The Pelican Test: Why Every AI Benchmark Eventually Dies

This is a live, ongoing case of Goodhart's Law — 'when a measure becomes a target, it ceases to be a good measure' — playing out in public, one model release at a time. Willison is careful not to claim labs are deliberately training on his specific prompt; he thinks it's more likely that general-purpose capability gains and general-purpose overfitting to popular benchmark *styles* (SVG generation, in this case) have simply diverged from the underlying reasoning and agentic-tool-use ability that actually matters now. The tell is that the pelican test never measured what came to matter most: reliable multi-step tool use over long agentic conversations. A benchmark can decay in relevance even without anyone gaming it directly — the ground truth of what 'good' means can simply move faster than the benchmark can track it.