AI Laws That Restrict, Filter, or Police Inquiry

Full report

On this page

Executive summary

Across the jurisdictions reviewed, I did not find a major enacted democratic-law analogue that literally criminalizes “unlawful questions” or “unlawful thought” as such. What I did find is a spectrum of laws and bills that can functionally push providers to screen prompts, outputs, user activity, identity traits, or recommendation flows for illegality, falsity, harmfulness, or age-inappropriateness. The clearest and most direct example is China, where the 2023 Generative AI Measures expressly require providers to ensure outputs do not include broad categories of prohibited content, to stop generation/transmission upon finding unlawful content, to restrict or terminate users engaged in unlawful activity, to preserve records, and to report to authorities. The UK and India have adopted or attempted regimes that pressure intermediaries to assess, trace, or suppress unlawful or “fake, false or misleading” content; the EU generally moves in the opposite direction on “thought-like” inference by banning or tightly restricting some AI practices such as emotion recognition in workplaces and schools and criminal-risk prediction, while still imposing stronger illegal-content risk mitigation duties on large platforms. The United States remains fragmented: enacted federal law is narrow and targeted, while the most aggressive current U.S. proposals are aimed at minors, deepfakes, and recommendation systems rather than ideas as such.

The jurisdictions also differ sharply in what is being judged. In the EU, the most consequential AI law for “mental” or affective inference is the AI Act, which prohibits some uses precisely because they intrude too far into people’s inner states, including emotion recognition in workplaces and schools, social scoring, and AI-based criminal-risk prediction. In China, by contrast, the law is not mainly trying to protect the forum internum; it is trying to preserve state control over information. China’s rules make “lawfulness” and political acceptability a gating condition for generative AI output, and they combine content restrictions with filing, security assessment, record-retention, and reporting duties. That is the closest major statutory structure I found to a system in which AI providers are legally expected to judge whether a user’s inquiry or output crosses into prohibited ideological or political territory.

A second major pattern is that many laws are not framed as “AI censorship” but nevertheless encourage AI-assisted review and monitoring. The EU’s Digital Services Act requires very large platforms and search engines to identify and mitigate systemic risks including the spread of illegal content and threats to fundamental rights; the UK’s Online Safety Act requires risk assessments around illegal harms and has already produced document requests, fines, and age-assurance requirements; India’s intermediary rules and later amendments have pushed toward traceability and government-linked fact-checking or synthetic-content controls; state-level U.S. child-safety laws increasingly require age-estimation and feed restrictions, which almost inevitably means more user classification and behavioral analysis. In other words, the practical compliance path often runs through automated detection, human review, age inference, or risk scoring, even where the statute itself does not say “monitor thought.”

The strongest due process and user-rights protections in the set reviewed are generally found in the EU’s platform regime, not because it is libertarian, but because it pairs moderation/risk duties with transparency, notice, appeals, and some user-choice rights such as non-profiled feeds. The weakest protections are in the most politically centralized systems. China’s Generative AI Measures contain privacy and complaint-handling clauses, but those protections sit alongside mandatory reporting, security assessment, broad prohibited-content rules, and administrative sanctions. India’s most controversial 2023 “fact-check unit” move was stayed by the Supreme Court and later held unconstitutional by the Bombay High Court precisely because of vagueness, breadth, and free-speech concerns. That litigation is significant because it shows courts can still resist laws that shift from regulating unlawful conduct toward regulating what users may say about the government.

The broader cognitive-liberty concern that regulation can slide from “illegal content” toward policing permissible thought was also raised in the user-provided background materials, which criticized frameworks conditioning protection on “lawful thought” or “lawful inquiry.” That concern is analytically relevant here even though most public statutes do not use that exact phrase.

What counts as restriction in this report

This report uses a deliberately narrow test for whether a law “restricts, censors, or limits AI.” A law is included when it does one or more of the following: requires providers to filter outputs or user activity for illegal or prohibited categories; requires platforms to assess, collect, or preserve user data or prompt/output records for compliance; compels age estimation, recommendation restrictions, or traceability that alter who may ask or receive certain information; bans AI systems that infer or judge affective or behavioral states; or creates takedown and liability regimes that are likely to induce automated or human review of user communications. That approach matters because many modern speech-control and behavior-control laws operate through platform obligations, not direct punishment of the end user.

