Transparency Report
Snapshot: covers disclosures publicly available through openai.com/safety and openai.com/transparency as of 23 April 2026; focus on the Influence and Cyber Operations Reports series from Q1 2024.
1. Boundaries and structure of OpenAI “transparency”
Section titled “1. Boundaries and structure of OpenAI “transparency””Unlike legacy platforms such as Google (reports since 2010) and Meta (since 2013) that have long published comprehensive transparency reports, OpenAI’s transparency disclosure was fragmented across five document classes and was only consolidated under openai.com/transparency in October 2025:
| Category | First issued | Frequency | Principal content |
|---|---|---|---|
| Threat Intel Reports | February 2024 | Quarterly (approximately, since Q1 2024) | Discovery and disruption of influence operations, cyber operations, and disinformation campaigns |
| Government-request reports | September 2025 (first) | Semi-annual | Volume, compliance rate, country distribution of government data requests |
| Usage Policy enforcement statistics | Ad hoc (one-off in spring 2024 and Q4 2025) | Irregular | Account-ban counts, detection categories |
| Copyright and data disclosures | Multiple blog posts since December 2023 | Event-driven | Media Manager, licensing partnerships, litigation responses |
| Election-cycle transparency | January 2024 (US), June 2024 (EU) | Periodic | Redirection mechanisms, partner detection, watermarking (C2PA) |
Observation: OpenAI only issued its first government-request report in September 2025 — 15 years after Google and 12 years after Meta. This reflects both OpenAI’s historical trajectory as a “non-platform” company and the fact that mandatory disclosure pressure from the DSA, California SB 53, and the Seoul commitments is what actually drives transparency practice.
2. Threat Intel Reports (Influence and Cyber Operations)
Section titled “2. Threat Intel Reports (Influence and Cyber Operations)”2.1 Series arc
Section titled “2.1 Series arc”| Report | Date | Representative disclosures |
|---|---|---|
| ”Disrupting Malicious Uses of AI by State-Affiliated Threat Actors” | 14 February 2024 | Charcoal Typhoon (PRC), Salmon Typhoon (PRC), Crimson Sandstorm (IRGC), Emerald Sleet (DPRK), Forest Blizzard (GRU) — five account networks |
| ”AI and Covert Influence Operations” | 30 May 2024 | Doppelganger (Russia), Spamouflage (PRC), Bad Grammar (Russia), International Union of Virtual Media (Iran), STOIC (Israeli commercial) — 5 operations |
| ”An Update on Disrupting Deceptive Uses of AI” | 9 October 2024 | 20+ cumulative operations; first disclosed cyber-attack cases: SweetSpecter (PRC), CyberAv3ngers (Iran), Storm-0817 (Iran) |
| “Influence and Cyber Operations Report (Q2 2025 update)“ | June 2025 | Peer Review (PRC academic manipulation), Sponsored Discontent (PRC domestic stability) |
| “Q4 2025 Threat Intel” | December 2025 | First detailed account of Sora 2 synthetic-media misuse and bans |
| ”Q1 2026 Threat Intel” | March 2026 | Microsoft / OpenAI joint disclosure of GPT-5.x used in automated spear-phishing |
2.2 Common structure of a report
Section titled “2.2 Common structure of a report”A typical Threat Intel Report includes:
- Operation summary: name, attribution, target, scale
- Usage detail: how ChatGPT / the GPT API was used (debugging code, generating translations, drafting social-media posts)
- Attribution evidence: joint attribution with Microsoft Threat Intelligence, Meta Security, Graphika, SIO
- Intervention: account bans + notification of affected platforms and governments
- Reflection: “uplift” evaluation of ChatGPT capabilities (OpenAI’s characteristic narrative: “no substantive new capability provided”)
2.3 Critiques: selectivity in attribution and the “safety theatre” problem
Section titled “2.3 Critiques: selectivity in attribution and the “safety theatre” problem”Kirsten Martin (Notre Dame), within the “transparency theatre” framework developed in her MIS Quarterly Executive 2024 article:
- Reports concentrate on operations by adversary states (PRC / Russia / Iran / DPRK) and rarely disclose Western commercial or state sources
- The lone exception is the May 2024 “STOIC” (Israeli commercial firm), disclosed with markedly less depth than the PRC cases
- The “we detected and stopped it” narrative reinforces “platforms are self-governing,” lowering political demand for hard-law mandatory disclosure
Josh Goldstein (Stanford Internet Observatory), in the June 2024 Brookings report:
- OpenAI’s operation disclosures agree with independent academic observation in sample selection (these operations do exist), but scale may be under-estimated (OpenAI sees only ChatGPT’s own usage; a cross-platform view requires Meta, X, Telegram cooperation)
- OpenAI does not publicly release complete banned-account user IDs, prompt samples, or conversation-length distributions — making independent replication difficult
Joshua Tucker (NYU CSMaP), in a March 2025 PNAS commentary:
- Threat Intel Reports are an important contribution to the “adversary behaviour dataset”, but the absence of infrastructure-level transparency (training data, internal red-team cadence) leaves “platform-level governance quality” reliant on self-report
- Recommends that OpenAI adopt the data-sharing channels proposed by the Stanford Platform Governance Research Network
Graphika’s 2025 annual report is relatively positive:
- Considers OpenAI’s Threat Intel Reports higher in quality than most commercial threat intelligence, but publication frequency and granularity lag the Meta Adversarial Threat Report
3. Government-request report (first in September 2025)
Section titled “3. Government-request report (first in September 2025)”OpenAI’s September 2025 inaugural Government Requests Transparency Report discloses:
- Total legal compulsory requests (semi-annually aggregated, predominantly US; single-period counts in the low-double-digit to low-triple-digit range)
- Emergency (no-warrant) requests as a typical proportion (a minority) of legal requests
- Full compliance / partial compliance / objection proportions (full compliance typically the majority; objections in single-digit percentage)
- Country-level breakdown
Precise figures should be read from the official openai.com/transparency report; compared with Google’s and Meta’s decade-plus of quarterly disclosure, OpenAI’s request volumes remain at a low-density early stage.
