How AI Impacts Skill Formation

A randomized study found developers who used AI to learn a new Python library scored 17% lower on a skill quiz without coding any faster.

2026/01/28 by Judy Hanwen Shen, Alex Tamkin · 100 voices · 5 citations
#cs.CY #cs.AI #cs.HC

paper · pdf

Abstract

AI assistance produces significant productivity gains across professional domains, particularly for novice workers. Yet how this assistance affects the development of skills required to effectively supervise AI remains unclear. Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition in the process. We conduct randomized experiments to study how developers gained mastery of a new asynchronous programming library with and without the assistance of AI. We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library. We identify six distinct AI interaction patterns, three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance. Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-critical domains.

Summary

Researchers ran a randomized experiment where 52 developers learned an unfamiliar Python library (Trio, for asynchronous programming) either with or without a chat-based AI assistant. People who used the AI assistant scored about 17% lower on a follow-up quiz testing debugging, code reading, and conceptual understanding, even though they did not finish the coding tasks any faster on average. Watching screen recordings of every participant, the researchers found that AI users who mainly asked conceptual questions (rather than having AI generate or debug their code for them) kept most of the learning benefit that the no-AI group got.

machine-generated · claude-sonnet-5

Outline

• Introduction — Frames the question: does relying on AI to complete unfamiliar tasks undermine the skills workers need to later supervise AI-generated work?
• Background — Reviews prior findings on AI productivity gains (biggest for novices), cognitive offloading, skill retention after AI use, and overreliance on fallible AI output.
• Framework — Treats AI-assisted task completion as a different 'learning path' than doing the task unaided, and sets two research questions on productivity and skill formation.
• Methods — Describes the randomized, pre-registered design: an unfamiliar library (Trio), a warm-up task, a timed AI/no-AI coding task, and a 14-question, 27-point skill quiz.
• Results: Pilot Studies — Four pilots identify and fix problems (participants secretly using AI in the control group, quiz questions giving away each other's answers) before the main study.
• Results: Main Study — AI use did not speed up task completion but cut quiz scores by 4.15 points (17%, Cohen's d=0.738, p=0.01); the gap held across all experience levels and was largest on debugging questions.
• Qualitative Analysis — Codes screen recordings into six AI-interaction patterns: three built on code delegation or AI-driven debugging scored under 40% on the quiz, three built on asking conceptual questions or explanations scored 65% or higher.
• Discussion — Argues AI-driven productivity gains are not free when new skills are on the line, and that how people use AI (not just whether they use it) determines the learning cost.

machine-generated · claude-sonnet-5

Claims

• Using AI assistance to complete unfamiliar coding tasks reduces subsequent quiz performance on debugging, code reading, and conceptual understanding by about 17% (2 grade points on a 27-point quiz; Cohen's d=0.738, p=0.010), a result that holds after controlling for warm-up task time. [experiment]
• AI assistance does not produce a statistically significant improvement in overall task completion time in this setting. [experiment]
• The skill-formation gap between AI and no-AI groups appears at every level of prior coding experience, not just among novices. [experiment]
• Participants who encountered more errors during the task (especially Trio-specific errors like TypeError and RuntimeWarning) later scored better on debugging and conceptual quiz questions, consistent with independent error-resolution driving skill formation. [experiment]
• Among AI users, interaction patterns dominated by code generation or AI-driven debugging (fastest, most delegated) score far lower on the skill quiz (24-39%) than patterns built around conceptual questions or requested explanations (65-86%). [experiment]
• How AI-generated code is adopted (pasted directly vs. manually retyped) changes completion speed but not quiz score, suggesting cognitive effort matters more for learning than time spent. [experiment]

machine-generated · claude-sonnet-5

Key figure

Figure 1 — Left: workers who used AI assistance to complete tasks with an unfamiliar Python library scored notably lower on tests of conceptual understanding, code reading, and debugging than workers who did the same tasks without AI. Right: of six distinct ways participants used the AI assistant, three that kept them actively thinking (asking conceptual questions or requesting explanations) preserved learning, while three built on delegating code writing or debugging to AI did not.

machine-generated · claude-sonnet-5

Glossary

Trio

A Python library for asynchronous, concurrent programming that participants had never used before the study.

