Why is AI governance benchmarking broken, and what is the fix?

As reported by the Stanford HAI AI Index 2026, a 2026 survey of roughly 195 AI governance benchmarks, and the Future of Life Institute AI Safety Index, governance benchmarks are fragmented, non-reproducible, and unsigned, and leaderboards rank models whose differences are statistically indistinguishable. The fix is a signed, reproducible measurement card: every proportion carries a Wilson 95 percent confidence interval, overlapping results record a TIE rather than a ranking, UNMEASURED is a first-class status, each card is Ed25519-signed, and every number is recomputable from published rows.

AI Governance Benchmarking Is Broken. Here Is the Signed, Reproducible Fix.

In 2026 the field said out loud what practitioners had suspected for years: AI governance measurement is not yet trustworthy as evidence. As reported by the Stanford HAI Artificial Intelligence Index Report 2026, average foundation-model transparency scored roughly 40 out of 100 — a documented gap between what developers disclose and what a purchaser or regulator can independently verify. As reported by a 2026 survey of roughly 195 AI governance benchmarks, the landscape is fragmented, non-comparable, and non-reproducible: benchmarks rarely publish their splits, scoring, or provenance, so two numbers that look alike cannot be trusted to mean the same thing. And as reported by the Future of Life Institute AI Safety Index, leading developers still score poorly on independent safety assessment. Three independent sources, one conclusion: the numbers exist, but the method behind them usually cannot be audited.

Why current benchmarks fail

The failures are structural, not cosmetic. Splits are undisclosed, so a score cannot be recomputed by anyone outside the lab that produced it. Runs are non-reproducible, because the frozen item bank, the grader, and the harness version are not published alongside the number. Results are unsigned, so there is no tamper-evident record that the score you are reading is the score that was actually produced. And most damaging for procurement: leaderboards present a ranked order even when the gap between adjacent models sits inside the margin of error — a difference that is statistically indistinguishable is shown as first place and second place. A ranking built on noise is worse than no ranking, because it looks like evidence.

The fix: signed, CI-bounded, reproducible measurement cards

The Council of AI answer is a measurement card that a stranger can audit end to end. Every proportion carries a Wilson 95 percent confidence interval — the interval used because it holds up near 0 and 1, per Evan Miller, Adding Error Bars to Evals, arXiv:2411.00640, and Wilson (1927). No model is called a leader unless its interval separates from the fleet (the separated_leaders rule), and head-to-head claims use the paired McNemar test. When two results overlap, the card records a TIE: statistically indistinguishable, never a ranking. Each card is Ed25519-signed under did:web:csoai.org#estate-chain-1 over a canonical body hash, so the number is tamper-evident. Measurement is three-state: only a MEASURED axis earns a number, UNMEASURED is a first-class status shown before anything is measured, and a status that cannot be checked renders UNTESTED rather than a fabricated score. Grading is deterministic, and every number is recomputable from the published rows — anyone re-derives the score from the frozen bank and harness. See the published methodology, the live scoreboard, the offline card verifier, and the open datasets at huggingface.co/datasets/Nicholastempleman/govbench and huggingface.co/csoai.

The honest boundary

We measure; we do not certify. A signed card states what was measured, when, against which frozen bank, and what it never proves — it is not a quality verdict, not a compliance determination, and not investment advice. The determination of whether a system is compliant, safe, or fit for a given market stays with the competent authorities. A measurement body earns its standing by making its own method auditable, not by ruling on outcomes it has no authority to decide.

Standards and sources referenced

Stanford HAI, Artificial Intelligence Index Report 2026 (model transparency finding)
2026 survey of roughly 195 AI governance benchmarks (fragmentation, non-comparability, non-reproducibility)
Future of Life Institute, AI Safety Index (2026)
Evan Miller, Adding Error Bars to Evals, arXiv:2411.00640 (Wilson intervals for evals)
Wilson, E. B. (1927) — confidence interval for a proportion
NIST AI RMF MEASURE function; ISO/IEC 42001:2023
Council of AI measurement methodology (councilof.ai/methodology)