← Signals
Recentyearscience

The AI Index records both fast benchmark movement and doubts about benchmarks

Stanford HAI's 2026 AI Index reported resolution on SWE-bench Verified rising from around 60% to near 100% within a year, accuracy on the OSWorld computer-use benchmark rising from roughly 12% to 66.3%, and a thirty-percentage-point gain on Humanity's Last Exam — while stating that benchmarks face growing questions about their reliability, and recording that robots succeed at only about 12% of real household tasks. Publishing the rate of progress and the doubt about how it is measured in the same document is the correct posture, and it is the posture this Journey takes.

Historical — it happened, and the record is settled.

Subjects

Evidence

Partly verified — Core facts are sourced; some optional detail is deliberately absent.

Recorded on Benchmark saturation

  1. ImageNet is assembled and released
  2. GLUE is saturated and SuperGLUE is built to replace it
  3. A human-parity claim in translation fails under document-level evaluation
  4. Payroll data shows a relative decline in entry-level employment in AI-exposed occupations
  5. An audit finds the most-cited preference leaderboard is optimised rather than merely measured
  6. A systematic review of 445 benchmarks reports pervasive validity failures
  7. A randomised trial finds experienced developers slower with AI assistance
  8. The Commission proposes deferring the high-risk obligations
  9. The second International AI Safety Report reports mixed labour findings
  10. The AI Index records both fast benchmark movement and doubts about benchmarks

Related

Related because they share a subject in the Atlas — never because the text looks similar.