It is equally important to separate direct prompt/output control from adjacent surveillance or inference. China’s generative AI regime is direct: it governs what the model may generate and how the provider must react when a user tries to generate prohibited material. The EU AI Act is indirect in a different way: it does not tell a chatbot to reject political questions, but it does prohibit some uses of AI to infer emotion, categorize protected traits, or assess criminal propensity, thereby limiting the deployment of systems that purport to read minds or dispositions. New York and California’s youth-feed laws are different again: they do not define political ideas as unlawful, but they require or encourage age estimation and algorithmic feed control, which can change what information minors can discover and how they are profiled.

Jurisdictional findings

European Union

The AI Act is the most consequential enacted AI law in the EU, but it is not a classic censorship statute. Its logic is mainly rights-protective and risk-based. The Commission’s official summary states that the Act prohibits, among other things, “social scoring,” “individual criminal offence risk assessment or prediction,” “emotion recognition in workplaces and education institutions,” biometric categorization that deduces protected characteristics, and certain forms of real-time remote biometric identification. Those prohibitions became applicable from 2 February 2025. The same official summary also says that high-risk AI systems in education, employment, access to essential services, law enforcement, migration, and justice must meet logging, documentation, human oversight, cybersecurity, and risk-mitigation duties. For generative AI, the Act requires transparency so that users know they are interacting with a machine and so that AI-generated content can be identified and, in some cases, labeled. These are restrictions on AI deployment and inference, but not on “questions” as such.

The Digital Services Act is where the EU comes closest to an indirect regime of AI-assisted content governance. The Commission’s official DSA page says large platforms and search engines must ensure they do not amplify illegal content and “shape opinion at scale”; specifically, the largest platforms must identify and analyze systemic risks including “the spreading of illegal content, as defined in national or EU laws,” “threats to fundamental rights, such as freedom of expression,” and other public-interest harms. The DSA also creates user-side procedural protections: explanations for content removals, appeals, out-of-court dispute settlement, and non-personalized feed options for users of very large platforms. So the EU model is dual: stronger moderation and risk duties on platforms, paired with stronger transparency and appeal rights than many other jurisdictions provide.

Known enforcement matters under the DSA already show how this regime works in practice. The Commission accepted commitments from TikTok to permanently withdraw the “TikTok Lite” rewards feature after concerns about addictiveness, especially for minors, and later issued its first non-compliance decision and €120 million fine against X for transparency-related DSA violations. Those cases were not framed as “thoughtcrime,” but they demonstrate that the DSA is an active enforcement system capable of reshaping platform design and moderation infrastructure. By contrast, as of the material reviewed here, the AI Act itself had implementation guidance and staged applicability, but no mature public enforcement record comparable to the DSA’s first cases.

United Kingdom

The Online Safety Act 2023 is not an AI-specific law, but it is highly relevant to AI systems that operate as search tools, chat interfaces, or image/video services. Reuters reported that Ofcom required platforms to conduct illegal-content risk assessments by 31 March 2025, with risks spanning terrorism, hate crime, child sexual exploitation, and financial fraud. The Act covers social media, search, and pornography services, and secondary work under the Act has pushed them toward age assurance and documented risk management. In practical compliance terms, those duties incentivize AI or human systems to review user content, recommendations, or access requests for legal categories of harm.

Enforcement has begun and illustrates the law’s coercive reach. Reuters reported that the UK issued its first Online Safety Act fine against 4chan in October 2025 for failing to provide required information about illegal-content risks and risk-assessment documentation; Ofcom also warned that persistent non-compliance could bring escalated penalties and access restrictions. Separate reporting in late 2025 described a larger Ofcom fine against an adult-site provider for inadequate age verification. These are not AI-prompt cases, but they show that the OSA can compel platforms to build systems for user classification, content-risk assessment, and compliance documentation, all of which can spill over into AI-based search or recommendation systems.

The strongest free-expression criticism of the OSA is not that it explicitly punishes “unlawful thought,” but that its broad illegal-harms and children’s-safety duties can create incentives for over-removal, privacy-invasive age verification, and conservative moderation choices. Recent empirical work found that the OSA’s implementation milestones coincided with a measurable rise in UK VPN-related discussion and privacy concerns, especially once age-verification duties for adult content came into force. That does not prove censorship in a legal sense, but it is evidence that platform-safety laws can impose real privacy and access costs even when they are formally aimed at harms rather than ideas.

