A September 2026 paper testing axe-core against 250 real page-criterion records found the tool most teams gate their CI on misses close to two-thirds of the accessibility violations actually present, because it was built never to cry wolf. Agentic auditors that operate a page instead of parsing it catch far more, including every keyboard trap axe-core missed, but flag almost half their findings wrongly. Figma’s AI accessibility checker, one design layer earlier in the pipeline, draws the same read-only line: it can audit contrast and touch targets but has no way to test whether a keyboard user can escape a modal.
We trace both scores back to the study and to Deque’s own counter-argument about how “coverage” gets measured, and lay out why the gap is behavioral rather than visual — a test no rule-based checker, however precise, is built to run.