IA générative

Cobus Greyling put the developer productivity gain from generative tooling at up to 55% in November 2025. The Trésor-Éco note 391 (DG Trésor, June 2026) is narrower: +26% for developers, +14% in customer service, and no significant aggregate employment effect, adoption still sitting near 20% of EU firms. That note carries the one solid negative signal: a 16% drop in employment among exposed 22-25 year olds in the US, with no causality established.

The recurring argument concerns displacement rather than volume. Fred Plais argues the bottleneck did not disappear, it moved upstream, into deciding what to build. SFEIR reads the process side the same way: implementation compressed from weeks to hours, and of Marty Cagan's four big risks the SDLC natively covers only technical feasibility.

Philippe Ensarguet names the cost. Open source ran on an implicit contract in which effort signalled understanding of a project. IA générative decoupled the two by producing plausible-looking contributions with zero comprehension and zero effort. Ghostty issues permanent bans for AI-generated code, tldraw auto-closes external PRs, and cURL shut down its bug bounty programme under the volume of empty submissions.

The field experiment run by Dell'Acqua, Mollick, Lakhani and colleagues (791 professionals, Organization Science 2026) found an individual with AI matching a cross-functional pair without it, +0.37σ against +0.24σ. Human pairs without AI still identified their own best idea more often, roughly 50% against 37%. Evaluative judgment stayed with the humans.