Structured concurrency

A programming style, used by Trio, where concurrently running tasks are always nested inside a clear parent scope so none can outlive or escape it.

Nursery

Trio's construct for launching and supervising a group of concurrent tasks together.

Cognitive offloading

Letting a tool (like an AI assistant) do mental work for you instead of working through it yourself, which can weaken learning.

Between-subjects experiment

A study design where each participant is randomly assigned to only one condition (here, AI or no-AI), so the two groups are compared rather than the same person tried both.

Cohen's d

A standardized measure of how big a difference between two groups is, independent of sample size; larger values mean a bigger effect.

Overreliance

Accepting an AI's output as correct without checking it, even when the AI is wrong.

Local item dependence

A flaw in a test where the answer to one question can be figured out from another question, which inflates scores without reflecting real understanding.

machine-generated · claude-sonnet-5

Audience

Engineering managers, technical onboarding/L&D designers, CS educators, and researchers studying human-AI collaboration who need evidence on whether AI coding assistants help or hurt on-the-job skill acquisition.

prerequisites: Basic familiarity with Python and asynchronous programming concepts, Comfort reading experimental results reported as effect sizes and p-values, No machine learning background required

machine-generated · claude-sonnet-5

Open questions

• Would the skill-formation gap be even larger with agentic or autocomplete-style AI coding tools, where users don't have to type out explicit chat queries?
  The authors argue that composing a chat query forces some minimal thinking time that agentic or autocomplete tools skip entirely, and speculate this could make skill loss worse in the increasingly common agentic coding tools they did not test.
• Do the six AI-interaction personas (e.g., 'Conceptual Inquiry' vs. 'AI Delegation') replicate in a larger sample, given each pattern group in this study had only 2 to 7 participants?
  The paper's practical recommendation (ask conceptual questions, don't delegate) rests on a typology built from very small per-pattern counts drawn from one 52-person study.
• Does the observed skill deficit persist over time or fade with further practice, and does it hold for learning tasks longer than the 35-minute, two-task window used here?
  The study measures only an immediate post-task quiz on a short learning exercise; the authors note task duration may itself explain why AI produced no time savings, leaving open whether the skill-formation results generalize to longer, real on-the-job learning.

machine-generated · claude-sonnet-5

Supplementary links

• https://trio.readthedocs.io/en/stable/ [other]
• https://osf.io/w49e7 [other]
• https://github.com/safety-research/how-ai-impacts-skill-formation [code]

machine-generated · claude-sonnet-5

Citations

• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
• AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking

Cited by

• Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
• AI Assistance Reduces Persistence and Hurts Independent Performance
• Psychological Competence as a Missing Dimension in AI Evaluation
• The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks
• The Effortless Trap: Productive Struggle, AI, and the Illusion of Learning

