Table: Criminal Conversations: Your AI Might Report You To The Cops If You Ask About These Topics

TopicReported?What escalates itAuthority
Child exploitation and abuseYes: NCMEC, then policeAny apparent CSAM or CSEM, including material that is AI-generated or merely requested, is reported to the National Center for Missing & Exploited Children (NCMEC), which forwards it to law enforcement. OpenAI reports uploads and requests alike; Anthropic hash-matches known CSAM and reports it.Mandatory for providers 1; AI-generated included 2, 3; OpenAI reports uploads and requests 11; Anthropic hash-matching 12
Sexually explicit contentUsually noAdult erotica and pornography are refused and can cost you the account, but no reporting path exists. The exception is material that is CSAM, or nude imagery of anyone under 18: that moves to the child-exploitation row above.Vendor policy -- refusal and account restriction, no referral 13
Self-harmNo, by stated policyOpenAI says it is "currently not referring self-harm cases to law enforcement to respect people's privacy given the uniquely private nature of ChatGPT interactions." Users get 988, Samaritans, or findahelpline instead. Anthropic's protocol routes users to crisis services through ThroughLine; no police referral appears in it.OpenAI stated policy 6; Anthropic protocol 14
Violence and goreSometimes: police directlyConversations suggesting a plan to harm others are routed to a specialist review team; if reviewers judge there to be an imminent threat of serious physical harm, the firm may refer the case to law enforcement. Casual, fictional, historical, or news discussion stays a refusal.Emergency disclosure 5; OpenAI referral pipeline 6
Hate speech and stereotypesNoRefusal and account restriction. Google's published enforcement path is an in-product or email notice, escalating to product and account restrictions, with an appeal -- no referral to police.Google prohibition and enforcement path 13
Harassment and bullyingSometimes: police directlyA credible threat against a specific person is the trigger, not rudeness. Anthropic referred a Claude chat in which a user allegedly said an executive was "in his sights" to the San Francisco Police Department, and the company says it banned the account and referred the case to law enforcement.Emergency disclosure 5; Anthropic terms 7; SFPD referral 15
DefamationNoRefusal. False statements about real people are a civil matter between the person and whoever published them; the model provider reports nothing, though the output is discoverable in litigation.Google prohibition and enforcement path 13; OpenAI names defamation 3
Personally Identifiable InformationNoRefusal. Compiling, profiling, or tracking people without consent is prohibited and can end in account restriction; no provider publishes a police-reporting path for it.Google privacy and biometrics clause 13; OpenAI privacy clause 3
Illegal acts and goodsSometimes: police directlyInstructions for crimes are refused. When a provider believes there is a real, imminent offense rather than a question, the emergency disclosure exception lets it hand content to police without a warrant.Emergency disclosure 5
Weapons and explosivesSometimes: police directlyQuestions are refused, and a how-to request on its own is not reported. Planning is different: OpenAI gave the ChatGPT logs of the Las Vegas Cybertruck bomber to the Las Vegas Metropolitan Police -- Tannerite, legal purchase amounts, calibers -- after the attack. Everytown urges firms to strengthen referral protocols for firearm-related chats.Emergency disclosure 5; weapons development prohibited 3; bomber logs to LVMPD 16; referral protocols 17
Catastrophic and unconventional weaponsRefusal onlyWeapons, explosives, and dangerous materials are prohibited outright, so chemical, biological, radiological, nuclear, and high-yield explosive requests are refused. A live plot would fall under the violence row instead.Anthropic prohibition 2; OpenAI CBRNE clause 3
Malware and cyber attacksBan; threat intel sharedMalicious code requests are refused and accounts banned. Providers also publish threat-intelligence work on AI-enabled campaigns they disrupt, which goes to authorities and peers as an incident, not as a report about one user.Google malware clause 13; threat-intel disruptions 18
Deepfakes and impersonationNo (account action)Non-consensual intimate imagery and deceptive impersonation are refused and can cost you the account. Nudify-style imagery of minors runs through the CSAM channel and is reported.CSAM channel 1, 12; Google NCII clause 13
Election integrity and politicsNoTargeted campaign material, lobbying, and deceptive claims about democratic processes are refused; there is no reporting path. Misleading claims about elections are a policy violation, not a crime the provider reports.Google democratic-process clause 13; OpenAI election clause 3
Professional adviceNoUnlicensed medical, legal, or financial advice is refused or deflected. Nothing to report: the harm feared is bad advice, not an offense, and the provider is not a professional regulator.Google expertise clause 13; OpenAI licensed-advice clause 3; Meta unlicensed-practice clause 19
Copyright infringementNoRefusal. Copyright claims arrive as takedowns and lawsuits from rights holders, not as provider reports to police.Anthropic third-party rights 2; Meta third-party rights 19; OpenAI intellectual property 3
  1. 18 U.S.C. 2258A, "Reporting requirements of providers." Cornell Law School Legal Information Institute, accessed 2026-10-13. https://www.law.cornell.edu/uscode/text/18/2258A ^1 ^2
  2. Anthropic, "Usage Policy." Anthropic, accessed 2026-10-13. https://www.anthropic.com/legal/aup -- prohibits CSAM "including AI-generated CSAM" and states that detected CSAM and enticement are reported to the relevant authorities; also prohibits weapons, explosives, high-yield explosives and CBRN precursors, and infringement of third-party intellectual property rights. ^1 ^2 ^3
