AI coding assistants may be preventing novice developers from gaining real expertise

An essay titled "AI Coding will Prevent Expertise," published on larsfaye.com and subtitled "The need for ongoing friction in long-term skill formation," opens with a quote from OpenAI's Sam Altman: "We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter." The piece is a follow-up to an earlier essay by the same author, "Agentic Coding is a Trap," which introduced what it called the "skilled orchestrator paradox": the skills needed to manage AI coding agents are the same skills that continued use of those agents can erode. The author argues that developers with years or decades of experience predating AI tooling benefit most from these models today, because their knowledge has already "ossified," while developers who entered the field alongside LLMs are pushed, and sometimes required, to use coding assistants that themselves demand a background of expertise to use safely.

The essay calls this the "Expert Novice" problem. The industry tells newcomers they will be "left behind" if they skip AI tools, while also insisting that getting good results from those tools requires "higher-order thinking": writing robust specs, applying sound design patterns, and reviewing every output closely enough to understand it before shipping. That judgment, the essay argues, is itself a product of the friction and experience that AI tools let developers skip, creating a loop where the tools demand an expertise that using them prevents a person from building.

To support this, the essay cites a study recently referenced by JetBrains, a major developer-tools company: "The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers," which observed individual novice programmers in live coding sessions under varying levels of AI assistance. Its findings say participants who leaned on heavy AI assistance "often skipped crucial planning stages, finding that because they hadn't reasoned themselves into this position, Copilot had," and finished with an "illusion of competence rather than true understanding." The most confident, AI-reliant novices ended up having "skipped crucial steps in the programming problem-solving process, and were now lost." By contrast, participants who limited their AI use developed what the study calls "negative expertise," the ability to ignore incorrect or unhelpful suggestions, and used the tools mainly to accelerate solutions they had already worked out themselves. Even one well-prepared student who otherwise planned carefully was derailed after being "enticed by Copilot into quickly producing code," and ended up needing the same tool to fix a bug it had introduced. The essay frames this as "inverted learning": because novices do not yet know what they do not know, an LLM's accommodating, self-directed responses can convince them they understand more than they do.

A second data point comes from the University of Pennsylvania's 2025 study "Generative AI without guardrails can harm learning," which followed 1,000 students learning mathematics with an LLM. Students who used the AI as a crutch performed 17% worse than students working from just a textbook, despite believing they were doing well. A separate "Tutor" version of the study, where students first asked the AI for help and then attempted a problem on their own, saw that group perform 127% better during AI-assisted practice sessions, though on the actual test that group scored about the same as the textbook-only group. The essay also cites Anthropic's 2026 study "How AI assistance impacts the formation of coding skills," which it says reached a similar conclusion: for novice workers, cognitive effort, and even "getting painfully stuck," appears important for building mastery with AI tools.

Under the heading "Pipeline Collapse," the essay questions the broader industry bet that this foundational knowledge will simply stop mattering as models keep improving and take over the role of "developers" themselves. The essay quotes David Cramer, co-founder at Sentry, a performance and error tracking platform, from a recent interview, in the context of AI-generated code: "I think there's a type of person ... that inherently believes that LLM will get better enough that they will go back and fix this stuff, that it will be able to clean up all the junk that's been stacked up along the way. I don't think that's true. I think it's a science experiment."

As a remedy, the essay points to research suggesting that "dialogic AI systems can meaningfully stimulate reflective, critical and independent thinking" when used as a sparring partner rather than an answer generator, similar to the UPenn Tutor setup. It closes by invoking Joel Spolsky's 2002 essay "The Law of Leaky Abstractions," which argued that code-generation tools that abstract things away still leak, and that the abstractions "save us time working, but they don't save us time learning." The author's own proposed approach, "Friction First," is to learn a language's fundamentals before its frameworks, for instance not starting Java with Spring Boot. The text available for this retelling cuts off mid-sentence at that point, before the essay's specific recommendation for learning JavaScript and its closing argument.

Key facts

The essay, published on larsfaye.com, argues that heavy reliance on AI coding assistants stops novice developers from building the hands-on "friction" that produces real expertise, while developers with years or decades of pre-AI experience benefit most from the same tools.
A study JetBrains recently cited, "The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers," found novices who leaned on heavy AI assistance skipped planning steps and finished with an "illusion of competence," while those who limited AI use developed "negative expertise" and performed best.
In UPenn's 2025 study following 1,000 students learning math with an LLM, a group using AI as a crutch performed 17% worse than a textbook-only group, while a separate "Tutor" group, who asked for AI help before solving problems, scored 127% better in practice sessions but about the same as the textbook group on the final test.
Anthropic's 2026 study on coding-skill formation is cited reaching a similar conclusion: cognitive effort, including "getting painfully stuck," appears important for novices to build mastery with AI tools.
Sentry co-founder David Cramer is quoted, in the context of AI-generated code, doubting that models will "go back and fix" the debt piling up from heavy code generation, calling it "a science experiment."

Why it matters

The essay extends the author's earlier "skilled orchestrator paradox": the judgment needed to steer, audit and correct AI coding agents is the same judgment that leaning on those agents can prevent someone from ever developing. If the pattern holds, the industry may be training a generation of engineers who can operate at the surface of AI tools without the deeper expertise to catch their mistakes, even as the same industry treats deep code generation as part of what the essay calls a trillion-dollar bet that this expertise will stop mattering.

Who it affects

Novice and junior developers who entered the field after AI coding assistants became standard, and who lack the pre-AI experience the essay says protects veterans from skill loss, are the direct subject. Managers and teams that mandate AI-assisted workflows for junior staff are implicated too, since the essay argues the pressure to "keep pace" is what pushes novices toward the heaviest, least supervised AI use. The underlying studies' own subjects, including the 1,000 students in the UPenn experiment, show the same dynamic in a classroom setting rather than a workplace.

How to use it

There is no product here, only a proposed habit: use AI as a Socratic sparring partner rather than an answer generator, mirroring the UPenn study's "Tutor" format, where asking for help first and attempting a problem second outperformed letting the model generate the solution outright. The essay's own "Friction First" approach is to learn a language's fundamentals, for example core Java, before its frameworks, such as Spring Boot, rather than let a generation tool abstract the mechanics away from the start.

How solid is it

This is an argued essay, not a primary study: it leans on three outside sources, a study cited by JetBrains, the University of Pennsylvania's 2025 study, and Anthropic's 2026 study, without naming the researchers behind any of them or detailing their full methodology. The essay's own text carries no author name, and the version available for this retelling cuts off mid-sentence before its conclusion, so its final recommendation cannot be reported.

Risks and caveats

The two headline UPenn figures measure different things and should not be read on one scale: the 17% figure describes a crutch-using group's performance against a textbook-only group, while the 127% figure describes a separate Tutor group's improvement in AI-assisted practice sessions specifically, a group that then scored about the same as the textbook group on the actual test. The essay frames David Cramer's quote as being about AI-generated code, but that framing is the essay's own; the quote itself does not name AI coding tools directly.

“Participants thought it was like having a personal tutor. From the data in our study ... we observed that they did not, in fact, use GenAI tools like a personal tutor. In fact, it was quite the opposite.”

— study cited by JetBrains, "The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers"