Two Minute Papers · 2026-07-16 · notable
Two Minute Papers: 'The Dangerous Illusion of AI Coding Skills'
Károly Zsolnai-Fehér covers new research showing that developers overestimate how much AI coding assistants speed them up — self-reported gains outpace measured productivity by a wide margin.
Two Minute Papers unpacks the gap between how much faster developers think AI tools make them and how much faster they actually work.
What is it?
'The Dangerous Illusion of AI Coding Skills' is Károly Zsolnai-Fehér's July 16 video on the growing research literature showing a persistent gap between developers' self-reported productivity from AI coding assistants and the productivity measured in controlled studies. The video walks through recent field trials, RCTs, and telemetry work that all point the same direction.
How does it work?
The video pulls together several studies — including METR's randomized trial with experienced open-source contributors — that measured wall-clock task completion with and without AI assistance and consistently found the AI condition slower even though participants expected and reported the opposite. Zsolnai-Fehér contrasts that with the productivity benefits AI tools DO deliver (task type, seniority, and codebase familiarity all shift the effect) so viewers can separate the hype from the measurable wins.
Why does it matter?
Two Minute Papers is where a lot of practicing engineers first meet new empirical results, and the coding-agent productivity question is now a live budgeting decision at almost every team. Framing the 'illusion' clearly gives engineers language to push back on unqualified productivity claims and to design their own before/after measurements instead of trusting self-report.
Who is it for?
Engineering leaders, coding-agent power users, and researchers who care about how AI is actually changing developer productivity.