Skip to content

Usage Policy

Snapshot: based on the 29 October 2025 version of openai.com/policies/usage-policies, with reference to ChatGPT Enterprise Privacy and the Commercial Terms (August 2025 amendment).

1. Current structure (after the October 2025 rewrite)

Section titled “1. Current structure (after the October 2025 rewrite)”

OpenAI has reorganised its user- and developer-facing policy into a three-layer parallel document system:

TierDocumentScopePositioning
Universal termsUniversal PoliciesAll product usersCompliance and anti-abuse floor
Usage policyUsage PoliciesChatGPT end users and API customersScenario-specific prohibitions and duties
Building termsBuilding with our productsDevelopers building on the OpenAI APIDeployment responsibility, content labelling, disclosure obligations

This structure was carved out of the single pre-2024 Usage Policy document. The official rationale is to “make developer duties and consumer duties clearer”; the critical reading is that splitting the document means no single text lets the reader know all prohibited conduct at one glance, and increases the complexity of cross-references among clauses.

Core list of prohibitions and restrictions

Section titled “Core list of prohibitions and restrictions”

The clear prohibitions retained in the October 2025 version fall into roughly six categories:

  1. Weapons and critical infrastructure: “weaponisation” development of military weapons systems (see §4 on the military-prohibition erosion), CBRN weapon R&D, and attacks on critical infrastructure
  2. Child safety: CSAM (child sexual abuse material), sexualisation of minors, grooming-style interactions
  3. Political integrity: impersonating candidates or officials, mass political-persuasion campaigns, automated generation of election disinformation (more generalised wording than the 2024 version)
  4. Privacy and data: unauthorised facial recognition, biometric scraping, re-identification from public data
  5. Automated high-risk decisions: terminal decisions without human review in law, medicine, finance, employment, etc.
  6. Financial and fraud: financial fraud, academic fraud, forged identity documents

2. Version history

Section titled “2. Version history”
VersionDateKey change
First releaseMarch 2022Published with the public release of the GPT-3 API
2023-03 updateMarch 2023Added “prohibition on generating targeted election-campaign materials”
Major 2024-01 update10 January 2024Quietly deleted the explicit prohibition on “using our services for military and warfare” (retaining only “weapons development” and “harm to others”)
2024-10October 2024Unified default behaviour across the three product lines (ChatGPT / API / GPT Store)
2025-10 rewrite29 October 2025Language simplification, clause consolidation; converted the “prohibited categories” into a “principles + examples” structure

The January 2024 change was first reported by The Intercept (“OpenAI Quietly Deletes Ban on Using ChatGPT for ‘Military and Warfare’”) and marks the transition of OpenAI from a “universal mission” narrative to “national-security capitalism.”

3. The “simplification–blurring” trade-off in the October 2025 rewrite

Section titled “3. The “simplification–blurring” trade-off in the October 2025 rewrite”

The October 2025 rewrite is not mere linguistic polishing: it consolidates a large number of previously itemised prohibited scenarios (e.g. “prohibition on minors’ insurance pricing,” “prohibition on micro-targeted political advertising”) into a small set of principles (“no unlawful activity or activity that infringes others”).

Official narrative vs. critical reading

Section titled “Official narrative vs. critical reading”

OpenAI’s official explanation (blog post Updates to our Usage Policies):

  • “Clearer, more readable”
  • “Avoiding redundant enumeration of conduct already covered by law”
  • “Providing enterprise customers with a consistent global baseline”

Academic and industry critiques enter from three angles:

  1. Contextual integrity (Nissenbaum, 2010): Helen Nissenbaum’s contextual-integrity theory notes that norms derive binding force from context-specific commitments; dissolving specific clauses into “comply with local law” effectively outsources policy to jurisdictional plurality, which loses predictability in cross-border deployment.

