When AI Coding Boosts Skill Yet Encourages Laziness: A Developer’s Paradox

Personal Wake‑Up Call

At 2 a.m. on a Thursday, the author fell asleep after closing the AI coding assistant Cursor, only to wake at 5 a.m. with a mental prompt: “Maybe tweaking the function signature will make the model understand.” The habit of thinking in prompts continued even after the computer was turned off.

A Counter‑Intuitive Study

In July 2025, the research institute METR conducted a controlled experiment with 16 senior open‑source developers (average involvement in projects with >20 k stars and codebases of millions of lines). Participants fixed bugs, added features, and refactored code in their own repositories, either with AI tools or without. The result: developers using AI completed tasks 19 % slower, yet they subjectively felt they were 20 % faster. This perception‑reality inversion occurred among highly experienced engineers, not interns or novices.

The Dopamine Trap

The article attributes the misplaced satisfaction to “intermittent dopamine reinforcement.” Traditional programming involves a long feedback loop (write → test → debug → fix), often taking minutes or hours. AI coding compresses this to seconds: a natural‑language prompt yields a code snippet that “almost works.” The brain reacts more strongly to “almost successful” outcomes than to outright failure, encouraging repeated prompting. A Fastly survey of 791 professional developers found that nearly 80 % say AI makes programming more enjoyable, but enjoyment does not equal efficiency, as the METR study demonstrates.

Senior Developers Are the Most At‑Risk

Contrary to intuition, senior engineers rely on AI more than juniors. Fastly data shows one‑third of senior developers report that over half of the code they ship is AI‑generated, compared with only 13 % of junior developers. CodeRabbit’s analysis reveals that AI‑generated code contains 1.7 × the bugs of hand‑written code, with severe bugs 1.4 × higher. Seniors assume they can “spot‑check” the output because they have read thousands of lines before, but a follow‑up Fastly finding indicates that roughly 30 % of seniors spend as much time reviewing AI code as the time saved, nullifying any speed advantage.

Short‑Circuiting the Thought Process

The workflow has shifted from “understand → design → code” to “prompt → see result → edit.” This truncates the mental model building stage. A comment on Hacker News likens AI‑generated code to a black box: developers cannot interrogate the reasoning behind the code, and thus lose the ability to troubleshoot when problems arise. A veteran with 20 years of experience echoed this, noting that teams end up with runnable code they do not understand.

Embracing, Not Abandoning, AI

The author does not advocate abandoning AI. Inspired by Andrej Karpathy’s concept of “agentic engineering,” the idea is to use AI agents under supervision, preserving code quality while retaining speed. The core shift is that speed and rigor need no longer be mutually exclusive.

Three Personal Rules

Weekly “AI‑free” day : spend a full day writing all code by hand to ensure skills stay sharp.

Review AI code as an interview : treat each line as if a candidate must explain it before it ships.

Measure learning, not lines : output volume is AI’s advantage; personal value lies in judgment, architecture sense, and boundary intuition.

Conclusion

AI coding does make the author stronger by automating repetitive tasks and freeing mental bandwidth for higher‑level decisions. At the same time, it fosters laziness, eroding the deep system‑level intuition that seasoned engineers rely on. Recognizing this “leakage” is essential to plug the gap before skill decay becomes irreversible.

References: METR RCT study (2025), Fastly developer survey (2025), CodeRabbit report, “The Dopamine Trap of Vibe Coding” (2026), “Why Vibe Coding Won’t Build More Successful Products” (2026).