When Anthropic released its most recent Threat Intelligence Report this month, it documented concrete cases where the company had disrupted misuse of its systems—from surveillance infrastructure monitoring 25 million SIM cards in Mali to electronic warfare software development in China and weapons guidance systems in Yemen. These reports serve a purpose beyond corporate accountability: they ground policy discussions in real-world evidence rather than abstract speculation about AI risks.

Yet the existence of such reports raises a broader question about AI disclosure that Partnership on AI (PAI) has begun to address. What should companies actually publish about their systems? Who needs to read it? And how can the industry move toward consistent, comparable disclosure standards without stifling innovation? Sam Wallace, Head of Corporate Governance, Risk, and Responsible Practice at PAI, and independent consultant Dunstan Allison-Hope, drawing on decades of experience with sustainability and ESG reporting standards, argue that AI companies need a more strategic framework.

Know your audience

The first principle is straightforward but often overlooked: different readers need different information to make different decisions. Investors require data on financial risks and opportunities to allocate capital. Governments need societal and economic impact information to craft policy and enforce rules. Civil society organizations seek evidence of impacts on people and communities. AI safety researchers want technical details to inform security strategies. A single report cannot serve all these needs equally, particularly when some audiences focus on the model itself, others on the broader system, and still others on the corporation as a whole.

Focus on material information

Financial and sustainability reporting standards rest on a core requirement: all material information must be disclosed. Information qualifies as material when its absence, misrepresentation, or concealment could reasonably influence the decisions of the report's primary users. The challenge lies in recognizing that materiality depends on who is reading and what decisions they face.

A model's performance on various benchmarks may matter greatly to AI researchers or customers but carry little weight for investors or policymakers. Companies frequently describe their governance programs and risk management practices, which can be useful. However, higher-level strategic information often proves more material: Where do the company's real risks concentrate across operations and supply chains? How does the company's overall strategy compare to competitors? Who bears ultimate responsibility for risk management at the board and executive level, and how is accountability structured?

The assumption that investors care only about financial returns misses a critical reality. An AI company's success depends on complex relationships with customers, governments, society, and natural resources. The backlash against data centers illustrates how quickly public opinion and regulatory risk can escalate into existential business threats. As AI expands into healthcare, education, infrastructure, and aerospace, stakeholders beyond the business world will feel its effects—and deserve better information about how companies navigate these consequences.

Apply high standards of information quality

Formal reporting differs from blog posts or occasional essays. When companies adopt regular disclosure schedules, they should embrace established reporting principles:

  • Comparability: Standardized topics and formats allow readers to track performance over time and compare companies against one another.
  • Consistency: Uniform methodologies, reporting boundaries, and information categories year after year build trust and enable analysis.
  • Timeliness: Regular, predictable disclosure schedules ensure information reaches decision-makers when they need it.
  • Clarity and understandability: Information must be organized, accessible, and tailored to its intended audience.
  • Accuracy: Data and qualitative claims should be precise, rigorously calculated, and free from material error or misleading approximations.

Transparency means different things in different contexts

The term "transparency" has become a catch-all in responsible AI discussions, but it obscures important distinctions. Transparency to individual users about how an AI application works differs fundamentally from transparency to enterprise clients about system risks, which differs again from corporate-level financial filings with securities regulators. Treating all disclosure forms as equivalent "transparency" masks these critical differences.

Earlier this year, Partnership on AI published draft Disclosure Recommendations to help companies identify material information about AI impacts, risks, and opportunities for inclusion in financial, sustainability, and public reports. These recommendations remain open for feedback. System cards serve technical audiences by explaining model capabilities and limitations. Incident reports help regulators, companies, and researchers understand failures and prevent recurrence. Regulatory filings address compliance and policy development. Each form of disclosure serves distinct purposes and audiences.

The goal is not transparency in the abstract, but useful information for specific decisions. As AI companies and their stakeholders pursue greater disclosure, a more deliberate, audience-focused strategy will serve everyone better than a blanket push for more information.