Discussions

• ‘Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition… We find that AI use impairs conceptual understanding, code reading, and debugging abilit [bsky, 333 points, 6 comments]
• How AI impacts skill formation [hn, 236 points, 5 comments]
• The real problem is that the only people qualified to drive an agentic coding model are people who are intimately familiar with the codebase and *familiarity with the codebase drops when you vibe code [bsky, 200 points, 4 comments]
• I think often of a tweet I saw that was like "if your kid actually reads books, they're gonna be one of the smartest people in like ten years" arxiv.org/abs/2601.20245 [bsky, 73 points, 2 comments]
• Interesting paper, especially interesting it's coming from researchers at Anthropic arxiv.org/pdf/2601.20245 [bsky, 69 points, 3 comments]
• How AI Impacts Skill Formation [lobsters, 53 points, 10 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average" arxiv.org/abs/2601.20245 [bsky, 49 points, 1 comments]
• Essentially employees will get rewarded for deskilling themselves. (see e.g. arxiv.org/abs/2601.20245) I assume the end-goal is to broadly convert programming into a badly paid low-skill job. [bsky, 31 points, 1 comments]
• Lots of half-baked commentary about Anthropic’s Jan 28 paper about AI coding assistants arxiv.org/pdf/2601.20245 . Firstly it is not primarily about productivity. But how AI use affects on-the-job lea [bsky, 22 points, 1 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average." arxiv.org/abs/2601.20245 [bsky, 13 points, 1 comments]
• How AI Impacts Skill Formation arxiv.org/abs/2601.20245 Studied by challenging mostly university-educated Python programmers to use async Python (Trio), so likely quite conceptually challenging. LLM u [bsky, 13 points, 0 comments]
• "How AI Impacts Skill Formation" from Anthropic researchers: > We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency ga [bsky, 12 points, 4 comments]
• Imagine if we relied on oil companies to publish evidence that CO2 emissions cause climate change. This statement against interest by Anthropic illustrates the epistemic vulnerability in which funding [bsky, 9 points, 0 comments]
• 1. AI-assisted coding shows no increase in development velocity. Composing prompts and reading replies is as time-consuming as writing the code. 2. AI-assisted coding impairs developer growth and inhi [bsky, 8 points, 1 comments]
• From the preprint itself (on which the above blog post is based): arxiv.org/abs/2601.20245 [bsky, 8 points, 0 comments]
• interesting paper on how the use of AI assistance in coding affects productivity and skill formation (not a rosy picture): arxiv.org/abs/2601.20245 [bsky, 8 points, 1 comments]
• "How AI Impacts Skill Formation", by Judy Hanwen Shen and Alex Tamkin, both working at Anthropic: arxiv.org/abs/2601.20245 [bsky, 8 points, 0 comments]
• You do have to give Anthropic credit here. It is rare for a lab to publish data questioning its own tools. Meta constantly buries internal findings that challenge their business model. This kind of tr [bsky, 8 points, 0 comments]
• Read the full paper here: [2601.20245] How AI Impacts Skill Formation: arxiv.org/abs/2601.20245 [bsky, 8 points, 0 comments]
• AI is bad episode 109480948: même si elle permet des gains en productivité, l'utilisation de l'IA nuit à l'apprentissage. Cf cette étude qui vient de sortir arxiv.org/pdf/2601.202... #onestfoutus [bsky, 7 points, 1 comments]
• And this is from Anthropic... We need to get LLMs out of learning contexts "Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully ado [bsky, 7 points, 0 comments]
• For today's reading group @elisabassignana.bsky.social presented "How AI Impacts Skill Formation" by Judy Hanwen Shen & Alex Tamkin (2026). Paper: arxiv.org/pdf/2601.20245 #NLProc [bsky, 7 points, 0 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks [bsky, 6 points, 1 comments]
• 3/ Das ist eines der grossen Risiken von KI. Alle stürzen sich in halsbrecherischem Tempo auf die ChatGPT, Claude und Co. Im Irrglauben, dass sie dadurch kompetenter und produktiver werden. Die Studie [bsky, 5 points, 0 comments]