  2. “Vagueness → enforcement discretion”: Daniel Solove (GWU) has repeatedly noted in the privacy-law context that the vaguer the terms, the greater the deference regulated parties owe to enforcers. The same logic applies to OpenAI: deleting specific prohibited categories does not loosen actual enforcement but transfers decision-making from contractual clauses to the discretion of the internal Trust & Safety team.

  3. Direct contrast with the Anthropic AUP: Anthropic’s Usage Policy (May 2025 edition) moved in the opposite direction — it enumerates High-Risk Use Cases in greater detail and sets out “how enterprise customers can obtain carveouts through the Commercial Terms.” The two labs’ 2025 trajectories are nearly opposite: OpenAI removes specificity and enlarges its own discretion; Anthropic adds specificity and makes customer outcomes more predictable.

The most important question about a use policy is not what it prohibits, but who decides when something falls within a prohibition. — adapted from Kirsten Martin, Ethics of Data and Analytics (2022)

4. The factual disappearance of the military prohibition

Section titled “4. The factual disappearance of the military prohibition”

Before 2023, the OpenAI Usage Policy expressly prohibited “activity that has high risk of physical harm, including: weapons development; military and warfare.” The 10 January 2024 update removed the “military and warfare” phrasing, retaining only “weapons development” and “harm to others.”

From “removing the prohibition” to “active engagement”

Section titled “From “removing the prohibition” to “active engagement””
DateEvent
10 January 2024Usage Policy removes the “military and warfare” phrasing
Late 2024Collaboration with Anduril Industries on counter-drone systems for the US military
Mid-2025OTA contract signed with the US Department of Defense (“Civil Harmony” and other frontier-AI deployments for logistics and cyber defence; specific dollar amounts should be verified against official announcements)
October 2025The Usage Policy rewrite adds “national-security use by the US government and allies” as a permitted scenario
Q1 2026GSA / CIA contract rumours (unconfirmed)

Critiques (Stuart Russell BBC interview 2024; Future of Life Institute 2025 Annual Report):

  • The pivot from “we don’t do military” to “we are a national-security tool” is accompanied by no public value-based re-argumentation
  • A stark contrast with the 2018 Google Project Maven episode (employees protested → company withdrew)
  • It suggests that post-2023 OpenAI’s corporate culture no longer contains internal dissent at Maven-scale

5. Enterprise fragmentation: one policy, many faces

Section titled “5. Enterprise fragmentation: one policy, many faces”

The Usage Policy is not monolithic — across OpenAI’s product tiers the clause set that actually binds varies by product:

ProductDefault Usage PolicyOverridable by contractDefault data use
ChatGPT Free / PlusFully applicableNoMay be used for training (Plus can opt out)
ChatGPT TeamFully applicablePartiallyNot used for training
ChatGPT EnterpriseSome clauses customisable by MSAYesNot used for training
API (default)Fully applicableNo (standard terms)Not used for training (default since March 2023)
API EnterpriseCovered by Commercial TermsYesNot used for training

Practical implication: ChatGPT Enterprise customers may, through the combined Master Services Agreement (MSA) and Commercial Terms, obtain carveouts from some default Usage Policy clauses (e.g. default limits on certain “high-risk automated decisions”). This is openly discussed in enterprise legal-compliance circles but is not disclosed in the text of the Usage Policy — for external observers, the public policy and the enterprise-contract policy are two different systems.

6. Enforcement: the internal Trust & Safety mechanism

Section titled “6. Enforcement: the internal Trust & Safety mechanism”

Enforcement of the Usage Policy depends on OpenAI’s Trust & Safety team (industry-median size, comparable to peer organisations at Meta and Google; exact headcount not publicly disclosed). Enforcement mechanisms include:

  1. Automated detection: the Moderation API (published 2023, upgraded in 2024 to omni-moderation-latest) classifies inputs / outputs
  2. Human review: escalation to humans may lead to API-key suspension, account ban, organisation-wide ban
  3. GPT Store removal: removal of offending custom GPTs from the GPT Store (cumulative 2024–2025 removal counts not disclosed)
  4. Ecosystem eviction: for B2B products calling the API, wholesale revocation of access

