back to blog
ANALYSISNovember 14, 2026·8 min read

Is ChatGPT censored? What OpenAI restricts, what changed, and where it's blocked

Is ChatGPT censored? How OpenAI's rules have changed, what published studies say about political bias, and where ChatGPT isn't available.

By Unrestricted team · Editorial standards

ChatGPT is censored in one sense: OpenAI decides what ChatGPT won't produce and publishes those rules in its Model Spec, alongside separate Usage Policies for the people who use it.[1]In another sense it isn't: these are one company's product rules, not a government topic ban, and the spec's default guideline says its models should never avoid a topic solely because it is sensitive or controversial.[1]

Two more things sit under the same question. One regulator, Italy's data protection authority, temporarily stopped ChatGPT in 2023 over privacy rather than content,[2]and OpenAI doesn't offer API access in every country.[3]For how a chatbot's refusals are produced in the first place, see how AI content filters work. Below: what OpenAI restricts, a dated timeline of how the rules changed, what studies say about political bias, and where ChatGPT is blocked.

What OpenAI restricts by design

OpenAI's main published rules for ChatGPT are in two documents. The Model Spec describes how its models should behave, and the Usage Policies, in OpenAI's words, "outline our expectations for how people should use the API and ChatGPT."[1] The Model Spec sorts what its models should not generate into three tiers:[1]

  1. 01

    Prohibited: never, in any context

    Only sexual content involving minors, including transformations, such as translations, of content a user supplies. OpenAI says it keeps this tier to that one item "to maximize freedom for our users."

  2. 02

    Restricted: context doesn't unlock it

    Information hazards, such as detailed, actionable steps toward chemical, biological, radiological or nuclear weapons; content designed to manipulate the political views of specific individuals or demographic groups; and rules on creators' rights and people's privacy, including sensitive personal data. An informational or historical purpose doesn't unlock this tier; the only exception is transforming restricted content the user supplied.

  3. 03

    Sensitive: allowed in some contexts

    Erotica, depictions of illegal or non-consensual sexual activity and extreme gore, allowed only in scientific, historical, news, artistic or similar contexts. Content tied to violent extremism and hateful content aimed at protected groups sit here too, allowed only when discussed critically, analytically or factually, in contexts such as news reporting, history or education.

The sensitive tier also rules out gratuitous abuse or harassment of individuals, unless a user explicitly asks for it in a context that breaks no other rule. Outside the three tiers, the spec asks for extra care around self-harm, delusions and mania, illegal activity, and legal, medical or financial advice, and it adds protections for teens aged 13 to 17.[1]The spec is frank about one reason for all this: alongside being useful and preventing serious harm, a stated goal is to "maintain OpenAI's license to operate by protecting it from legal and reputational harm." [1]

So ChatGPT is filtered in the plain sense that a company decides what its product will say. But the same document argues against dodging subjects: "Refusing to discuss a polarizing topic or event is itself a form of agenda."[1]OpenAI also says the public version of the spec "may not include every detail" and that its "production models do not yet fully reflect the Model Spec," so a written rule and the answer you get can differ.[1]

If ChatGPT just turned you down, the cause can be one of these rules or a harmless request read as a risky one. Our guide to ChatGPT's content policy maps the refusal messages to what each can mean, and what to try without a jailbreak.

How the rules changed since 2024: a dated timeline

OpenAI first published the Model Spec in May 2024 and keeps a public changelog of every version since.[4]The timeline below follows the Model Spec, plus the GPT-5 system card, where OpenAI described a new safety training approach. It doesn't track changes to the separate Usage Policies.

Changes to OpenAI's Model Spec and safety approach, May 2024 to October 2026
DateDocumentWhat changed
May 8, 2024Model Spec, first releaseRules included "Don't respond with NSFW content," which the spec said may include erotica, extreme gore, slurs and unsolicited profanity. OpenAI said it was exploring whether it could allow such content in age-appropriate contexts. The spec already said the assistant "must not encourage or enable self-harm." [5]
February 12, 2025Model Spec, second release (the first open-source one, in the public domain)Sorted disallowed content into tiers, with only sexual content involving minors labeled prohibited; moved erotica and gore from a flat NSFW rule to sensitive content allowed in some contexts; and added "Don't have an agenda," "No topic is off limits" and a commitment to intellectual freedom.[4][5][6]
April 11, 2025Model SpecSmall fixes, such as narrowing the exception for white lies to pleasantries.[4]
August 7, 2025GPT-5 system cardIntroduced "safe-completions": instead of training models mainly to comply in full or refuse outright, GPT-5 models aim to give the most helpful answer the safety rules allow.[7]
September 12, 2025Model SpecMoved the guidance from hard refusals to safe completions in most cases, added red-line principles, and added rules for agents that take actions.[4]
October 27, 2025Model SpecExtended the self-harm rules to cover delusions and mania, and added a section on respecting real-world ties.[4]
December 18, 2025Model SpecAdded an Under-18 Safety Mode with extra guidance for teens aged 13 to 17.[4]
August 18, 2026Model SpecClarified guidance on relational interactions for teens and on requests built on false premises, and added a section on being clear about what the model can and can't do.[4]
October 29, 2026Model Spec, current versionReorganized the spec around five values and renamed its subject from "the assistant" to "GPT." Its commentary still says OpenAI is "exploring" how to allow erotica and gore in age-appropriate contexts.[4][1]