• I don't know enough to talk about the surgery-related de-skilling in the nature article going around, but they also make claims about software, and they're based on this paper arxiv.org/abs/2601.20245 [bsky, 4 points, 1 comments]
• For more details on this research, see the full paper: https://arxiv.org/abs/2601.20245 [bsky, 4 points, 0 comments]
• I need to read through this more carefully but the abstract sure is interesting and conditionally validates my feelings that LLMs have no place in education. arxiv.org/abs/2601.20245 [bsky, 4 points, 0 comments]
• How AI Impacts Skill Formation arxiv.org/pdf/2601.20245 In this paper, researchers validate my view that debugging and mental modeling are the true fault lines of AI use. Those who offload the "thinki [bsky, 4 points, 1 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average." arxiv.org/abs/2601.20245 [bsky, 3 points, 0 comments]
• C'est au tour d'anthropic d'arriver à des conclusions plus que mitigées sur l'acquisition de compétences en parallèle de l'utilisation d'un LLM. arxiv.org/pdf/2601.20245 "Novice workers who rely heavi [bsky, 2 points, 1 comments]
• Andrew Koh just told me about this cool experiment that seems the perfect motivation for our paper: arxiv.org/abs/2601.20245 [bsky, 2 points, 0 comments]
• This is a very interesting study on the use of AI coding and how it impacts skill development. Use of AI improved productivity somewhat, but not materially. Some of the reasons for non-AI coders going [bsky, 2 points, 1 comments]
• AIツールがユーザーのスキル習得にどう影響するかを調べたランダム化比較試験。 新しいPythonライブラリを学ぶタスクで… "" https://arxiv.org/pdf/2601.20245 [bsky, 2 points, 0 comments]
• Interesting study from *Anthropic* that found that AI use didn’t significantly speed up development on a novel task (w/ new Python async lib) but it did hurt understanding. Unsurprisingly, AI seems be [bsky, 2 points, 0 comments]
• Your experience is consistent with the observations in this paper. The situation was even worse (I am guessing you are above avg at learning things). IME AI coding assistants provide speedup for thing [bsky, 2 points, 0 comments]
• Some notes on "How AI Impacts Skill Formation". I read the blog, not the preprint, but I'll link both anyway. www.anthropic.com/research/AI-... arxiv.org/abs/2601.20245 [bsky, 2 points, 1 comments]
• How AI Impacts Skill Formation - "Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill form [bsky, 2 points, 0 comments]
• An interesting study on AI and learning skills: AI-enhanced productivity is not a shortcut to competence. (Authors are part of the Anthropic Fellows Program.) arxiv.org/abs/2601.20245 [bsky, 2 points, 0 comments]
• Don't mind me, just posting the links to the arXiv pages [How AI Impacts Skill Formation] arxiv.org/abs/2601.20245 [Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for [bsky, 1 points, 1 comments]
• I recommend reading the paper to get the full context [bsky, 1 points, 0 comments]
• Works Cited: Wolters Kluwer. Patients, Doctors, and Nurses on AI: Similar Tools, Different Pathways, One Destination (Wolters Kluwer, 2026). Budzyń, K. et al. Lancet Gastroenterol. Hepatol. 10, 896–90 [bsky, 1 points, 1 comments]
• AI makes you stupid. arxiv.org/abs/2601.20245 [bsky, 1 points, 0 comments]
• Studien finner förövrigt inga produktivitets-förbättringar när det kommer till ren tids-förbrukning, för tid spenderat med att koda ersattes mer med att interagera med chatbotten. Använder man coding- [bsky, 1 points, 2 comments]
• "How AI Impacts Skill Formation": arxiv.org/pdf/2601.202... Study run by Anthropic. Shows minimal improvement in time taken for coding task but _significant_ reduction in comprehension. The time saved [bsky, 1 points, 0 comments]
• AI Usage for software development tends to make you a poorer programmer. Meaning human in the loop tends to get less effective over time. arxiv.org/pdf/2601.20245 [bsky, 1 points, 1 comments]
• "Our main finding is that using AI to complete tasks that require a new skill reduces skill formation. Contrary to our initial hypothesis, we did not observe a significant performance boost in task co [bsky, 1 points, 1 comments]
• All the "Code with AI or be left behind!" crap? Yeah, it's crap. "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency [bsky, 1 points, 0 comments]