China

China is the jurisdiction in this review where the law comes closest to a direct system of state-mandated AI output censorship and user-activity control. The official Interim Measures for the Management of Generative Artificial Intelligence Services, issued in July 2023 and effective 15 August 2023, provide that providers and users of generative AI services must obey law and ethics and, in Article 4, must “uphold socialist core values” and must not generate content that incites subversion of state power, overthrows the socialist system, endangers national security or interests, damages the national image, incites separatism, undermines national unity or social stability, promotes terrorism or extremism, promotes ethnic hatred or discrimination, or includes violence, obscenity, pornography, or false and harmful information prohibited by law. This is the clearest enacted statutory text I found that directly requires AI systems to reject politically and ideologically disfavored content.

The same measures go further in Article 14. When providers discover unlawful content, they must “timely take measures such as stopping generation, stopping transmission, and elimination,” carry out rectification such as model optimization, and report to competent authorities. When a user is found using the service for unlawful activity, the provider must take measures such as warnings, function restrictions, suspension, or termination of service, preserve relevant records, and report to authorities. Article 17 requires providers of generative AI services with “public opinion properties or social mobilization capabilities” to undergo security assessment and algorithm filing. Article 19 requires providers to explain training-data sources, scale, categories, annotation rules, and algorithmic mechanisms during supervision and inspections. Those provisions do not merely regulate outputs; they create a reporting, filing, auditing, and user-sanctioning architecture.

China’s rules also contain limited user/privacy safeguards. Article 11 says providers must protect user input information and usage records, may not collect unnecessary personal information, may not unlawfully retain personally identifying input information and usage records, and may not unlawfully provide them to others; the same provision also requires handling user requests to access, copy, correct, supplement, or delete personal information. Article 15 requires complaints and reporting mechanisms. But these safeguards coexist with mandatory record preservation and reporting when illegality is found, which means privacy protection is conditional and subordinate to the censorship/security framework.

China’s 2023 generative AI rules sit on top of earlier platform-governance rules that already targeted algorithmic recommendation and synthetic media. Reputable summaries of the earlier framework note that China’s algorithm-recommendation regime requires providers to file certain algorithms with the CAC and prohibits using such systems to disseminate illegal or harmful content, while deep-synthesis rules require labeling and management of manipulated content. Reputable analysis also reports that the CAC’s compliance activity has included audits focused on politically sensitive topics. That combination—broad prohibited-content categories, filing and security review, politically sensitive audits, and takedown/reporting duties—makes China the most mature example in this survey of law that can force AI or humans to judge whether user activity crosses into state-defined impermissible inquiry.

India

India still does not have a single comprehensive AI statute comparable to the EU AI Act or China’s generative AI rules, but its intermediary regime has repeatedly moved in a direction that can pressure platforms to trace users, suppress unlawful content, and review misinformation-related speech. The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 require significant social media intermediaries to identify the “first originator of the information” in certain circumstances, a traceability obligation that has been one of the most controversial parts of the rules because of its implications for encrypted services and anonymous speech. The same rules were expressly conceived by the government in part to prevent the spread of fake news and other misuse of platforms.

The most speech-sensitive attempt went further in 2023, when India amended the IT Rules to create a government fact-checking mechanism for online content about the government. Reuters reported that the amendment required platforms to make “reasonable efforts” not to host information about the government that a government-appointed fact-checking unit deemed “fake, false or misleading.” That moved the system closer to state-directed truth policing. The measure immediately raised constitutional concerns because it would have made safe-harbor protection depend on compliance with a volatile government truth label for speech on public affairs.

Indian courts pushed back. Reuters reported that the Supreme Court stayed the fact-check unit in March 2024, holding that serious constitutional questions were raised about its effect on freedom of speech. Reputable legal summaries of the later litigation report that the Bombay High Court ultimately held the rule unconstitutional in September 2024, finding it vague, overbroad, and chilling. That makes India an important example of both attempted state-linked AI/content control and institutional resistance to it. It also means that, in India, the law closest to authorizing review of “unlawful” or disallowed discourse about state affairs was attempted and stayed/struck, not finally entrenched.