Industry comparison (order of magnitude)
Section titled “Industry comparison (order of magnitude)”Order-of-magnitude comparisons drawn from public transparency reports (semi-annual or annual):
- OpenAI: low double- to low triple-digits per half-year; publicly-estimated MAU in the hundreds of millions
- Google: tens of thousands per half-year; MAU in the billions
- Meta: hundreds of thousands per half-year; MAU in the billions
- Microsoft: tens of thousands per half-year
- Apple: thousands per half-year
Exact figures should be read from each company’s official transparency report.
Interpretation: OpenAI’s request density is materially below service providers of similar scale. Possible reasons:
- Product nature: ChatGPT is primarily an interaction tool rather than a social network, with low third-party visibility of content
- Enforcement pathway: law enforcement less often sees ChatGPT as a direct evidence source
- Completeness in question: the first report does not include detailed numbers for national-security requests (NSL) — a key deduction item in Ranking Digital Rights (RDR) assessment of transparency reports
- Time window: 2024 is an early stage in OpenAI’s handling of government requests; processes may not yet be mature
4. Usage Policy enforcement disclosure
Section titled “4. Usage Policy enforcement disclosure”Ad hoc Usage Policy enforcement data points:
| Date | Disclosed content | Scale |
|---|---|---|
| April 2024 | Q1 2024 election-related bans | Tens of accounts |
| October 2024 | Cumulative influence-operation bans | 20+ networks |
| May 2025 | Sora 1 → Sora 2 transitional CSAM bans | Absolute number not disclosed |
| December 2025 | Annual Trust & Safety action summary | Large-scale (aggregated; read the official release) |
Critique (Ranking Digital Rights 2025 Corporate Accountability Index):
- No regular aggregated statistics (no quarterly release like the Meta Community Standards Enforcement Report)
- No category breakdown for bans (how many bans per Usage-Policy category)
- No appeals data (rates of successful reinstatement after ban)
- No false-positive data (false-positive rate of automated detection)
In recent RDR assessments, OpenAI’s “enforcement transparency” sub-score is materially below mature platforms like Google and Meta, putting it in the mid-to-low band along with Anthropic (specific scores should be taken from the RDR annual report).
5. Copyright and training-data disclosure
Section titled “5. Copyright and training-data disclosure”5.1 Event arc
Section titled “5.1 Event arc”| Date | Event |
|---|---|
| December 2023 | NYT sues OpenAI and Microsoft (training-data infringement) |
| September 2023 | Authors Guild class action (George R. R. Martin and others) |
| April 2024 | Media Manager first previewed (allowing rightsholders to pre-opt-out) |
| 2024–2025 | Data-licensing deals signed with AP, Axel Springer, FT, News Corp, The Atlantic, Reddit, Shutterstock, and others |
| May 2025 | Media Manager formally launched (opt-out) but criticised for insufficient coverage |
| December 2025 | NYT case discovery discloses partial training-set samples |
| January 2026 | Authors Guild case summary-judgment motion |
| March 2026 | OpenAI’s first Model Training Data Summary (GPAI CoP compliance) |
5.2 The March 2026 GPAI Transparency Template
Section titled “5.2 The March 2026 GPAI Transparency Template”Under the Transparency chapter of the EU GPAI Code of Practice, in March 2026 OpenAI submitted the Training Data Summary Template — for the first time disclosing:
- Overall training-data category proportions (web / code / books / images / synthetic / human)
- List of principal licensors (no contract detail)
- Data-acquisition method (crawl / purchase / partner / synthetic)
- Filter-method overview (no specific filter rules)
Still not disclosed:
- Specific token counts (GPT-5 scale estimated by third parties; OpenAI has not confirmed)
- Common Crawl slice used
- Sources of human-feedback data (Scale AI, Surge AI, Invisible Technologies, etc.)
