TL;DR
Most of the world’s largest companies are silent on the AI risks they face. A new study from MIT FutureTech reveals that while companies occasionally acknowledge operational risks like discrimination, they almost entirely ignore the systemic, high-severity threats—such as dangerous capabilities and competitive "races"—that experts fear most.
What Problem Does This Paper Address?
As AI adoption accelerates, governance frameworks (like the EU AI Act or the Hiroshima AI Process) increasingly rely on a critical assumption: that companies can accurately identify, manage, and disclose the risks their AI systems create. However, internal corporate risk management is a "black box" to the public and regulators.
The authors sought to determine if public disclosure—the only visible signal of corporate awareness—actually reflects the reality of AI risk. By comparing what companies say in their annual reports and ESG filings against expert consensus on risk severity, the paper highlights a massive "recognition gap" that could undermine global AI safety efforts.
Methods
The researchers employed a multi-stage, tech-forward pipeline to map the disclosure landscape:
- Sampling: They analyzed 200 companies, including the top 100 global firms by market cap and a random sample of the next 900.
- Document Harvesting: Using LLM agents and systematic searches, they collected 539 public documents (annual reports, AI policies, etc.) from early 2026.
- Classification: They used GPT-5.2 to code thousands of text passages against the MIT AI Risk Repository, a taxonomy of 24 distinct AI risks.
- Benchmarking: These corporate disclosures were then compared to a "Delphi study" involving 272 AI risk experts who rated the same 24 risks based on severity, sector vulnerability, and corporate responsibility.
Datasets
- MIT AI Risk Repository: A comprehensive taxonomy of 24 AI risk subdomains (e.g., "AI misalignment," "power centralization," "toxic content").
- Corporate Corpus: 539 public-facing documents from 200 of the world's largest companies.
- Expert Delphi Study: A companion dataset of judgments from 272 international experts regarding the likelihood and impact of AI harms between 2025 and 2030.
Key Insights and Ideas
- The "Operational vs. Systemic" Divide: Companies are relatively comfortable discussing "operational" risks that have clear legal or reputational precedents, such as discrimination (acknowledged by 41% of the top 100) and privacy. However, they are virtually silent on systemic risks like power centralization or competitive dynamics (racing to deploy at the expense of safety), which were mentioned by 0% of the top 100.
- Expert-Corporate Divergence: There is zero correlation between how severe experts judge a risk to be and how likely a company is to acknowledge it. In fact, some of the risks experts rated as most likely to cause catastrophic harm (e.g., dangerous capabilities) were among the least discussed by firms.
- The Misalignment Paradox: Even among general-purpose AI developers—the companies building the most powerful models—acknowledgment of "AI misalignment" stood at only 15%, while 85% of the same group acknowledged discrimination.
- Vulnerability Doesn't Guarantee Disclosure: While companies in "extremely vulnerable" sectors were slightly more likely to mention specific risks, 77% of risks in these high-vulnerability categories still went unacknowledged.
Key Results
The quantitative findings paint a picture of widespread non-disclosure:
- Total Silence: 32% of the top 100 companies and an estimated 62% of the top 1,000 companies did not publicly acknowledge a single AI risk.
- Concentrated Focus: Just five risks (discrimination, weapons/cyberattacks, fraud, privacy, and devaluation of human creativity) accounted for over half of all corporate acknowledgments.
- Sector Leaders: The Information sector was the most transparent, acknowledging an average of 7.1 risks, compared to nearly 0 in "Other services."
Limitations and Future Work
The authors acknowledge that disclosure is an imperfect proxy for practice. A company might be managing a risk internally but keeping it quiet to maintain a competitive advantage. Conversely, some disclosures might be "symbolic compliance" rather than genuine risk management.
Future Work: * Standardized Reporting: The paper argues for a "calibrated" disclosure scheme where companies report on a common core of risks plus specific risks tied to their sector and role (similar to Scope 1-3 emissions in climate reporting). * Beyond Public Docs: Future research could involve direct surveys or interviews to see if internal risk registers match public statements. * Expanding the Sample: Including non-publicly listed "frontier" developers (like OpenAI or Anthropic) and governance actors would provide a more complete view of the ecosystem.
Summary updated