What it is
The panel’s first published output is deliberately an assessment, not a set of recommendations. It is organized in four parts: (1) Why this moment? (defining AI and arguing the UN is the right forum); (2) What does the evidence show? (eight cross-cutting assessments); (3) Findings by domain (seven domains from science to child safety); and (4) Evidence gaps, mandate scope, and next steps. Its job is to establish a shared, credible baseline of what is known and unknown, the input the intergovernmental Global Dialogue then uses to debate what to do.
Eight key assessments
- AI capabilities are advancing faster than the ability to measure or govern them.
- Only a handful of actors have trained frontier AI models, a concentration of power and market.
- AI’s inputs and outcomes are geographically and linguistically uneven.
- The AI divide is not just about access, but about the capacity to influence how AI is developed.
- For AI to be useful, it must be supported by an enabling environment (data, compute, skills, institutions).
- Agentic AI (systems that act autonomously) is a governance step change.
- AI can erode shared reality and threaten information integrity.
- AI is transforming human rights, including children’s rights.
Seven domains
Part 3 assesses the evidence domain by domain:
- AI science, advances and trajectories;
- societal applications: science, health, education, agriculture;
- economic implications;
- security, systems and environmental implications;
- human rights, information and democracy;
- cultural and individual flourishing, autonomy and child safety;
- management practices and frameworks, capacity building and reliability.
Evidence the report cites
The report grounds its “faster than we can measure” theme in hard numbers on the pace of capability:
- Benchmark scores climbing steeply: e.g. “Humanity’s Last Exam” from 8% to 45% in 16 months; GPQA Diamond from ~36% (2023) to ~95%; FrontierMath from 19% (Jan 2025) to 88% (2026); and a gold-medal performance at the 2025 International Mathematical Olympiad.
- The length of software tasks AI can complete doubling every 4–7 months.
- Agentic AI in self-driving chemistry labs yielding >10× faster materials discovery.
- Of 7,000+ languages spoken, only ~1,000 have foundations for meaningful AI inclusion.
- Fewer than one third of developing countries have a national AI strategy.
- The “evidence dilemma”: evidence of AI’s impact lags the pace of its deployment.
Why it matters for the SDGs
The report is, in effect, a measurement document, and measurement is SDGCounting’s core concern. Its “evidence dilemma” is a direct warning for SDG accountability: if the impact of a general-purpose technology cannot be measured as fast as it deploys, neither can its effect on the Goals. The language and national-strategy gaps put hard figures on SDG 10’s AI divide, while the domain findings map onto SDG 3 (health), SDG 4 (education) and SDG 2 (agriculture). By separating scientific assessment from policy, the panel aims to give every country the same starting facts, a public good in its own right.
Read it
- Independent International Scientific Panel on AI, official page and report.
- The 19 June briefing to member states, how the co-chairs introduced the work.
- Panel overview.