back to blog
EXPLAINERNovember 13, 2026·7 min read

ChatGPT content policy: why ChatGPT refuses, and what to try

What ChatGPT's content policy restricts, why ChatGPT refuses or says it can't help, what each refusal message can mean, and fixes that need no jailbreak.

By Unrestricted team · Editorial standards

ChatGPT refuses, or seems to, for one of three main reasons. Your request matches something OpenAI's rules tell the model to decline; the model misreads a legitimate request as risky, which is over-refusal; or something other than the content is in the way, such as a usage limit, an unsupported country, an account problem or an outage.

ChatGPT's "content policy" or "guidelines" isn't one page. The two main OpenAI documents are the Usage Policies, which set out how people should use ChatGPT and the API, and the Model Spec, which sets how the model itself should behave, including what it declines. The Model Spec says it is complemented by the Usage Policies and by safety protocols such as testing and monitoring, and OpenAI's Terms of Use also apply.[1] For how filtering works across chatbots in general, start with how AI content filters work. (If you were looking for your school's or employer's ChatGPT guidelines, those are separate rules set by them.) This guide covers what OpenAI's rules restrict, what each kind of refusal can mean, and fixes that need no jailbreak.

Where ChatGPT's rules live

The Model Spec says the Usage Policies prohibit harmful uses such as spamming and scamming. Those are hard to address through the model's behavior because they depend on how content is used after it is generated, so OpenAI addresses them primarily at the system level. The Model Spec adds that the Usage Policies prohibit using the models to manipulate or deceive people.[1]The Model Spec also notes that OpenAI's Terms of Use prohibit using its services in ways that violate other people's rights.[1] You can read the current Usage Policies and Terms of Use on OpenAI's site.

The Model Spec is the rulebook for the model. The current version is dated October 29, 2026. It tells the model to assume good intent, to encourage intellectual freedom and to refuse only on grounds established by its chain of command: the ranked instructions from the Model Spec and OpenAI, then developers, then users. OpenAI also says its production models don't yet fully reflect the Model Spec.[1]

The rules change: the Model Spec's changelog runs from its first release in May 2024 to the October 2026 version, and the September 2025 update moved its guidance from hard refusals to "safe completions" in most cases: answers that give what they can within the rules.[2] The full history is in our timeline of how ChatGPT's rules changed.

Why ChatGPT refuses: what the Model Spec tells it to decline

The Model Spec sorts what it holds back into groups, and each group has different exceptions. These are the main ones, and only one is off-limits in every case.[1]

The second and third columns summarize OpenAI's Model Spec, October 29, 2026 version.
RuleWhat it coversWhat it still allows
Sexual content involving minorsNever produced, in any context, including rewrites of text you provide.[1]Non-graphic, factual discussion in educational contexts.
Information hazardsDetailed, actionable steps for illicit acts, harm to people or property, or large-scale harm, including any steps toward creating, obtaining, magnifying or deploying chemical, biological, radiological or nuclear weapons. The ban on these steps has no exception for informational or historical questions.[1]Neutral, factual information without step-by-step detail, when a topic also has legitimate uses.
Targeted political manipulationPolitical content designed to manipulate the views of specific people or demographic groups.[1]Political content for a broad or unspecified audience, on any political topic.
Creators' rightsRespecting creators' intellectual property: for example, no full lyrics of a song not in the public domain and no getting around paywalls.[1]Analysis, such as a song's key themes.
Private informationPrivate or sensitive information about people, even if it is online somewhere.[1]Generally public facts about public figures, such as an official's office phone number.
Erotica and extreme goreErotica, depictions of illegal or non-consensual sexual activity, and extreme gore, in text, audio or images.[1]Scientific, historical, news, artistic or other contexts where it is appropriate.
Illicit behaviorGuidance, instructions or actionable steps for illicit behavior, and, once you indicate illicit intent, even information it would give in another context.[1]Neutral facts about the consequences. When intent is unclear and nothing else rules it out, it should typically assume good intent and answer.
Violent extremism and hateContent that praises, endorses or aids violent extremist agendas, and hateful content targeting protected groups.[1]Critical, factual or educational discussion, such as news reporting or historical analysis.
Self-harm, delusions and maniaEncouraging or enabling self-harm, or affirming or escalating ungrounded beliefs, such as delusions.[1][2]A supportive reply that points to people and resources who can help.
Legal, medical and financial adviceDefinitive advice that a licensed professional should give.[1]Information, with a short note suggesting qualified help.

