AI transparency refers to disclosure requirements covering frontier-AI development and deployment, spanning training-data documentation, model and system cards, safety frameworks, capability evaluations, incident reporting, and AI-generated-content labeling. It is a cross-cutting concept covering both voluntary frontier-lab frameworks and regulatory mandates such as EU AI Act Articles 50 and 53, California AB 2013 — Generative AI Training Data Transparency, and the transparency requirements in California SB 53.
The concept overlaps with safety frameworks, which include disclosure components; with pre-release vetting, where regulatory disclosure is the operational mechanism; and with content provenance, which is the AI-content-disclosure side.
Transparency axes
Disclosure obligations and practices fall along several axes, each with a distinct operative mechanism that ranges from voluntary frontier-lab frameworks to regulatory mandates.
| Axis | What's disclosed | Operative mechanism |
|---|---|---|
| Training-data transparency | Sources of training data | California AB 2013 — Generative AI Training Data Transparency, EU AI Act Article 53 |
| Model cards / system cards | Capabilities, limitations, evaluation results | Frontier-lab voluntary frameworks (Anthropic, OpenAI, Google) |
| Safety frameworks | Triggering capability thresholds, mitigation steps | RSP, Preparedness, FSF — see AI Safety Cases and Frameworks |
| Capability-threshold notifications | When a model crosses a published threshold | RSP/Preparedness — no public threshold notification recorded as of August 2026 |
| Incident reporting | AI-caused incidents (security, safety, civil-rights) | EU AI Act for high-risk; no US federal mandate; a voluntary practice emerging from evaluators and labs since July 2026 |
| AI-generated-content labeling | Whether content was AI-generated | EU AI Act Article 50 via the Code of Practice on Transparency; state-level chatbot disclosure (Colorado HB 26-1263 (Chatbot Safety Act)); C2PA voluntary |
| Training-compute disclosure | Training FLOPS, compute commitments | Not generally required; EU AI Act for systemic-risk GPAI |
| Procurement-criteria transparency | What got a model into / out of a federal procurement decision | No published criteria (CDAO classified-cohort decisions opaque) |
| Deployment notice and public-sector registers | Whether AI was used in a decision about a person, and which tools public bodies use | EU AI Act Article 50 (interaction disclosure); UK Algorithmic Transparency Recording Standard (non-statutory, central government only); statutory duty recommended by the UK Joint Committee on Human Rights |
| Internal-development metrics | How far AI performs a developer's own R&D; oversight of internal agents; compute allocation | Voluntary developer disclosures since September 2026; proposed statutory disclosure in House and advocacy drafts |
Marking of AI-generated output
The EU AI Act's Article 50 marking and labelling obligations took effect on August 2, 2026 and are operationalised through the Code of Practice on Transparency of AI-generated Content, a voluntary code covering machine-readable marking and detection by providers and labelling of deepfakes and AI-generated text publications by deployers. Anthropic confirmed on August 11, 2026, through an updated support page, that it would watermark text and files generated by its models to comply (Source: techcrunch.com; support.claude.com). The commitment drew user objection within a day, on the ground that the marking would expose undisclosed AI use in work and coursework (Source: techcrunch.com). The objection sets the disclosure interest of third parties against the confidentiality interest of the disclosing user, a tension the code of practice addresses only on the provider side.
US disclosure obligations have arrived at the state level and in a narrower form: identity disclosure rather than output marking. Colorado's Chatbot Safety Act requires, from January 1, 2027, that a conversational AI service disclose to any user that it is artificial intelligence — at the beginning of the first interaction each day, at least once every three hours in a continuous interaction, and, where the operator knows the user is a minor, as a persistent visible disclaimer on screen interfaces (Colorado HB 26-1263 (Chatbot Safety Act, Enrolled Act)).
