Is AI Deskilling Developers? It's the Usage Pattern, Not the Tool.

On May 23, Mauro Bieg published "Is AI Causing a Repeat of Frontend's Lost Decade?" on his Mastro blog. The piece sat for six days, then surfaced on the Hacker News front page on May 29 with 329 points and 283 comments. Keith Wagner amplified the argument three days later with an extended note. The frame: AI coding tools are deskilling developers the same way JavaScript frameworks deskilled frontend across the past decade. A craft is being lost. The leak in the abstraction is where craft will still matter.

Bieg is right that deskilling is happening. He is wrong about the cause. The Anthropic Fellows ran a randomized controlled trial in January that nobody in this week's discourse is citing in full.

What did the Trio RCT actually find?

Fifty-two engineers, mostly junior, all with at least a year of Python experience, none with prior exposure to Trio. Trio is an async Python concurrency library chosen specifically because it was novel to the participants. Warm-up. Main task: code two features. Follow-up comprehension quiz administered without AI. Half the engineers had AI assistance during the main task; half coded by hand.

The AI-assisted group scored 50% on the quiz. The hand-coding control scored 67%. Cohen's d of 0.738. P of 0.01. Nearly two letter grades of comprehension gap. The largest drop was on debugging questions. The AI group finished only about two minutes faster than the control, a difference that was not statistically significant.

That 17-point headline is the number Addy Osmani uses in his March 14 Comprehension Debt essay. It is the anchor for ByteIota's March replication and for Daniel Storey's Victoria talk. It is the number the comprehension-debt practitioner cluster has been arguing around for two months. It is the right number to cite if your argument stops at AI is making developers worse at this.

It is not the structural finding.

What is the within-group decomposition nobody is citing?

The Anthropic Fellows did something more interesting after the headline test. They coded each engineer in the AI group by usage pattern. Three patterns clustered low: AI delegation (asking the AI to produce code and accepting it), progressive reliance (starting with hand-coding and shifting to delegation as the task grew), and iterative debugging (running the AI's code, hitting an error, asking the AI to fix it, repeat). All three averaged below 40% on the comprehension quiz.

Three other patterns clustered high: generation-then-comprehension (generate the code with AI, then study it carefully before integrating), hybrid code-explanation (alternate between asking the AI to write code and asking it to explain code), and conceptual inquiry (ask the AI questions about the problem without generating code with it). All three scored at or above 65%. The conceptual-inquiry engineers were the fastest among high-scoring patterns even though they never had the AI generate code at all. They could test concepts immediately, and the questions required them to hold the model of the system in their own head before asking.

Same group. Same tool. Same task. Twenty-five points of comprehension spread between what you do while it works. The within-group gap is larger than the between-group gap. Whether you used AI mattered by 17 points. How you used AI mattered by 25.

Why does this dissolve the deskilling frame?

The deskilling frame treats the tool as the cause. If the tool were the cause, the within-group spread should be small relative to the between-group spread, and it is not. The variable that dominates the outcome is not does this engineer use AI. It is how does this engineer use AI. Two engineers using the same Claude or Cursor or Copilot, looking at the same problem, can finish at the same time with a 25-point comprehension difference between them.

The Bauhaus analogy in Bieg's piece is correct on the trajectory. A craft is being lost. The analogy is wrong on the cause. The craft is being lost by delegation, not by the tool that enables delegation. Bauhaus did not argue against industrialization. It argued for craft inside industrialization, by insisting that the designer remain the author of the artifact even when the process scaled. The press did not deskill designers. The press deskilled designers who stopped designing and started feeding copy into the press.

The manifesto called this distinction before the RCT was public. Agents are collaborators, not delegates was published a year ago. AI amplifies your direction said the same thing in the contrast pair: not prompt and pray. Aim and build. What we wrote about vibe coding called the looks done vs actually done gap is now visible at the cognitive level, with a measurement to attach to it: looks-comprehended versus actually-comprehended, with a 25-point spread per usage mode.

What does this mean for builders?

The remedy that follows from the measurement is concrete. Ask the AI to explain what it is suggesting before you accept it. Ask it for two approaches and the trade-offs between them. Generate the code last, after you understand what you are about to read. Use the AI as a tutor for the parts of the problem you do not know yet, and as a peer reviewer for the parts you do. The inquiry modes that scored above 65% share a structure: the engineer holds the model of the system in their own head and uses the AI to extend or test the model, rather than to replace it.

The remedy that does not follow from the measurement is use AI less. The hand-coding control scored 67%, which is better than the AI delegation group's sub-40%, but it is also lower than the AI inquiry group's 65%-or-higher. Inquiry modes were competitive with hand-coding on comprehension and faster on completion. The right floor is not no AI. It is AI as a tool you ask questions of. The deskilling discourse, by treating the tool as the cause, points the remedy at the wrong variable.

For the domain expert showing up to a build with prompts and prototypes and intent, the question the deskilling discourse will plant is should I be using AI at all? Am I making myself worse at the thing I came here to do? That is exactly the wrong question, but it is the question Bieg's piece sets up. The right question is what am I doing while the AI does what it does? If you are asking why, you preserve the expertise that brought you to the work. If you are accepting what, you erode it. The three layers of collaboration frame holds: the agent-to-agent layer can move faster, but whose hand is on the orchestrator is the question that determines whether comprehension survives. The RCT shows the choice is yours, and the consequence is measurable.

There is one more reason the deskilling frame keeps surfacing despite the measurement that contradicts it. Discourse rewards generalization. "AI deskills developers" is a clean headline. "Engineers who delegate to AI lose comprehension; engineers who inquire of AI maintain it" is a longer headline that requires the reader to think about what they are doing while they use the tool. The cleaner headline travels further. The longer one is what builders need to read.