I spend hours every day with AI tools in my editor, but I have trouble reconciling what I feel with the numbers people throw around online.
Depending on which link you click, AI has either revolutionized engineering or barely moved the needle. Industry roundups like Digital Applied’s collection of AI coding adoption statistics assemble dozens of distinct data points from across the industry, but they combine metrics that have almost nothing to do with each other: raw tool adoption, developer satisfaction, lines of code, and self-reported time savings.
Those figures get cited interchangeably, even when they point in opposite directions. For instance, an interview on ShiftMag highlights that 93% of developers use AI, yet overall team productivity gains remain stuck around 10%. Opening an AI autocomplete tab or asking a chat window to format a regular expression counts as usage. It does not mean the engineering organization is moving twice as fast.
The gap widens further when researchers move away from self-reported surveys and watch people write complex software. The METR study on experienced open-source developers evaluated how AI assistance actually impacts experienced programmers tackling real repository tasks (unverified: the exact measured time difference between developers with and without AI tools in METR’s July 2025 benchmark). Their findings show a far more complicated picture than synthetic code-generation benchmarks suggest. Experienced engineers spend most of their effort understanding existing systems, debugging edge cases, and verifying correctness, areas where current tools offer less leverage than they do on greenfield snippets.
When someone claims AI makes developers 20%, 50%, or 90% more productive, they are usually answering a very specific question that has little to do with day-to-day software engineering. Until we distinguish between typing speed, tool adoption, and genuine problem resolution, none of these numbers tell us how much to rely on the output in front of us.