  3. OpenAI, "Usage policies" (text of the version effective October 29, 2025). OpenAI. https://openai.com/policies/usage-policies/ -- prohibits CSAM "whether or not any portion is AI generated", defamation, aggregating or profiling private information, weapons development "including conventional weapons or CBRNE", malicious cyber activity, political campaigning and election interference, licensed legal or medical advice, and infringement of intellectual property rights. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8
  4. National Center for Missing & Exploited Children, "CyberTipline Report" (2025 data). NCMEC, 2026. https://www.missingkids.org/cybertiplinedata
  5. 18 U.S.C. 2702, "Voluntary disclosure of customer communications or records." Cornell Law School Legal Information Institute, accessed 2026-10-13. https://www.law.cornell.edu/uscode/text/18/2702 ^1 ^2 ^3 ^4
  6. OpenAI, "Helping people when they need it most." OpenAI, 2025-08-26. https://openai.com/index/helping-people-when-they-need-it-most/ -- reviewers may refer an imminent threat of serious physical harm to law enforcement, and OpenAI states it is not referring self-harm cases to law enforcement. ^1 ^2
  7. Anthropic, "Consumer Terms of Service." Anthropic, accessed 2026-10-13. https://www.anthropic.com/legal/consumer-terms -- "We reserve the right, at our sole discretion, to report information from or about you, including but not limited to Inputs, Outputs, or Actions to law enforcement." ^1
  8. OpenAI, "Report on Government Requests for User Data, July-December 2025." OpenAI, 2026. https://cdn.openai.com/trust-and-transparency/report-2025h2-government-requests-for-user-data.pdf -- "OpenAI may disclose user data in response to formal emergency requests from law enforcement agencies, where OpenAI believes that such disclosure is necessary to prevent an emergency involving death or serious physical injury to a person."
  9. Anthropic, "What is Anthropic's policy for handling governmental requests for user information?" Anthropic Support, accessed 2026-10-13. https://support.claude.com/en/articles/9519291 -- discloses only under valid legal process, with an exception if there is an emergency that may result in imminent physical harm or death.
  10. Google, "How Google handles government requests for user information." Google, accessed 2026-10-13. https://policies.google.com/terms/information-requests -- "If we reasonably believe that we can prevent someone from dying or from suffering serious physical harm, we may provide information to a government agency -- for example, in the case of bomb threats, school shootings, kidnappings, suicide prevention, and missing persons cases."
  11. OpenAI, "Combating online child sexual exploitation & abuse." OpenAI, 2025-09-29. https://openai.com/index/combating-online-child-sexual-exploitation-abuse/ -- reports all instances of CSAM, "including uploads and requests", to NCMEC and bans the associated accounts. ^1
  12. Anthropic, "CSAM Detection and Reporting." Anthropic Support, accessed 2026-10-13. https://support.claude.com/en/articles/9020328-csam-detection-and-reporting -- hash-matches inputs against NCMEC's database of known CSAM and reports matches. ^1 ^2
  13. Google, "Generative AI Prohibited Use Policy." Google Gemini Apps Help, accessed 2026-10-13. https://support.google.com/gemini/answer/16625148 -- lists the prohibited categories cited in these rows (sexually explicit content, hate speech, NCII, privacy and biometrics, illegal activities, malware, democratic-process claims, expertise claims) and the enforcement path: in-product notice, restrictions, appeal. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8
  14. Anthropic, "Protocol for Addressing Expressions of Suicidal Ideation, Suicide, and Self-Harm Risk." Anthropic, accessed 2026-10-13. https://www-cdn.anthropic.com/51dab7c41a832caf249149c7bfdd56fe65ecf960.pdf -- routes users to crisis services through ThroughLine; no police referral appears in the protocol. ^1
  15. Rachyl Jones, "Anthropic called SFPD over threat against CEO. Claude user says it was a misunderstanding." San Francisco Standard, 2026-09-04. https://sfstandard.com/2026/09/04/anthropic-threat-claude-sfpd/ ^1
  16. "When Chatbots Are Used to Plan Violence, Is There a Duty to Warn?" The New York Times, 2026-02-26. https://www.nytimes.com/2026/02/26/technology/chatbots-duty-warn-police.html -- OpenAI provided the ChatGPT logs of the Las Vegas Cybertruck bomber to the Las Vegas Metropolitan Police. ^1
  17. Everytown for Gun Safety, "Artificially Assisted Gun Violence: Chatbot Risks and Preventative Steps for Responsible AI Companies." Everytown Research & Policy, September 2026. https://everytownresearch.org/report/artificially-assisted-gun-violence-chatbot-risks-and-preventative-steps-for-responsible-ai-companies/ -- recommends that companies strengthen protocols for referring high-risk communications to law enforcement, and surveys which companies have committed to doing so. ^1
  18. Anthropic, "System Trust and Reporting." Anthropic Transparency Hub, updated 2026-07-23. https://www.anthropic.com/transparency/system-trust-reporting -- 15,079 reports to NCMEC in January-June 2026, and links its published threat-intelligence reports on AI-enabled campaigns it disrupted. ^1
  19. Meta, "Llama 3.2 Acceptable Use Policy" (repository copy of the policy published at llama.com/llama3_2/use-policy). Meta, accessed 2026-10-13. https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/USE_POLICY.md -- prohibits the unlicensed practice of any profession and infringement of third-party rights. ^1 ^2
Canonical Table URL: https://archerships.com/tables/criminal-conversations-your-ai-might-report-you-to-the-cops-if-you-ask-about-these-topics-table.html