For users ages 13 to 17, the Model Spec's under-18 rules apply on top, and some requests an adult could make get declined, including risky activities that may be legal for adults.[1]

Refusal messages and what each can mean

A refusal doesn't always look like one. OpenAI says its models starting with GPT-5 prefer safe completions over hard refusals in most cases, while older models typically give a short, neutral refusal.[1]The Model Spec's template for a plain refusal is "Sorry, I can't help with [X]", and it gives "Sorry, I can't help with that" as the typical refusal from older models, so that is the only message quoted below; the other rows describe what you see.[1]

What you seeWhat it can meanWhat to try
"Sorry, I can't help with that."A plain refusal. The Model Spec uses this style when a request states illicit intent, and OpenAI says older models typically refuse this way.[1]If your purpose is legitimate, say what it is and ask for general information. That helps with ambiguous requests, not restricted ones.
A partial answer that says it can't cover everythingA safe completion: the model briefly explains why it can't give a full answer, then gives as much as its rules allow.[1]Ask for the allowed parts on their own, or for more depth on what it did answer.
A refusal to give full song lyrics or a paywalled articleThe creators' rights rule. The Model Spec's examples are the full lyrics of a song not in the public domain and getting around a paywall.[1]Ask for an analysis, such as the song's key themes, or for the specific information you need.
An answer that leaves out a person's private detailsThe privacy rule: no private or sensitive information about people, even if it is online.[1]Ask for public information, such as an official's office contact details.
An error, a cap, or a blocked accountNot a content refusal: a plan limit, a location block, account action or a service problem. Each is covered in the next sections.Check your plan's limits and OpenAI's status page before rewording anything.[3][4]

Over-cautious refusals

Sometimes the request is fine and the model misreads it: the XSTest test suite was built to catch safe prompts that get refused because they "use similar language to unsafe prompts or mention sensitive topics".[5] The OR-Bench authors add that stronger safety often comes with over-refusal as a side effect,[6] and what that costs is covered in our piece on AI refusals and the alignment tax.

Refusals that aren't about content: limits, region, account

OpenAI's pricing page lists Free-plan text chats as unlimited, subject to abuse guardrails, but limits messages with uploads, image creation, voice chats, deep research, and memory and context, and says limits also apply on Plus.[3] Hitting one of those caps can look like a refusal; how to tell which limit you hit is in our guide to AI message caps and rate limits.

Location can be the block. OpenAI limits where its services can be used: its API, for example, has its own supported-countries list, and its developer docs warn that accessing or offering access to its services outside the listed countries and territories may result in an account being blocked or suspended.[7] API errors are labeled, too: the error-code guide lists a 403 for an unsupported country, region or territory, a 429 for sending requests too quickly or hitting a credit, spend or usage limit, and a 503 when a model is temporarily overloaded.[8]

Account action is separate from a refusal. The Model Spec treats misuse of what the model provides as a Usage Policies matter, "which may result in actions against the user's account", and says OpenAI uses automated monitoring to detect potential usage policy violations.[1]A warning or a deactivated account is something to take up with OpenAI; rewording a prompt won't fix it.

And sometimes nothing is refusing at all. When replies stall or error out, check OpenAI's status page, which tracks ChatGPT's components one by one, such as Conversations, Login and Voice mode.[4]

What to try, without jailbreaks

The Model Spec tells ChatGPT never to ask about your intent in order to decide whether to refuse. When intent is unclear and nothing else rules a request out, it should typically assume good intent and comply, so stating your purpose can help when a request looks risky. These fixes work within OpenAI's rules, and the first four are based on what the Model Spec says.[1]

  1. 01

    Say what it's for

    The Model Spec says the same information can be refused in one framing and given in another: it declines tips for getting away with shoplifting but gives shoplifting-deterrence tips. Put your real purpose (a novel, a class, a safety check at work) in your first message. A stated purpose won't unlock restricted content, though: the Model Spec says there's no "good cause" exception for things like information hazards or private data.

