Programmer Skills & AI Coding Assistants

Programmer cognitive skills and AI coding assistants.

Anthropic's 2026 study found AI-assisted developers scored about 17 points lower (50% vs 67%) on an immediate comprehension quiz when learning a new library. The pattern is consistent across the dev-skill-atrophy research line.

Published May 24, 2026Updated May 25, 2026Reviewed by Senwitt Editorial Team

Summary

Anthropic's 2026 study measured ~17 points lower comprehension (50% vs 67%) when developers learned a new library with AI assistance versus by hand.

A SEPARATE 2025 METR trial (experienced open-source developers — not the Anthropic study) found those developers took ~19% longer on their own tasks with AI tools allowed.

The drop is concentrated in NEW skill formation — skills developers already deeply have are largely preserved.

Developers using AI for conceptual inquiry scored 65%+ on comprehension; those delegating code generation scored under 40%.

The practical response is to use AI deliberately, read every line of AI output, and keep daily unmediated practice on the calendar.

Does AI make developers worse at coding?

The published 2026 research is unusually clean on this. Anthropic's study found developers using AI coding assistance scored about 17 points lower (50% vs 67%) on an immediate comprehension quiz when learning a new library. (A separate 2025 METR trial of experienced developers — a different study — found they took about 19% longer on their own tasks with AI tools allowed; we cite the two separately.) The drop Anthropic measured is concentrated in new skill formation — skills developers already deeply have are largely preserved. The mechanism is encoding: writing code yourself encodes the path; letting AI generate it encodes the output. The fix is calibrated AI use, not abstinence.

What the source says

The Anthropic 2026 protocol randomized 52 mostly-junior developers to learn an unfamiliar Python library (Trio). Some had AI coding assistance available throughout; others coded by hand. Comprehension was measured immediately afterward via a quiz — the authors note this captures immediate comprehension, not long-term retention or on-the-job performance.

The headline finding: the AI-assisted group averaged 50% on the quiz versus 67% for the hand-coders — about a 17-point gap, nearly two letter grades. (The often-quoted “19% slower” figure comes from a separate 2025 METR trial of experienced developers, not this study; we keep the two apart.)

The complementary Psychology Today coverage of cognitive offloading sharpens the conclusion: AI-related skill drift is concentrated innewskill formation. The skills you already have stay reasonably intact; the skills you would have built by struggling through a task on your own are the ones that don't get built when AI handles the struggle.

Practitioner coverage from VirtusLab, Addy Osmani's Substack, Futurism, CIO, and InfoQ all reach consistent conclusions. The phrase "cognitive debt" (originally from the MIT essay-writing study) gets applied directly to engineering by VirtusLab: shipped code that no one understands is a deferred cost that compounds.

What the source does not say

The Anthropic study does not show that AI coding tools are net-negative for engineering productivity. The opposite is widely documented — AI tools ship output faster on average. The question is about skill formation underneath productivity, not about whether the tools are good.

It also does not recommend removing AI from engineering workflows. The published advice across every reputable source is calibration, not abstinence.

And it does not show that experienced engineers lose existing skills. The atrophy pattern is concentrated in early-career developers and in new-skill formation, not in deep skills that have been built over years.

What this means for daily practice

For practicing engineers, the four-habit pattern across the published advice is consistent: use AI for conceptual inquiry more than for code generation, read every line of AI output before committing it, do at least one non-trivial unaided thing per day, and pair-program with a human regularly.

For engineering managers, the structural moves matter: protect non-AI windows in onboarding (where skill formation actually happens), reward understanding alongside velocity in code review, and make pair- programming a default again.

Senwitt's Code Skill is the daily- practice surface for engineers who want to keep the underlying reasoning, prediction, and bug-spotting skills in regular use. The developers persona page is the full walkthrough.

From Senwitt · advertisement

The text above is editorial. What follows is a promotional message from Senwitt, the maker of this site. Senwitt is a brain-exercise app and is not a medical product. Read the full disclaimer in the footer.

References — canonical order.

1.Kosmyna, N., Hauptmann, E., Yuan, Y.T., Situ, J., Liao, X.-H., Beresnitzky, A.V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. arxiv.org/abs/2506.08872. Primary anchor.

2.Stanković, M., Hirche, E., Kollatzsch, S., & Doetsch, J.N. (2026). Comment on: Your Brain on ChatGPT. arXiv:2601.00856. arxiv.org/abs/2601.00856. The methodological critique — paired with Kosmyna.

3.Risko, E.F. & Gilbert, S.J. (2016). “Cognitive Offloading.” Trends in Cognitive Sciences 20(9):676–688. DOI: 10.1016/j.tics.2016.07.002.

4.Sparrow, B., Liu, J., & Wegner, D.M. (2011). “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips.” Science 333(6043):776–778. DOI: 10.1126/science.1207745.

5.Simons, D.J., Boot, W.R., Charness, N., Gathercole, S.E., Chabris, C.F., Hambrick, D.Z., et al. (2016). “Do ‘Brain-Training’ Programs Work?” Psychological Science in the Public Interest 17(3):103–186. DOI: 10.1177/1529100616661983.

6.FTC v. Lumos Labs, Inc. (2016). “Lumosity to Pay $2 Million to Settle FTC Deceptive Advertising Charges for Its ‘Brain Training’ Program.” Stipulated $50M judgment, suspended on payment of $2M. ftc.gov press release (Jan 5 2016).

7.Max Planck Institute for Human Development & Stanford Center on Longevity (2014). A Consensus on the Brain Training Industry from the Scientific Community. Signed by 70 neuroscientists/psychologists. longevity.stanford.edu.

Not brain training. Brain exercise.

Senwitt is a daily brain exercise app, not a brain training program. We do not claim to improve general cognition, prevent cognitive decline, or treat any condition. Independent scientific consensus — the 2014 Stanford Center on Longevity / Max Planck Institute statement signed by 70 neuroscientists, the 2016 Simons et al. review in Psychological Science in the Public Interest, and the FTC's 2016 settlement with Lumos Labs — has concluded that “brain training” claims are not supported by the evidence. Senwitt is built on a different premise: skills you actively practice get sharper; skills you stop practicing fade.

Why we avoid old brain-training claims