Known large-scale enforcement events:

DateActionScale
February 2024Bans on North Korean, Iranian, Russian, and Chinese state-actor accounts5 operations, dozens of accounts
May 2024Influence-operation accounts including “Bad Grammar” and “Doppelganger”5 networks
October 2024First Influence and Cyber Operations Report aggregation20+ operations to date
February 2025Disclosure of the “Peer Review” influence operation targeting ChatGPT (PRC-linked)1 operation
June 2025Bans on Sora 2 CSAM generatorsNot disclosed

See the transparency report for detail.

7. Industry practice: how the policy shapes the product

Section titled “7. Industry practice: how the policy shapes the product”

Usage-Policy clauses directly determine engineering decisions; the following coupling chains are observable from public signals across 2023–2026:

  1. “Prohibit automated high-risk decisions” → ChatGPT Enterprise mandatory logs + audit export; default “not-a-decision advice” disclaimers for medical / legal scenarios
  2. Child-safety floor → age-based filtering in Sora video generation, the DALL-E face policy, and the ChatGPT minors-conversation mode (launched September 2025)
  3. “Election disinformation” → during the 2024 US election, ChatGPT redirected election-related queries to CanIVote.org (later criticised for “avoiding substantive answers”)
  4. “CSAM zero tolerance” → CSAM hash scanning and training-data filtering in collaboration with Thorn and NCMEC (April 2024 joint statement)
  5. Residual weapons prohibition → still prohibits consumer-level weapon-development use; national-security exceptions handled via Commercial Terms rather than the public Policy

8. Comparison with peer frontier labs

Section titled “8. Comparison with peer frontier labs”
DimensionOpenAI Usage Policies (October 2025)Anthropic AUP (May 2025)Google AI Usage RequirementsxAI Acceptable Use
StructureUniversal / Usage / Building three tiersSingle document + High-Risk annexBy product (Gemini, Vertex)Minimal single document
Military usePermits national-security useExpress prohibition of weaponisationProhibits core control of weapons systemsNo explicit prohibition
Political contentPrinciples-based ban on “manipulation”Ban on substantive manipulation + allows genuine draftingBans “deceptive political content”Weak constraints
CSAMZero toleranceZero toleranceZero toleranceZero tolerance
Degree of specificityLow (post-October 2025)HighMediumVery low
Enterprise-carveout pathwayMSA + Commercial TermsCommercial TermsEnterprise contractsOpaque

Cross-firm conclusion: Anthropic’s AUP remains the industry ceiling on specificity and public enforcement disclosure; OpenAI, after the October 2025 rewrite, is closer to Google’s “principles + industry guidance” model; xAI systematically refuses to make the AUP a verifiable commitment.

9. Regulatory interface: the Usage Policy and hard law

Section titled “9. Regulatory interface: the Usage Policy and hard law”

The Usage Policy does not sit in a vacuum; it overlaps with the following hard-law regimes:

  • EU AI Act Art. 5 (prohibited uses), Art. 51–55 (GPAI systemic-risk obligations): the Usage Policy’s coverage of prohibited uses is typically stricter than the AI Act
  • EU DSA Art. 14, 34–35 (VLOP Terms & Conditions transparency and risk assessment): after ChatGPT was designated a VLOP, Usage Policy amendments require prior risk assessment (OpenAI’s October 2025 rewrite did not publish a complete risk assessment, which scholars have questioned)
  • California SB 53 §22757.11 (frontier-developer obligations): requires disclosure of “foreseeable material risk” mitigation, with the Usage Policy as one of the cited compliance documents
  • China’s Generative AI Interim Measures 《生成式人工智能服务管理暂行办法》 Art. 4, 10, 14: OpenAI does not offer services in mainland China, but indirect access through the Azure international edition triggered a Cyberspace Administration of China (CAC) notice in October 2024 (details not public)

10. Further reading

Section titled “10. Further reading”