• Here you go: arxiv.org/abs/2601.20245 It's a decent study, and they go into detail on the various outcomes of AI-use versus no-AI. But pay attention to the section detailing the pilot studies, which p [bsky, 1 points, 0 comments]
• “Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition in the process… We find that AI use impairs conceptual understanding, code reading, and d [bsky, 1 points, 0 comments]
• Eu não vi o @jeffquesado.ulivre.dev colocar no perfil dele, então vou colocar aqui e dar os louros a ele, pq ele que mandou no grupo do ZAP ZAP "Participants who fully delegated coding tasks showed so [bsky, 1 points, 1 comments]
• AI-enhanced productivity is not a shortcut to competence. How AI Impacts Skill Formation (Judy Hanwen Shen & Alex Tamkin) arxiv.org/pdf/2601.20245 [bsky, 1 points, 0 comments]
• "Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-c [bsky, 1 points, 0 comments]
• 🤖👨‍💻 Дослідження Anthropic показало, що ШІ допомагає швидше писати код, але знижує рівень засвоєння навичок у junior-розробників. Ну це логічно... arxiv.org/abs/2601.20245 [bsky, 1 points, 0 comments]
• Is it telling that I wanted to have AI summarize this for me? Probably yes. arxiv.org/abs/2601.20245 [bsky, 1 points, 0 comments]
• ok found one I was looking for: arxiv.org/abs/2601.20245 [bsky, 1 points, 1 comments]
• @getajobmike Yeah, even Anthropic's own research shows this: https://arxiv.org/pdf/2601.20245 [bsky, 1 points, 0 comments]
• Anthropic just published this research gem: “We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average.” a [bsky, 1 points, 0 comments]
• "Treating #AI as a goal (rather than a means) takes the focus away from humans, teams, and skill-building, and makes work fully dependent on automation." Some powerful implications for #education and [bsky, 1 points, 0 comments]
• “Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation.” arxiv.org/abs/2601.20245 [bsky, 1 points, 1 comments]
• How AI Impacts Skill Formation arxiv.org/abs/2601.20245 Python経験はあるが非同期プログラミングフレームワークの経験はないプログラマ52人を、生成AI利用群と非利用群にランダムに振り分けてコーディング課題を実施し、その後で理解度テストを実施した。 ①利用群は非利用群に比べて、17%のスコア低下。 ②プログラミングの経験年数に関わらずスコア [bsky, 1 points, 2 comments]
• How AI Impacts Skill Formation https://arxiv.org/abs/2601.20245 + https://www.anthropic.com/research/AI-assistance-coding-skills [bsky, 0 points, 1 comments]
• Relevant paper from Anthropic: How AI Impacts Skill Formation (Feb 3, 2026) arxiv.org/pdf/2601.20245 [bsky, 0 points, 0 comments]
• Fascinating study about how AI Coding impacts skill formation. Novices learn less from just reading the code, leading to a lack of skill formation (and dependency) in developers who use AI. https://ar [bsky, 0 points, 1 comments]
• Anthropic: AI Coding shows no productivity gains; impairs skill development https://arxiv.org/abs/2601.20245 (https://news.ycombinator.com/item?id=46821360) [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation https://arxiv.org/abs/2601.20245 (https://news.ycombinator.com/item?id=46821360) [bsky, 0 points, 0 comments]
• “6 distinct AI interaction patterns, 3 .. involve cognitive engagement & preserve learning outcomes even when participants receive AI assistance .. AI-enhanced productivity .. not a shortcut to compet [bsky, 0 points, 0 comments]
• "Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-c [bsky, 0 points, 0 comments]
• The effect of AI tools on coding velocity and skill acquisition. Summary: 1) It makes *novice* devs significantly faster. It has little impact on mid to senior devs code velocity. 2) It SIGNIFICANTLY [bsky, 0 points, 1 comments]
• Interesting. And, somehow, not that surprising... arxiv.org/abs/2601.20245 > We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation [bsky, 0 points, 0 comments]
• How AI impacts skill formation https://arxiv.org/abs/2601.20245 https://news.ycombinator.com/item?id=46821360 [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation https://arxiv.org/abs/2601.20245 https://news.ycombinator.com/item?id=46821360 [bsky, 0 points, 0 comments]
• Anthropic: AI Coding shows no productivity gains; impairs skill development https://arxiv.org/abs/2601.20245 [bsky, 0 points, 0 comments]
