policy8 min read

AI Transparency in 2026: What Companies Actually Disclose and What They Hide

MisalignAI Research Team|June 16, 2026|

transparencypolicygovernanceai-safety

Why transparency reports matter now

AI transparency reports have moved from voluntary corporate communications to expected governance artifacts. Regulators in the European Union, the United States, and China now reference disclosure obligations in draft rules. Investors ask about them during due diligence. Civil society organizations use them to compare claims against evidence. For anyone evaluating an AI system, a transparency report is the first place to look for whether the builder treats safety as a public accountability issue or a private marketing exercise.

The problem is that transparency reports are not standardized. One company may publish a detailed system card with training data sources, energy estimates, red-team findings, and incident response timelines. Another may release a glossy PDF that repeats high-level safety commitments without showing the methodology, dates, or limitations behind the numbers. The gap between what is disclosed and what is needed to make an informed deployment decision remains wide.

What major AI labs disclose in 2026

OpenAI publishes system cards for major model releases. These cards include evaluation results, red-team summaries, and known limitations. The 2024 GPT-4o system card, for example, reported multimodal safety tests, refusal behavior, and bias evaluation. However, critics note that the underlying training data composition, compute budgets, and detailed incident logs remain outside public view. OpenAI's transparency model is selective: enough to satisfy headline readers, often insufficient for independent auditors.

Anthropic has taken a different approach with its Responsible Scaling Policy and published safety evaluations. The Claude 3.5 Sonnet system card included more detailed red-team methodology than some competitors, and Anthropic has shared pre-deployment evaluation frameworks. Still, the full training corpus, fine-tuning data, and internal safety incident database are not publicly available. The company has argued that some information must remain confidential for competitive and security reasons, a tension that exists across the industry.

Google's Gemini model cards and AI principles reports cover a broader portfolio, including image, video, and audio systems. The advantage of Google's approach is scale: the reports touch on many products. The disadvantage is depth. Because Google ships many models, individual cards can feel standardized, and it is harder to find the specific evaluation that matters for a particular deployment context. Google's 2024 image-generation pause, documented in a blog post, showed how a reactive disclosure can be more informative than a pre-planned card when an incident forces the company to explain what went wrong.

Meta's approach leans toward open model releases with accompanying research papers rather than detailed system cards. The Llama family of models is accompanied by technical reports that describe architecture, training scale, and some safety benchmarks. The transparency gap here is different. Because Meta releases weights for some models, downstream users can perform their own evaluations. But the upstream training data, filtering decisions, and pre-deployment safety testing are less documented than for closed API providers. This shifts the burden of safety evaluation from the builder to the deployer, which is itself a transparency problem.

Microsoft publishes AI transparency notes for its Copilot products and Azure services. These notes are deployment-oriented, focusing on intended use, limitations, and fairness considerations. They are useful for enterprise customers who need procurement documentation. They are less useful for researchers who want to understand model behavior across edge cases or compare systems on a common benchmark.

Chinese companies are also entering the transparency conversation. DeepSeek's technical reports for its reasoning models include training scale, architecture details, and benchmark results. The level of technical detail is competitive with Western labs, though the reports focus more on capability than on safety evaluation, red-team findings, or incident disclosure. Zhipu AI has published model cards for its ChatGLM series, including safety test results and bias evaluation. The trend suggests that Chinese frontier labs recognize transparency as a global expectation, even if the emphasis and framing differ.

What a transparency report should include

A useful transparency report should answer questions that matter to deployers, regulators, and affected users. Based on guidance from Stanford HAI's Foundation Model Transparency Index, the Mozilla Foundation's AI transparency research, and the AI Now Institute's accountability work, the following elements should be present:

Model training data scale and scope. What data was used, how much, from what time period, and what filtering was applied? This is not about revealing trade secrets. It is about giving users a way to estimate whether the model's knowledge overlaps with their use case.

Energy and environmental impact. Training large models consumes significant energy. A transparency report should estimate training and inference carbon footprint, hardware used, and any offset or mitigation claims. This is increasingly required by corporate sustainability standards and procurement rules.

Safety evaluation results. The report should describe pre-deployment testing, including red-team exercises, adversarial robustness tests, bias evaluation, and truthfulness checks. The methodology should be specific enough to reproduce or critique.

Red-team findings and mitigations. Red teaming is now standard for frontier models, but the value depends on what is shared. A transparency report should list the categories of risks discovered, the severity of those risks, and the mitigations applied before release. If risks were accepted rather than fixed, the report should explain why.

Incident reporting and post-deployment monitoring. What happens after launch? A transparency report should describe the monitoring system, what kinds of incidents are tracked, how users can report problems, and how the company responds to documented failures.

Copyright and intellectual property compliance. With lawsuits challenging training data use, transparency reports should explain the company's position on copyrighted material, licensing practices, and any opt-out mechanisms for content creators.

The problems with current transparency reports

Selective disclosure is the most common issue. Companies report what makes them look good and omit what does not. A strong jailbreak resistance score may be reported while a weak long-context stability result is buried. A successful red-team exercise may be described while a failed one is never mentioned.

Lack of comparability is another problem. Each company uses its own benchmarks, its own test sets, and its own scoring methodology. This makes it impossible to compare two models side by side without running independent evaluations, which most users cannot do. The [MisalignAI model score methodology](/posts/model-score-methodology-preview) and the [AI model card reading guide](/posts/ai-model-card-reading-guide) both address this problem by encouraging readers to look for standard dimensions rather than headline numbers.

Lack of third-party verification undermines trust. When a company evaluates its own model, the incentives are clear. Independent audits, academic evaluations, and regulatory inspections add credibility. Very few transparency reports in 2026 include third-party validation of the core safety claims. The Stanford HAI Foundation Model Transparency Index has documented this gap across dozens of model releases, showing that even basic documentation like training data sourcing is often missing.

What to demand going forward

Users, regulators, and enterprise buyers should treat transparency reports as starting points, not endings. Ask for the missing pieces. If a report omits energy estimates, ask why. If safety testing is described in vague terms, request the methodology. If red-team findings are summarized without detail, push for the full categories. The Mozilla Foundation has argued that transparency should be treated as a minimum viable accountability mechanism, not a marketing checklist.

AI Now Institute research has shown that transparency without accountability is insufficient. The report must connect to a decision-making process: who reviewed the findings, what changed as a result, and what happens if the report turns out to be incomplete or misleading? Without those connections, transparency becomes theater.

For readers of MisalignAI, the practical lesson is simple. Read the dimensions before the headline. Check the dates. Look for independent confirmation. And treat every transparency report as a claim that still needs to be verified.

Source notes

Source support: Stanford HAI's Foundation Model Transparency Index provides the comparative framework for evaluating model documentation. Mozilla Foundation AI transparency research frames disclosure as an accountability prerequisite. AI Now Institute work on algorithmic accountability informs the governance and verification discussion.