AI Transparency Starts With the Audience

Sam Wallace is Head of Corporate Governance, Risk, and Responsible Practice at Partnership on AI (PAI), which published the disclosure recommendations discussed in this article. Dunstan Allison-Hope is an independent consultant, and PAI is among his clients.

Earlier this month, Anthropic published its latest Threat Intelligence Report, outlining real-world threats the company disrupted and the trends these cases reveal. These threats spanned topics including surveillance (such as a domestic surveillance system monitoring roughly 25 million SIM cards in Mali), cybersecurity and electronic warfare (such as developing software for electronic warfare and suppression of air defenses in China), and conventional weapons (such as developing guidance, navigation, and control software in Yemen).

Anthropic publishes Threat Intelligence Reports to help governments, cybersecurity teams, and peer AI developers identify and disrupt similar threats. The reports also inform evidence-based AI policy by grounding discussions in real-world cases and are part of Anthropic’s approach to corporate accountability.

Amid all the discussion about existential risk, reports like this remind us of the real near-term impacts and how they may evolve. As AI moves into more industries — from healthcare and education to infrastructure and aerospace — it creates new risks and opportunities. The concrete case studies offer far more specificity than “AI-will-kill-us-all” and show how addressing near-term risks also helps prepare for the long term.

Thus, while fascinating in their own right, these Threat Intelligence Reports underscore the need for a more strategic approach to AI-related disclosure, reporting, and transparency. What information about AI should companies publish, for whom, and what decisions might these reports inform? How can we move toward greater consistency and comparability of disclosure across the AI industry and over time, while remaining nimble amid AI's changing dynamics and impacts?

We have both spent decades addressing similar questions in sustainability/ESG reporting, where we have sought to enhance the quality of disclosures companies make on topics such as climate change, human rights, and community impacts. Based on this experience, we believe three key factors should shape a more strategic approach to reporting by AI companies.

One: Know the audience for the report

Companies should publish reports so that report readers can make informed decisions. However, a challenge in reporting is that multiple audiences need different information to make different decisions.

For example, investors need information about risks and opportunities for a company’s financial position so that they can allocate capital and manage portfolios. Governments need information about the impacts of a company or technology on society and the economy so that they can make good policies and enforce regulation effectively. Civil society organizations need information about a company's impacts on people so they can advocate for human rights protection, social justice, and community well-being. AI researchers need information to inform AI safety and security strategies.

Not all of these needs will be satisfied in a single report; in AI, some readers are interested in the AI model, some in the overall AI system, and some in the corporation.

Two: Focus on material information that helps someone make a decision

A core principle of financial and sustainability reporting standards is that all material information must be disclosed. In formal reporting, information is considered material if its omission, misstatement, or obscuring could reasonably be expected to influence the decisions of the report's primary users.

The challenge for companies is that what the report user considers material varies depending on the decisions they are trying to make and the reader's expertise.

For example, an AI system's performance against various benchmarks may be useful information to AI researchers or customers, but it's unlikely to mean much to investors or policymakers.

Companies often report on programs and practices they have in place. This can be useful, but more high-level information is more likely to be material. What are the real risks, and where do they concentrate in a company's operations and value chain? What is a company’s overall strategy, and how does it compare to its competitors? Who is ultimately responsible for risk management in the company’s management and board, and how is accountability structured?

Much of our work focuses on investors, and there’s a common assumption that investors only care about financial returns. Whether that's true or not, a company’s ability to navigate the AI age depends on a complex set of relationships with customers, governments, society, and natural resources.

The backlash against data centers is one example of how risks from public opinion and regulation can quickly snowball into major threats to the AI business. We can only expect more of this as AI revolutionizes additional industries and, inevitably, impacts stakeholders beyond businesses. We all deserve better information from companies about how they are navigating these complex issues.

Three: Apply high standards of information quality

Formal reporting differs from other forms of communication because it should adhere to principles of information quality, such as consistency, comparability, and timeliness. Formal reporting on a regular schedule differs from publishing a blog post or long-form essay every so often, and AI companies should adopt several well-developed reporting principles when shaping their disclosure strategies. For example:

• Comparability: Use of standardized topics and disclosure formats so readers can track performance over time and benchmark against other companies.
• Consistency: Use of uniform methodologies, reporting boundaries, and categories of information year over year.
• Timeliness: Disclosures should be made on a regular, predictable schedule so information is available when decisions are made.
• Clarity / Understandability: Information should be organized, clear, and accessible to its intended audience.
• Accuracy: Data and qualitative statements should be precise, rigorously calculated, and free from material error or misleading approximations.

Bringing it together

Transparency is often promoted as a core principle in responsible AI, but in practice its meaning and implications vary significantly by context. For example, transparency with individual users about their use of an AI application differs greatly from transparency with enterprise deployers about the risks of an AI system, and different again from corporate-level SEC (Securities and Exchange Commission) filings.

Earlier this year, Partnership on AI published draft Disclosure Recommendations to help companies identify material information about the impacts, risks, and opportunities associated with AI to include in their financial, sustainability, and other public reports. These draft Disclosure Recommendations are currently open for feedback and point to a broader question about what information is actually useful to the people making AI-related decisions.

Corporate disclosure is only one part of that question. System cards, regulator-required reports, and incident reports serve different purposes and audiences. A system card may help a technical audience understand model capabilities and limitations, while an incident report may help regulators, companies or researchers understand what went wrong and how to prevent recurrence. Treating all of these forms of disclosure simply as “transparency” can obscure these important differences.

The goal is not only greater transparency in the abstract, but more useful information for specific decisions. We have mapped these different forms of reporting according to their purpose, intended audience, and usefulness to that audience. As AI companies and their stakeholders rush into an agenda of increased transparency, we encourage a more strategic approach.