  2. 02

    Ask how it works, not for a procedure

    When a topic has both harmful and legitimate uses, the Model Spec tells the model to give neutral, factual information without step-by-step detail. Questions about how something works, why it's dangerous or what the rules are fit that better than a request for a method.

  3. 03

    Paste the text you need help with

    The Model Spec lets ChatGPT handle limited requests to translate, summarize, classify or fix the grammar of restricted or sensitive text you provide, even text it wouldn't write itself, though never sexual content involving minors. It covers only text you paste in, not links or titles. It won't add new material its rules disallow, and other rules, such as the one on illicit behavior, still apply.

  4. 04

    Split a mixed request

    When part of a request is off-limits, the model should typically give a partial answer and briefly explain why. Ask for the allowed parts separately so they aren't cut short.

  5. 05

    Rule out the non-content causes

    Before rewording anything, check the status page, your plan's limits and whether a location or account problem is in the way.

  6. 06

    Don't invent a purpose

    Context helps only when it's true. The Model Spec treats misuse of what the model gives you as a matter for the Usage Policies, which may result in action against your account.

What doesn't belong on this list is a jailbreak, a prompt written to talk the model out of its rules. Why those keep breaking is covered in our explainer on ChatGPT DAN and other jailbreaks.

When a less restrictive assistant fits

If your question is legal, you've given the context and ChatGPT still declines or waters down the answer, a less restrictive assistant can be the better tool. If the reason is one of the hard lines in the table above, such as sexual content involving minors or help toward chemical, biological, radiological or nuclear weapons, switching tools isn't the answer, and shouldn't be.

That gap is what Unrestricted is for. It answers through 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. It still has limits. The model provider refuses some prompts, and our Terms ban some uses, including 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. A less restrictive model can also be wrong, so check what matters.

We compare the two in our comparison of Unrestricted and ChatGPT. For how refusals are produced across chatbots, from training to classifiers, see our guide to AI content filters.

Frequently asked

  • Why does ChatGPT say "Sorry, I can't help with that"?

    It's the short, neutral refusal OpenAI's Model Spec describes, used for example when a request states an illicit purpose. OpenAI says older models typically refuse this way, while models from GPT-5 on prefer a partial, safe answer in most cases. If your question is legitimate, say what it's for and ask for general information.

  • What are ChatGPT's guidelines?

    The two main OpenAI documents are the Usage Policies, which cover how people should use ChatGPT and the API, and the Model Spec, which covers how the model behaves and what it declines. The Model Spec says it is complemented by the Usage Policies and by safety protocols such as testing and monitoring, and OpenAI's Terms of Use also apply. The current Model Spec is dated October 29, 2026. Schools and employers can set their own ChatGPT guidelines, which are separate rules.

  • What triggers ChatGPT's content policy?

    Requests the Model Spec rules out: sexual content involving minors in any context, detailed steps toward illicit or dangerous acts including chemical, biological, radiological or nuclear weapons, political content designed to manipulate specific people or groups, full lyrics of songs not in the public domain or getting around paywalls, private information about people, and erotica or extreme gore outside contexts such as news, history or art. Stating an illicit purpose triggers a refusal even for information it would otherwise share. Harmless prompts can trigger it too, through over-refusal.

  • Why won't ChatGPT answer my question?

    If you get an error or no reply rather than a refusal, the content may not be the problem. Check OpenAI's status page for an outage and your plan's limits; a location or account problem can also block you. If it is a refusal, add your purpose and ask for general information rather than a step-by-step method.

Sources

Checked November 2026.

  1. Model Spec (October 29, 2026), OpenAI
  2. Model Spec changelog, OpenAI, on GitHub
  3. ChatGPT pricing (plan features and limits), OpenAI
  4. OpenAI Status, OpenAI
  5. XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models (Röttger et al., NAACL 2024), arXiv
  6. OR-Bench: An Over-Refusal Benchmark for Large Language Models (Cui et al., ICML 2025), arXiv
  7. Supported countries and territories (API), OpenAI developer docs
  8. Error codes, OpenAI developer docs