AI Transparency Is on the Decline (Stanford HAI 2025)

Authors: Stanford HAI researchers (presentation summary) Year: 2025 Venue: Stanford HAI / FMTI briefing Link: https://hai.stanford.edu Tags: papers transparency governance policy ai Created: 2026-04-22 04:22:18Z Last Updated: 2026-04-22 04:22:18Z


Summary

This source argues that AI transparency is declining despite rapid adoption of foundation models. It frames transparency as a prerequisite for accountability, trust, and democratic oversight, and highlights major disclosure gaps in training data, safety testing, governance, and deployment.

Core Ideas

  • Transparency means stakeholders can inspect how systems are built, tested, governed, and deployed.
  • The Foundation Model Transparency Index (FMTI) indicates low and declining disclosure quality across major providers.
  • Opaque systems create risks in hiring, lending, healthcare, education, and public-sector decisions.
  • Competitive pressure, legal liability fears, and weak legal obligations incentivize non-disclosure.

Why It Matters

  • Missing transparency weakens independent audits and policy design.
  • Public institutions may deploy AI tools without meaningful oversight.
  • The presentation treats transparency as a civic right, not just a product feature.

Policy and Action Notes

  • Stronger AI Transparency Regulation is positioned as the missing enforcement layer.
  • Recommended interventions include mandatory model/system disclosures, independent auditors, public registries, and whistleblower protections.
  • Practical actions are listed for individuals, schools, and companies to push for explainability and accountability.

Sources

  • sources/AI_Transparency_Bright.pptx
  • Stanford HAI, FMTI 2025 briefing materials