Read together, the documents moved in two directions. Rules about topics and viewpoints loosened in 2025: erotica and gore moved from a flat NSFW rule to a tier allowed in some contexts, only sexual content involving minors is labeled prohibited, "no topic is off limits" became a stated principle, and the guidance moved from hard refusals toward safe, partial answers in most cases.[4][5][6] Over the same period OpenAI added protections for teens aged 13 to 17 and extended its existing self-harm rules to cover delusions and mania.[4][5] None of this shows how a given model answers on a given day; for that you need dated tests with saved transcripts.

Political bias: OpenAI's stated goal and published studies

OpenAI's stated goal is neutrality by default. The Model Spec tells its models to "fairly describe significant views" and present the strongest arguments for each position, and says that in ChatGPT, personalization must not "introduce biases or omissions that could undermine users' ability to form informed opinions."[1]Peer-reviewed studies have measured something different: a lean in older models' answers to political questionnaires.

Peer-reviewed studies of political lean in ChatGPT and other chatbots
StudyModels and methodWhat it reported
Motoki, Pinho Neto and Rodrigues, Public Choice (online August 2023)The authors report asking ChatGPT the Political Compass questions 100 times each, in random order, by default and while impersonating partisansA "significant and systematic political bias" toward the Democrats in the US, Lula in Brazil and the Labour Party in the UK[8]
Rozado, PLOS ONE (July 2024)24 conversational models (a set that included OpenAI's GPT-4) on 11 political orientation tests, run December 2023 to January 2024Most conversational models gave answers that most tests rated left of center; results for base models without chat training were inconclusive; fine-tuning could steer a model's position[9]

Both studies have limits. They used multiple-choice questionnaires. Rozado disputes the criticism that such test results are close to random, but calls the concern about forcing models to pick from preset answers "more valid" and suggests that future work look at open-ended answers. In his tests, GPT-4 had an invalid-response rate of 33%, mostly because it often declined to answer a test question, saying it lacked political preferences. His paper also says the results shouldn't be read as evidence that developers deliberately inject political preferences.[9]Critics Arvind Narayanan and Sayash Kapoor argued that Motoki et al. actually queried text-davinci-003, an older model not used in ChatGPT, and that the study's forced-choice prompt, with dozens of questions at once, was far from real use.[10]Both studies also tested models from 2023 and early 2024, so neither says how today's models score. A lean is also a different complaint from censorship: one is about which views an answer favors, the other about what it won't say.

Over-refusal is not the same as censorship

Some of what gets called censorship is over-refusal: a model declining a harmless request because it resembles a risky one, which is a calibration problem rather than a rule against the topic. What that costs the people asking, and what published research measures, is covered in AI refusals and the alignment tax.

Where ChatGPT isn't available, and what governments have done

Not every block is about content. OpenAI publishes a list of countries and territories that "currently support access to our API services," and warns that accessing or offering access to its services outside them "may result in your account being blocked or suspended."[3]China, Hong Kong, Russia, Iran, North Korea, Cuba, Syria, Belarus and Venezuela aren't on that list. The page is written for API access, so check OpenAI's help pages before relying on it for ChatGPT in a specific country.

Where local law requires changes, the Model Spec says the model must be open about it: content left out for legal reasons beyond the spec's own limits "must be transparently indicated to the user," along with the type of information removed and why.[1]Its example for "No topic is off limits" is a question about Tiananmen Square, where an answer that talks around the 1989 protests is marked as one that "censors / evades the point of the question."[1] For getting sourced answers on dark or contested history from any chatbot, see AI for history questions.

In Italy, a government regulator did stop ChatGPT, and its reasons were about personal data, not topics:

Italy's data protection authority (the Garante) and ChatGPT
DateWhat happened
March 30 to 31, 2023The Garante imposed an immediate temporary limitation on OpenAI's processing of Italian users' data. It cited no legal basis for mass collection of personal data to train the models, no information for users, inaccurate personal data and no age verification.[2]
April 28, 2023OpenAI made ChatGPT available in Italy again after publishing a notice on how it uses personal data for training, adding a form to opt out of that processing, and asking for a date of birth at sign-up.[11]
December 20, 2024The Garante announced decision No. 755 of November 2, 2024, closing its investigation with a fine of 15 million euros and an order to run a six-month information campaign. It passed any continuing issues to Ireland's data protection authority (the DPC), which had become OpenAI's lead authority under the GDPR.[12]
March 18, 2026The Court of Rome upheld OpenAI's appeal and annulled decision No. 755 (judgment No. 4153/2026). It found that, under the GDPR's one-stop-shop rules, the Garante wasn't the competent authority, and it didn't rule on the appeal's other grounds. The Garante temporarily removed the decision from its website.[12][13]

This is background information, not legal advice.

