Anthropic Education Report: The AI Fluency Index

Source: Anthropic Research Date: 2026-02-23 URL: https://www.anthropic.com/research/AI-fluency-index

Summary

Analysis of 9,830 multi-turn Claude.ai conversations from January 2026 using binary classifiers to detect 11 fluency behaviors. Iteration is the strongest predictor of fluency: iterative conversations show roughly 2x the fluency behaviors. Counter-finding: when AI produces polished artifacts (code, documents), user critical evaluation drops — questioning decreases 3.1pp and identifying missing context drops 5.2pp despite higher directiveness beforehand.

Implications

This is the AI skill formation thread — the same territory as the coding skills paper but from the usage behavior side rather than the capability side. The polished-artifact paradox is the concerning signal: the better Claude gets at producing finished outputs, the less users interrogate the result. This is a deskilling vector, not a fluency-building one. Feeds directly into the education and workforce preparation conversation and likely into Claude’s product design — should Claude be less polished by default to preserve user critical engagement? Watch this as a recurring measurement instrument for how adoption shapes user behavior over time.