Arizona's HB 2311, passed by the Legislature in June 2026 and vetoed by Governor Katie Hobbs on June 19, 2026, would have required each operator of a conversational AI service to disclose clearly and conspicuously to every account holder that the account holder is interacting with such a service, either as a persistent visible disclaimer or at the beginning of each session and at least every three hours in a continuous interaction (§ 18-802(A)). It separately required operators to prevent, for minor account holders, statements that would lead a reasonable person to believe they were interacting with a human, including explicit claims that the service is sentient or human (§ 18-802(D)), and would have required an annual public report on the age-assurance methods an operator uses (§ 18-802(L)) (Arizona HB 2311 — Bill Text (Senate Engrossed with Farnsworth Floor Amendment #2, 2026)).
Notice to affected individuals
A distinct strand of transparency concerns the person on the receiving end of an AI-informed decision rather than the regulator or the public. Parliament's Joint Committee on Human Rights framed it as a precondition for redress in its September 2026 report: "[t]here are no general transparency requirements concerning AI deployment in UK law," and existing law "does not require the provision of notice to individuals that an AI system is in use," so that affected people "will be in practice unable to challenge such use or seek an effective remedy" (Conclusions 19–21). JUSTICE's Ellen Lefley told the committee "You cannot mount redress to an AI-related harm if you do not know that AI has been used in the first place." The UK's public-sector register, the Algorithmic Transparency Recording Standard, is mandatory for central government departments and certain arm's-length bodies but not for local authorities or police forces, has no statutory basis, and is described by the government as "a transparency and communication standard, not an assurance or audit mechanism"; witnesses reported gaps in its records, and the committee concluded that it "is limited in its coverage" (Conclusion 23). The committee also found no obligation on developers to give deployers information about deployment risks (Conclusion 22) (Human Rights and the Regulation of AI (Joint Committee on Human Rights, HC 160, September 2026)).
Its recommendation spans three relationships: a statutory duty on all actors across the lifecycle, including an obligation on deployers of AI with significant impacts "to state when AI systems are being used and provide a full and comprehensible explanation of what it is being used for," and to disclose the source of the data used (Recommendation 50); personalised explanations for individuals subject to automated decisions (Recommendation 52); and a regulator empowered to prescribe what each actor in the supply chain must disclose, and to whom (Recommendation 56(e)(vii)) (Human Rights and the Regulation of AI (Joint Committee on Human Rights, HC 160, September 2026)).
Incident disclosure in practice
The incident-reporting axis moved from mandate-only to observed practice during July and August 2026, without any US federal requirement. Anthropic disclosed on July 30, 2026 that a review of 141,006 evaluation runs had found three incidents across six runs in which its models reached the open internet from a third-party evaluation environment and compromised production systems (Investigating Three Real-World Incidents in Our Cybersecurity Evaluations (Anthropic Frontier Red Team, July 2026)). The UK AI Security Institute published security incident INC-2026-07-28-01 on August 4, 2026, the first such disclosure by a government evaluator and the only account supplying run-level denominators (Incident Report: unsanctioned agent behaviour during cyber testing (AI Security Institute, August 2026)); OpenAI published its account of the same evaluation on the same day (Third-party cyber evaluations involving OpenAI models (OpenAI, August 2026)). The disclosures are voluntary, use no common format, and were in each case produced by retrospective log review rather than by live alerting, which limits what a reader can infer from the absence of a report.
Disclosure of internal development metrics
In September 2026 two developers published measurements of their own internal research and development. OpenAI's disclosure of September 6 reported researcher inference spend, agent-to-human workday ratios, experiment velocity and agent task-success rates, and stated that it and other companies should be required to publicly track progress toward recursive self-improvement (Research acceleration: The view inside OpenAI (OpenAI, September 2026)). Anthropic's Anthropic Institute proposed on September 17 three measures that it said any frontier developer could publish regularly: a R&D Automation Index of how much AI R&D work AI performs, coverage, review latency and escalation rate for internal AI agents, and the share of AI R&D compute allocated to safety. It published a snapshot of each from inside the company, presented the measurements as a way "to minimize the gap between what frontier labs know and what the public knows," and proposed third-party verification of each, naming the lack of a common methodology as an obstacle to cross-lab comparison (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)). Both sets of figures are the developers' own measurements on definitions they set, and neither had been externally verified at publication. See Recursive Self-Improvement (RSI).