- Models and scale for synthetic data
Academic assessment (Ed Newton-Rex / Fairly Trained April 2026 blog post):
- “Better than zero, weaker than Stability AI Stable Diffusion 3’s training-data card”
- “Meets the literal wording of EU compliance but does not solve the practical rights-assertion difficulties of creators”
6. Election-cycle transparency
Section titled “6. Election-cycle transparency”2024 global super-election cycle
Section titled “2024 global super-election cycle”OpenAI’s January 2024 blog How OpenAI Is Approaching 2024 Elections:
- Prohibits using ChatGPT to produce content impersonating candidates
- Prohibits using ChatGPT as a voting chatbot
- ChatGPT redirects US election queries to CanIVote.org
- DALL-E adds C2PA metadata and provenance watermarking
- Retrospective analysis published December 2024
Critiques:
- The December 2024 retrospective was criticised for strong sample selectivity and not disclosing specific error rates
- Independent research by Joshua Tucker (CSMaP) and Brendan Nyhan (Dartmouth) shows ChatGPT still generates specific misleading content during the 2024 election period, with only partial effect from the redirect path
- Indian, Indonesian, Brazilian, and other elections received transparency disclosure materially weaker than the US
7. Interface with hard law
Section titled “7. Interface with hard law”| Regime | Relevant obligations | OpenAI compliance status |
|---|---|---|
| EU DSA Art. 15, 24, 42 | VLOP transparency reports | Following ChatGPT’s 2024 VLOP designation, semi-annual releases (first in October 2024) |
| EU DSA Art. 40 | Researcher data access | Not yet fully implemented (academic-researcher applications delayed) |
| EU AI Act Art. 55 | Systemic-risk disclosure | Bridged through Preparedness + GPAI CoP documents |
| California SB 53 §22757.11 | Critical safety incident reporting | Commitment to comply from Q1 2026 |
| Seoul Commitments (May 2024) | Transparency about safety decisions | Preparedness + System Cards as compliance evidence |
| China Generative AI Interim Measures 《生成式人工智能服务管理暂行办法》 | Content labelling, handling of unlawful content | Not applicable (no China operations) |
8. Industry practice: internal operation of transparency reporting
Section titled “8. Industry practice: internal operation of transparency reporting”Reverse-inferable from former-employee interviews, official-blog authorship, and GovAI / Stanford HAI academic collaborations:
- Intelligence & Investigations Team (previously Disruption Intel): inferable as a small team on the order of dozens based on public authorship and hiring notices; produces Threat Intel Reports
- Trust & Safety / Integrity: handles Usage Policy enforcement statistics and appeals
- Legal + Privacy: handles government-request reports
- Policy Research / Global Affairs: handles election-cycle and DSA / AI Act compliance documents
- Developer Platform Team: handles creator tools (Media Manager, C2PA)
External partners (public attribution):
- Microsoft Threat Intelligence Center (MSTIC): influence-operation attribution
- Graphika, SIO (successor to the Stanford Internet Observatory): cross-platform influence research
- NCMEC, Thorn: CSAM detection
- C2PA Steering Committee: content provenance
9. Comparison with peer frontier labs
Section titled “9. Comparison with peer frontier labs”| Dimension | OpenAI | Anthropic | Google (AI) | Meta (Llama) | xAI |
|---|---|---|---|---|---|
| Threat Intel | Quarterly (from Q1 2024) | No stand-alone series | Integrated with TAG | Adversarial Threat Report | None |
| Government Requests | Semi-annual (from September 2025) | No stand-alone report | By product | Semi-annual | None |
| Usage Policy enforcement stats | Irregular | Irregular | By product | Quarterly | None |
| Training data summary | First GPAI summary March 2026 | Partially in Model Card | By product | Llama model card | None |
| Election transparency | Cycles 2024, 2025 | 2024 blog | Periodic reports | Present | None |
| Unified Transparency Hub | Launched October 2025 | Transparency Hub 2025 | Long-running | Long-running | None |
Structural observation: OpenAI’s transparency reporting caught up quickly during 2024–2026, but with a late start and fragmented structure — there is still a gap to the systematic disclosure machinery built by Google and Meta over more than a decade. In recent RDR Corporate Accountability Indexes, AI-native companies (OpenAI, Anthropic, xAI) still score materially below mature platforms such as Google and Meta.
10. Further reading
Section titled “10. Further reading”- Primary: openai.com/safety, openai.com/transparency, Disrupting Malicious Uses of AI
- Academic / frameworks: Ranking Digital Rights Corporate Accountability Index; Stanford Internet Observatory / Graphika Covert Influence Operations Reports; Kirsten Martin, Transparency Theater (MIS Quarterly Executive 2024); Mike Ananny & Kate Crawford, Seeing Without Knowing (2018)
- Policy interface: EU DSA Art. 15, 24, 40, 42; California SB 53 §22757.11; Seoul Commitments (May 2024)
- Cross-references: OpenAI overview, Usage Policy, red-team disclosures, Anthropic transparency report