AI coding tools stretch developer hours

AI coding is starting to look like a classic productivity trap: when code gets easier to produce, teams quietly raise expectations faster than they improve review, testing, and recovery. The result is more throughput on paper and more cognitive load in practice.

• –Berkeley Haas found workers expanded their job scope, filled breaks with prompting, and ran parallel AI workflows, so the gain showed up as denser workdays rather than shorter ones
• –Google’s DORA data suggests higher AI use can come with more software delivery instability, which undercuts the simplistic “AI equals faster and better” narrative
• –Multitudes reported 27.2% more merged pull requests alongside a 19.6% rise in out-of-hours commits, a strong signal that velocity is being subsidized by personal time
• –Anthropic’s January research adds a longer-term concern: developers using AI to finish tasks may learn less about the underlying systems, especially debugging
• –The most useful idea in the piece is the need for an “AI practice” with pauses, sequencing, and human review, because unmanaged acceleration is now a real engineering risk