India also appears to be moving toward more direct regulation of synthetic or AI-generated content. Reporting in 2026 described draft or proposed amendments to the IT Rules that would require labeling of synthetically generated information and would shorten takedown windows. Because I did not locate the full official text in this research pass, I treat those 2026 amendments as in process and not yet fully verified here from primary text.

United States

At the federal level, the most important enacted AI-adjacent speech law I identified is the TAKE IT DOWN Act, signed in May 2025. AP reports that it criminalizes knowingly publishing or threatening to publish non-consensual intimate images, including AI-generated deepfake imagery, and requires covered platforms to remove reported content within 48 hours and prevent duplicates from reappearing. This is a genuine federal restriction on certain AI outputs and platform hosting, but its scope is narrow and tied to sexual privacy harms, not ideology or “unlawful questions.” Critics nonetheless warn that fast takedown deadlines can encourage over-removal and automated filtering.

The most consequential proposed federal bills are broader but still not direct “thought” laws. The NO FAKES Act would create a federal digital-replica right and safe-harbor system for unauthorized AI clones of voice or likeness; it remained pending in Congress in the material reviewed here. The Kids Online Safety Act was reintroduced in 2025 after passing the Senate in 2024 but dying without becoming law; its core method is a youth “duty of care” and design restrictions aimed at algorithmic recommendation and addictive features. The Kids Off Social Media Act would go further by banning under-13 access to social media and forbidding personalized recommendation systems for users under 17. These bills are best understood as algorithmic-access and harmful-design bills, not laws defining prohibited ideas, but their practical compliance model depends heavily on user-age inference, content categorization, and recommendation control.

At the state level, the picture is far more active. New York’s SAFE for Kids Act, signed in June 2024, prohibits “addictive” algorithmic feeds for users under 18 without verifiable parental consent and bans late-night notifications to covered minors; proposed rules later specified age and parental-consent verification mechanisms. California’s SB 976 similarly restricts “addictive feeds” for minors and nighttime or school-hour notifications; after litigation, the law was initially blocked in significant part and then largely upheld by the Ninth Circuit in 2025, with some provisions still in dispute. These laws do not punish questions, but they do require platforms to judge age and to alter what information can be recommended to minors.

California also enacted some of the strongest election deepfake laws in 2024, but at least one was promptly enjoined. AP reported that a federal judge temporarily blocked a California law enabling suits over election-related deepfakes, finding the measure likely violated the First Amendment. That matters because it shows a common U.S. pattern: the more a statute approaches direct regulation of false or manipulative political media, the more likely it is to collide with constitutional speech protections.

Finally, Colorado’s AI Act is one of the most important enacted state AI statutes, but it sits in a different category. It regulates “high-risk” AI systems that play a substantial role in “consequential decisions” in areas like employment, education, finance, government services, health care, housing, insurance, and legal services, and it is aimed at preventing “algorithmic discrimination.” It therefore legalizes and structures some AI-based decision-making rather than censoring inquiry. Recent litigation against the law nonetheless shows how politically and constitutionally contested these frameworks already are.

Middle East and Latin America

In the Middle East, I did not identify, in this research pass, a regionally prominent enacted AI statute as direct and developed as China’s generative AI rules. The more common pattern is that general cybercrime and media laws—rather than dedicated AI acts—do the restrictive work and can then be applied to AI-generated content. Jordan’s 2023 cybercrime law has been widely criticized as threatening anonymity, expression, and access to information. The UAE also continues to police online content aggressively through cybercrime and media rules, while simultaneously moving ahead with ambitious state AI initiatives. That means the Middle Eastern risk profile is often high censorship capacity plus limited AI-specific statutory tailoring, rather than explicitly AI-native censorship law. I treat this part of the map as comparatively under-documented here and therefore lower-confidence than the China/EU/UK/US sections.

In Latin America, the most important current examples are still mostly proposed, especially in Brazil. Brazil’s PL 2630/2020—the “fake news” bill—remains pending and has been described as putting the burden on platforms and search services to find and report illegal material, create content-checking mechanisms, and increase accountability for disinformation. Critics argue that it could encourage censorship and heavy-handed moderation. Brazil’s dedicated AI framework bill, PL 2338/2023, is a separate in-process risk-based AI governance proposal, inspired in part by the EU model, with rights, risk categories, and governance obligations rather than overt “thought policing.” In short: Latin America is important because of large pending platform and AI-framework bills, but I did not identify a major enacted AI censorship statute in the region comparable to China’s rules.

