malware and getting around safety filters), hate speech, harassment, incitement to violence, sexually explicit content made for gratification, and scams, impersonation and other deception.[13]The page doesn't say. Google says it may make exceptions for educational, documentary, scientific or artistic reasons, or where harms are outweighed by substantial benefits to the public.[13]
Meta Llama 3 Acceptable Use Policy (no date on the page)Breaking the law or others' rights (including violence, terrorism, child exploitation, human trafficking and harassment), malware, activities that risk death or bodily harm (weapons, illegal drugs, critical infrastructure, self-harm), deception such as fraud, disinformation and impersonation, and failing to tell end users about known dangers of an AI system.[11]Anyone who accesses or uses Llama 3 agrees to it, including not letting others use it for these purposes. Meta lists an email address for reporting violations.[11]

All four rule out the sexual exploitation of children. These are written rules, not filter settings, so a model with fewer trained refusals doesn't change what they allow.

What "no filter AI" means

"No filter AI" means an AI chat with no content filter added between you and the model: no classifier screening your prompt or the reply. (The phrase also turns up for camera and photo apps without beauty filters; this guide is about chat.)

It can't mean a chatbot with no limits of any kind. The model's training still shapes what it says, whoever serves the model can still refuse some prompts, and usage policies, licenses and the law still apply. So a no-filter AI is best described by what was taken out: the added classifiers and, in an uncensored model, the refusals its makers trained or edited out.

Unfiltered AI chat

Unfiltered AI chat is the same idea seen from the chat window: a service that sends your message to the model and returns its reply with no moderation layer in between. What you get then depends on the model. A mainstream model keeps the refusals from its safety training even with the filters off, which is why Google can leave the Gemini API's extra filters off by default.[9] An uncensored model has had those refusals trained or edited out.

You can sometimes tell which layer stopped you. In Google's Gemini API, a classifier blocks rather than argues: a blocked prompt or reply makes the request fail with an HTTP 400 error whose code is "safety" (or, for a blocked prompt, "jailbreak", "prohibited_content" or "blocklist"), and the blocked content isn't returned.[9]A refusal the model writes out, with reasons or a partial answer, can come from its training or its instructions, but it doesn't prove which layer acted: OpenAI's moderation guide leaves it to each app to decide what to do with a flag, so an app can answer a classifier flag with a written refusal too.[5]

How uncensored models remove refusals

Uncensored models go after the training layer. Some are fine-tuned on instruction data filtered "to remove alignment and bias," as the Dolphin 2.6 Mistral 7B model card puts it,[14] and others have the single direction that mediates refusal removed from their weights;[15] both methods are covered in our guide to uncensored LLMs.

What changes without the filter stack

Here is how that looks at Unrestricted. We send chats to one uncensored open-weights model based on Dolphin-Mistral-24B, served by a third-party inference provider, and we add no filter or classifier of our own: nothing on our side screens your prompt before it reaches the model or grades the reply before you see it.

Some things stay. The model provider still refuses some prompts, which is a block on the provider's side rather than a filter of ours, and our Terms ban some uses: violence planning, sexual content involving minors or non-consensual depiction of real people, targeted malware, fraud impersonation and illegal use. Saved chats are stored on our servers, and we don't train models on your chats.

What changes is which layer can stop a request: with no classifier of our own, a refusal comes from the model provider, not from a filter we added. How filtered chatbots can hedge on legitimate questions, such as dark history topics, is covered in our piece on AI for history questions. A model with fewer refusals can also be wrong, so check what matters. Web search with citations is included at no extra cost, which helps with that.

Limits that remain

Taking out the filters doesn't take out the rules: usage policies, model licenses and the law still apply to what you do with the answers, and some requests stay off-limits everywhere. What a no-filter or uncensored AI still won't do, across the category, is in our guide to AI without restrictions.

Frequently asked

  • What is a no-filter AI?

    An AI chatbot with no content filter added on top of the model, meaning no classifier screening your prompts or its replies. It isn't an AI with no limits: the model's training, the provider's rules and the law still apply. The phrase is also used for camera apps without beauty filters, which is a different thing.

  • What is unfiltered AI chat?

    A chat service that passes your message to the model and returns its reply with no moderation layer in between. How much it answers then depends on the model: a mainstream model keeps the refusals from its safety training, while an uncensored model has had them trained or edited out. Usage policies and the law still apply either way.

  • Why are AI chatbots censored?

    Companies train and filter chatbots so they won't help with harms their policies rule out, such as weapons capable of mass harm or the sexual exploitation of children, and so use stays within their usage policies. The same layers can also catch harmless questions, which is called over-refusal. Refusing help with a weapon is a safety rule; hedging on a legitimate history question is over-refusal.

  • What is AI filtering?

    In AI chat, filtering means screening prompts and replies for content a provider doesn't allow, either by training the model to refuse or with separate classifier models such as OpenAI's moderation endpoint and Meta's Llama Guard. The term is also used for AI tools that filter spam, email or web content for businesses, which is a different kind of product.

Sources

Checked November 2026.

  1. Training language models to follow instructions with human feedback (Ouyang et al., 2022), arXiv
  2. Constitutional AI: Harmlessness from AI Feedback (Bai et al., 2022), arXiv
  3. Model Spec (October 29, 2026), OpenAI
  4. System prompts (release notes), Anthropic, Claude docs
  5. Moderation guide, OpenAI developer docs
  6. Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations (Inan et al., 2023), arXiv
  7. Constitutional Classifiers: Defending against universal jailbreaks (February 3, 2025), Anthropic
  8. Activating AI Safety Level 3 protections (May 22, 2025), Anthropic
  9. Safety settings (Gemini API), Google AI Studio docs
  10. Usage Policy (version taking effect November 12, 2026), Anthropic
  11. Meta Llama 3 Acceptable Use Policy, Meta
  12. Mistral-Small-24B-Instruct-2501 model card, Mistral AI, Hugging Face
  13. Generative AI Prohibited Use Policy (last modified December 17, 2024), Google
  14. dolphin-2.6-mistral-7b model card, Hugging Face
  15. Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., 2024), arXiv