• 📰 How AI Impacts Skill Formation 🔗 https://arxiv.org/abs/2601.20245 💬 Discuss on HN [bsky, 0 points, 0 comments]
• arxiv.org/pdf/2601.20245 [bsky, 0 points, 1 comments]
• How AI Impacts Skill Formation arxiv.org/abs/2601.20245 #AI [bsky, 0 points, 0 comments]
• "Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-c [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation https://lobste.rs/s/k6fjcb #vibecoding [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation arxiv.org/abs/2601.20245 via Anthropic Fellows Program "We find that #AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering [bsky, 0 points, 0 comments]
• arxiv.org/abs/2601.20245 The damages of vibe coding and AI assisted coding [bsky, 0 points, 0 comments]
• Important to note this is published by Anthropic's Fellows Program. I continue to appreciate the way Anthropic approaches AI, and they again show themselves to be one of the real adults in the room in [bsky, 0 points, 0 comments]
• ⚡ Hackernews Top story: How AI Impacts Skill Formation [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation https://arxiv.org/abs/2601.20245 (http://news.ycombinator.com/item?id=46821360) [bsky, 0 points, 0 comments]
• Do you think big tech will put in policy restrictions for employees? "You have to have [these] inhouse competencies before we let you use AI, and you'll have to take an annual exam to continue to qual [bsky, 0 points, 0 comments]
• IA é o novo ópio, mas ao invés de focar na China, estão se expandindo (inclusive com a participação deles também) 🫠 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2601.20245 [bsky, 0 points, 0 comments]
• Anthropic/Claude has studied how developers gained mastery with and without AI. "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering signifi [bsky, 0 points, 0 comments]
• C'est maintenant démontré : l'IA peut rendre plus bête si elle est mal utilisée. Étude complète en anglais à lire ici : arxiv.org/pdf/2601.20245 Résumé en français à lire là : www.linkedin.com/posts/a [bsky, 0 points, 1 comments]
• Jak AI wpływa na naukę nowych umiejętności? Słabo. Jeśli uczycie się czegoś nowego i chcecie, żeby zostało , uczcie się z różnych źródeł, w różny sposób i multimodalnie. Najważniejsze jest by wielokro [bsky, 0 points, 0 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks [bsky, 0 points, 1 comments]
• "We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average." arxiv.org/pdf/2601.20245 [bsky, 0 points, 0 comments]
• A must-read #paper 👇🏽 if you use #AI to boost your #productivity: “How AI Impacts Skill Formation”, Judy Hanwen Shen & Alex Tamkin, Anthropic (arxiv.org/abs/2601.20245). On HN: news.ycombinator.com/ [bsky, 0 points, 1 comments]
• arxiv.org/abs/2601.20245 How AI Impacts Skill Formation [bsky, 0 points, 0 comments]
• Preprint available here: arxiv.org/abs/2601.20245 From abstract: We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency [bsky, 0 points, 1 comments]
• > Our main finding is that using AI to complete tasks that require a new skill (i.e., knowledge of a new Python library) reduces skill formation > Work done as a part of the Anthropic Fellows Program [bsky, 0 points, 0 comments]
• Original Report: arxiv.org/abs/2601.20245 Anthropic’s Post: www.anthropic.com/research/AI-... [bsky, 0 points, 1 comments]
• Studies have shown! arxiv.org/abs/2601.20245 [bsky, 0 points, 1 comments]
• How AI Impacts Skill Formation https://arxiv.org/abs/2601.20245 [bsky, 0 points, 0 comments]
• On commence à avoir des résultats de recherche intéressants sur l'impact des IAG sur les compétences cognitives. Dans cette étude cela portait sur des buts de code sur une bibliothèque python. arxiv.o [bsky, 0 points, 0 comments]
• How AI Impacts Skill Formation https:// arxiv.org/abs/2601.20245 # ai # arxiv [mastodon, 0 points, 0 comments]

Related

• Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers’ awareness?
• From Teacher to Colleague: How Coding Experience Shapes Developer Perceptions of AI Tools
• Echoes of AI: Investigating the Downstream Effects of AI Assistants on Software Maintainability
• Towards Decoding Developer Cognition in the Age of AI Assistants
• Does Co-Development with AI Assistants Lead to More Maintainable Code? A Registered Report