Comparative tables

Core laws and bills

Which regimes come closest to policing inquiry or inner states

Timeline of major developments

timeline
    title Major developments in AI-adjacent censorship, monitoring, and restriction law
    2020 : China's online content ecosystem rules take effect, reinforcing platform responsibility for prohibited content
    2021 : India adopts IT Rules with traceability and intermediary compliance duties
         : EU Digital Services Act is adopted later in 2022, building illegal-content and systemic-risk duties
    2022 : China brings algorithmic recommendation and deep-synthesis regulation into force
    2023 : China issues Interim Measures for Generative AI Services
         : UK enacts Online Safety Act
         : India adds government fact-check-unit amendment to IT Rules
         : Brazil's PL 2630 remains hotly contested and stalls
    2024 : EU AI Act enters into force
         : New York signs SAFE for Kids Act
         : Colorado enacts AI Act
         : California enacts youth-addictive-feed law and election deepfake rules; some provisions are swiftly challenged
         : U.S. Senate passes KOSA, but it does not become law in that Congress
    2025 : EU AI Act prohibited-practice bans begin to apply in February
         : EU GPAI obligations begin to apply in August
         : UK OSA illegal-content duties and age-assurance rollout accelerate
         : U.S. enacts TAKE IT DOWN Act in May
         : India fact-check-unit regime remains blocked and is later held unconstitutional in Bombay High Court
         : California youth-feed law is largely upheld on appeal
    2026 : Colorado AI Act takes effect on June 30
         : India continues considering synthetic-content rule changes
         : Middle East and Latin America still show more movement in proposals and adjacent cyber/media law than in mature AI-specific censorship statutes

Assessment, gaps, and limitations

The highest-confidence conclusion is this: China is the paradigmatic case of enacted law requiring AI providers to judge and suppress disallowed categories of user inquiry and output. The UK and India provide the strongest examples in major democracies of laws or attempted rules that push platforms toward illegal-content review, traceability, or truth-policing, though India’s most aggressive move was at least temporarily checked by courts. The EU is more ambivalent: it imposes stronger platform moderation/risk duties than the U.S. but also creates some of the strongest legal limits on AI systems that try to infer internal states or dangerous propensities. The U.S. remains piecemeal, with the most concrete enacted controls focused on deepfakes, youth safety, and discriminatory consequential decisions rather than ideological speech.

Two uncertainties should be stated plainly. First, the phrase “unlawful thought” is more common in philosophical and rights discourse than in enacted statutes. The closest public-law equivalents tend to be laws about illegal content, false information, social stability, public order, or harmful recommendation systems—not explicit criminalization of ideas in a person’s head. Second, I was able to verify the strongest examples with primary or near-primary sources in the EU and China, and with strong reporting in the UK, India, and the U.S., but the Middle East and Latin America sections are comparatively less complete on AI-specific legislation because those regions presently rely more on adjacent cybercrime/media law or still-pending bills than on mature, directly comparable AI statutes.

A final substantive caution: laws that do not mention “thought” can still produce a thought-policing effect when they create strong incentives for providers to inspect prompts, outputs, recommendation behavior, age status, or politically sensitive topics in order to avoid fines, loss of safe harbor, or access blocking. The real legal question in the next few years will often be less “does the statute say thought?” and more “does the compliance architecture force platforms or models to act as judges of inquiry?” That is already happening in China, arguably happening in parts of the UK/India platform regime, and increasingly emerging in youth-safety and recommender-system laws across U.S. states.

Principal sources

The report prioritizes official or primary materials where available, especially the European Commission’s AI Act and DSA materials and the CAC’s official text of China’s Generative AI Measures. It also relies on high-quality reporting for implementation and litigation where official texts were hard to access or not fully machine-readable, especially Reuters and AP for the UK, India, and the United States. Key sources used include the European Commission’s AI Act page and DSA page, the official CAC text of the 2023 Generative AI Measures, Reuters reporting on the UK OSA and India’s fact-check-unit litigation, AP and Reuters reporting on U.S. federal and state measures, and reputable summaries or bill trackers for pending U.S. and Brazilian proposals.