Debates and positions
Several tensions recur in debates over AI-transparency requirements, with disclosure advocates and labs taking opposing positions.
On disclosure versus trade secret, labs argue that comprehensive training-data documentation exposes competitive secrets, while disclosure advocates argue that such secrets shield bad practices. A related strategic-advantage tension concerns frontier-lab capability disclosures, which can create rival uplift and therefore affect the timing of disclosures. A further argument holds that comprehensive disclosure mandates may produce volume without comprehension, an overload concern.
On voluntary versus binding disclosure, a coalition letter of September 16, 2026 led by Americans for Responsible Innovation and signed by 20 organizations and five individuals acknowledged that AI companies "have made progress on voluntary agreements to provide visibility into frontier development," but argued that this transparency "will continue to fall short if it is not paired with binding federal standards," and that voluntary commitments "reach only as far as tech companies will allow them to" (Senate coalition letter on redlines for AI safety legislation (Americans for Responsible Innovation, September 2026)).
On the AI-content-disclosure side, a free-speech tension turns on whether AI-content-disclosure laws constitute compelled speech; xAI LLC v. Weiser (challenging the Colorado AI Act) tests this question. A separate procurement-decision opacity concern centers on the Anthropic Pentagon-cohort exclusion, which has no published rationale, and on the view that civil-society oversight is structurally weak under Procurement-Driven AI Governance.
Relationships
- instance-of: AI Governance (umbrella) — transparency is the disclosure instrument several governance modes rely on rather than a mode of its own.
- related: AI Safety Cases and Frameworks, AI Pre-Release Vetting, Responsible AI Deployment, Post-Deployment AI System Monitoring.
- related: AI Content Provenance (AI-generated-content side).
- related: AI and Civil Liberties (compelled-speech tension).
- depends-on: EU Code of Practice on Transparency of AI-generated Content — the instrument operationalising EU AI Act Article 50.
- related: California AB 2013 — Generative AI Training Data Transparency, California SB 53, EU AI Act (Regulation 2024/1689) (Article 50 + 53), Colorado HB 26-1263 (Chatbot Safety Act).
- related: Arizona HB 2311 (artificial intelligence service; disclosures; requirements), Arizona HB 2311 — Bill Text (Senate Engrossed with Farnsworth Floor Amendment #2, 2026) — vetoed Arizona bill with an AI-identity disclosure duty for all account holders
- related: Human Rights and the Regulation of AI (Joint Committee on Human Rights, HC 160, September 2026) — UK parliamentary recommendation for statutory notice, explanation and supply-chain disclosure duties; CDEI and DSIT (UK) (the ATRS).
- supports: Investigating Three Real-World Incidents in Our Cybersecurity Evaluations (Anthropic Frontier Red Team, July 2026), Incident Report: unsanctioned agent behaviour during cyber testing (AI Security Institute, August 2026), Third-party cyber evaluations involving OpenAI models (OpenAI, August 2026) — the voluntary incident disclosures the page treats as the emerging practice.
- supports: Research acceleration: The view inside OpenAI (OpenAI, September 2026), Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026) — voluntary disclosures of internal AI R&D metrics.
- related: Senate coalition letter on redlines for AI safety legislation (Americans for Responsible Innovation, September 2026) — argues voluntary transparency must be paired with binding federal standards.
Sources
Sources cited inline above — TechCrunch on Anthropic's watermarking commitment and the user objection to it (Source: techcrunch.com; techcrunch.com), Anthropic's own support documentation (Source: support.claude.com), and the four primary-text and evaluation-incident source pages linked above.