How AI assistance impacts the formation of coding skills \ Anthropic

Research shows AI helps people do parts of their job faster. In an observational study of Claude.ai data, we found AI can speed up some tasks by 80%. But does this increased productivity come with trade-offs? Other research shows that when people use AI assistance, they become less engaged with their work and reduce the effort they put into doing it—in other words, they offload their thinking to AI.
It’s unclear whether this cognitive offloading can prevent people from growing their skills on the job, or—in the case of coding—understanding the systems they’re building. Our latest study, a randomized controlled trial with software developers as participants, investigates this potential downside of using AI at work.
This question has broad implications—for how to design AI products that facilitate learning, for how workplaces should approach AI policies, and for broader societal resilience, among others. We focused on coding, a field where AI tools have rapidly become standard. Here, AI creates a potential tension: as coding grows more automated and speeds up work, humans will still need the skills to catch errors, guide output, and ultimately provide oversight for AI deployed in high-stakes environments. Does AI provide a shortcut to both skill development and increased efficiency? Or do productivity increases from AI assistance undermine skill development?
In a randomized controlled trial, we examined 1) how quickly software developers picked up a new skill (in this case, a Python library) with and without AI assistance; and 2) whether using AI made them less likely to understand the code they’d just written.
We found that using AI assistance led to a statistically significant decrease in mastery. On a quiz that covered concepts they’d used just a few minutes before, participants in the AI group scored 17% lower than those who coded by hand, or the equivalent of nearly two letter grades. Using AI sped up the task slightly, but this didn’t reach the threshold of statistical significance.
Importantly, using AI assistance didn’t guarantee a lower score. How someone used AI influenced how much information they retained. The participants who showed stronger mastery used AI assistance not just to produce code but to build comprehension while doing so—whether by asking follow-up questions, requesting explanations, or posing conceptual questions while coding independently.
We recruited 52 (mostly junior) software engineers, each of whom had been using Python at least once a week for over a year. We also made sure they were at least somewhat familiar with AI coding assistance, and were unfamiliar with Trio, the Python library on which our tasks were based.
We split the study into three parts: a warm-up; the main task consisting of coding two different features using Trio (which requires understanding concepts related to asynchronous programming, a skill often learned in a professional setting); and a quiz. We told participants that a quiz would follow the task, but encouraged them to work as quickly as possible.
We designed the coding task to mimic how someone might learn a new tool through a self-guided tutorial. Each participant was given a problem description, starter code, and a brief explanation of the Trio concepts needed to solve it. We used an online coding platform with an AI assistant in the sidebar which had access to participants’ code and could at any time produce the correct code if asked.1
In our evaluation design, we drew on research in computer science education to identify four types of questions commonly used to assess mastery of coding skills:
Our assessment focused most heavily on debugging, code reading, and conceptual problems, as we considered these the most important for providing oversight of what is increasingly likely to be AI-generated code.
On average, participants in the AI group finished about two minutes faster, although the difference was not statistically significant. There was, however, a significant difference in test scores: the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group—or the equivalent of nearly two letter grades (Cohen's d=0.738, p=0.01). The largest gap in scores between the two groups was on debugging questions, suggesting that the ability to understand when code is incorrect and why it fails may be a particular area of concern if AI impedes coding development.
We were particularly interested in understanding how participants went about completing the tasks we designed. In our qualitative analysis, we manually annotated screen recordings to identify how much time participants spent composing queries, what types of questions they asked, the types of errors they made, and how much time they spent actively coding.
One surprising result was how much time participants spent interacting with the AI assistant. Several took up to 11 minutes (30% of the total time allotted) composing up to 15 queries. This helped to explain why, on average, participants using AI finished faster although the productivity improvement was not statistically significant. We expect AI would be more likely to significantly increase productivity when used on repetitive or familiar tasks.
Unsurprisingly, participants in the No AI group encountered more errors. These included errors in syntax and in Trio concepts, the latter of which mapped directly to topics tested on the evaluation. Our hypothesis is that the participants who encountered more Trio errors (namely, the control group) likely improved their debugging skills through resolving these errors independently.
We then grouped participants by how they interacted with AI, identifying distinct patterns that led to different outcomes in completion time and learning.
Low-scoring interaction patterns: The low-scoring patterns generally involved a heavy reliance on AI, either through code generation or debugging. The average quiz scores in this group were less than 40%. They showed less independent thinkin