So, is ChatGPT censored?

ChatGPT runs on published company rules: a short list of hard limits, a longer list of contexts that call for care, and a default guideline against dodging controversial topics. Whether that counts as censorship depends on your definition. It isn't a government topic ban, though OpenAI doesn't offer API access everywhere. Written rules are only one of the layers that shape a chatbot's answers, and our guide to AI content filters explains each one.

Frequently asked

  • Is ChatGPT censored?

    OpenAI sets limits on what ChatGPT will produce and publishes them in its Model Spec, so in that sense the company filters its answers. It isn't a government topic ban, and the spec's default guideline says its models should never avoid a topic solely because it is sensitive or controversial. The hard limits are sexual content involving minors, which is never allowed, and restricted content such as detailed, actionable steps toward weapons and sensitive personal data. Hateful content aimed at protected groups is off by default but allowed in critical or educational contexts.

  • Why is ChatGPT so restrictive?

    OpenAI's Model Spec lists preventing serious harm and protecting OpenAI's "license to operate" from legal and reputational harm among its goals, alongside being useful. Some refusals follow written rules, such as no detailed steps toward weapons. Others can be over-refusals, where a harmless request is read as a risky one, and OpenAI says its production models don't yet fully reflect the spec.

  • Is ChatGPT politically biased?

    Peer-reviewed studies of older versions found a lean. A study in Public Choice, published online in 2023, reported a systematic bias toward the Democrats in the US, Lula in Brazil and Labour in the UK, though critics argued it queried an older model not used in ChatGPT. A 2024 PLOS ONE study found that most of the 24 chatbots it tested (a set that included GPT-4) gave answers most political tests rated left of center. Both used multiple-choice questionnaires on models from 2023 and early 2024, and OpenAI's Model Spec tells its models to present significant views fairly without taking an editorial stance.

  • Is ChatGPT banned in some countries?

    OpenAI's list of countries supported for API access leaves out China, Russia and Iran, among others, and warns that accessing its services from elsewhere can get an account blocked or suspended; that list is written for the API, so check OpenAI's help pages for ChatGPT. In Italy, the data protection authority temporarily limited OpenAI's processing of Italian users' data in March 2023 over privacy concerns, and ChatGPT came back there at the end of April 2023 after OpenAI made changes.

  • Has ChatGPT gotten more censored?

    On paper, OpenAI's rules about topics loosened in 2025: the February 2025 Model Spec labeled only sexual content involving minors as prohibited, moved erotica and gore to a tier allowed in some contexts and said no topic is off limits, and later versions favor safe, partial answers over flat refusals. Over the same period OpenAI added protections for teens aged 13 to 17 and extended its existing self-harm rules to cover delusions and mania. The written rules don't show how a given model answers on a given day.

Sources

Checked November 2026.

  1. Model Spec (October 29, 2026 version), OpenAI
  2. Artificial intelligence: stop to ChatGPT by the Italian SA (March 31, 2023), Garante per la protezione dei dati personali
  3. Supported countries and territories, OpenAI API documentation
  4. Model Spec changelog, OpenAI, on GitHub
  5. Model Spec (May 8, 2024 version), OpenAI
  6. Model Spec (February 12, 2025 version), OpenAI
  7. GPT-5 System Card (August 7, 2025), OpenAI Deployment Safety Hub
  8. More human than human: measuring ChatGPT political bias, Fabio Motoki, Valdemar Pinho Neto and Victor Rodrigues, Public Choice, vol. 198 (2024), published online August 17, 2023
  9. The political preferences of LLMs, David Rozado, PLOS ONE, July 31, 2024
  10. Does ChatGPT have a liberal bias?, Arvind Narayanan and Sayash Kapoor, AI as Normal Technology, August 18, 2023
  11. ChatGPT: OpenAI reinstates service in Italy with enhanced transparency and rights for European users and non-users (April 28, 2023), Garante per la protezione dei dati personali
  12. ChatGPT: the Italian Data Protection Authority closes the preliminary investigation (December 20, 2024, with a note on the March 2026 judgment), Garante per la protezione dei dati personali
  13. Judgment in OpenAI OpCo v. Garante per la protezione dei dati personali, R.G. 4785/2025 (decided March 18, 2026), Tribunale Ordinario di Roma, copy hosted by Sapienza University of Rome