AI coding assistants are rapidly becoming standard in software teams, promising faster delivery and fewer tedious tasks. Yet new internal research from Anthropic suggests that the same tools that boost output can quietly hollow out the very skills developers rely on. The company is now warning that without guardrails, AI help risks turning into a long term tax on engineering competence rather than a pure productivity dividend.
At the center of the debate is a simple tension: AI can write working code in seconds, but learning to think like an engineer still takes years of deliberate practice. Anthropic’s own experiments and staff interviews indicate that when developers offload too much of that thinking to models, they may ship more features today while eroding their ability to debug, design, and mentor tomorrow.
The experiment that rattled Anthropic’s own engineers
Anthropic’s concerns are not abstract, they come from a randomized trial in which developers were asked to learn a new asynchronous Python library while some received AI assistance and others did not. In the study described in Anthropic research, participants leaned on an assistant to generate code for tasks using the Trio library. They appeared to progress quickly, but when the training wheels came off, those who had relied on AI showed weaker understanding of the underlying concepts than peers who had struggled through the exercises manually.
The company’s public write up of how AI assistance impacts the formation of coding skills notes that using a helper improved short term task performance but reduced knowledge retention, with the negative effect hovering around the threshold of statistical significance in the core experiment. That pattern is echoed in a broader working paper on AI assistance that finds significant productivity gains across domains, particularly for novices, yet warns that learning outcomes can suffer when people lean too heavily on generated answers. In a video presentation on How AI Impacts, Anthropic researchers describe how participants who received more direct code suggestions ended up with a shallower mental model of Trio’s concurrency primitives.
Productivity gains without real speedups
Inside Anthropic, the tension between feeling more productive and actually moving faster has become a recurring theme. An internal account of AI Productivity Paradox describes engineers who feel like they are “winning and losing at the Same Time,” shipping more features while worrying that their own expertise is stagnating. Konika Dhull, identified as Product and Data An in that account, reports Anthropic staff describing a sense that they are becoming orchestrators of tools rather than deep specialists in systems they once knew inside out.
External observers have seized on the same paradox. A widely shared discussion of Anthropic’s paper highlights that the company did not find a statistically significant speed up in development from AI assisted coding in its controlled Trio experiment, even though participants reported feeling more efficient. A separate summary from MEXC Exchange notes that the measured productivity gains in the randomized trial “weren’t statistically significant,” even as developers leaned heavily on the assistant. That disconnect between subjective smoothness and hard throughput metrics is at the heart of what Anthropic later labels, in a separate Conclusion, The Productivity Paradox.
Evidence that core skills are already slipping
Beyond lab experiments, there are early signs that everyday coding habits are shifting in ways that could weaken fundamentals. A report on how Anthropic study finds AI might be weakening core coding skills describes engineers who increasingly paste entire problem statements into assistants instead of reasoning through them. According to that coverage, Anthropic’s internal researchers worry that habits like this erode the mental muscles developed through manual problem solving, particularly around algorithmic thinking and debugging.
Other write ups have focused on concrete numbers. A crypto focused outlet summarizing the same randomized trial under the banner Anthropic Study Finds reports that skill retention dropped by 17% when developers used AI tools, and that the New randomized trial from Anthropic highlighted how quickly people began to over trust generated code. A separate analysis of how Anthropic research shows notes that Claude Code flipped the software world on its head, But the same internal data suggests that the more developers leaned on the assistant to write entire functions, the less confident they became in their own ability to reason about edge cases and performance.
“Not a shortcut to competence”
Anthropic’s own researchers have started to speak more bluntly about what their findings mean for teams racing to adopt AI. In one interview, they stress that AI tools are “Not a shortcut to competence,” warning that while assistants can improve developer productivity, the technology could also “inhibit skills formation” if used as a crutch. That warning is captured in a detailed analysis of how Anthropic researchers see AI reshaping the path from novice to expert.
Follow up commentary on Anthropic research underscores that skilled devs make better use of AI, but using AI is bad for learning skills, particularly for those still building an understanding of the Trio library. In other words, the very people who stand to gain the most from a productivity boost, junior engineers and career switchers, may also be the ones whose long term growth is most at risk if they never learn to solve problems without autocomplete. That is why Anthropic’s own write up on Does AI assistance impact skill formation leans heavily on the idea that organizations must treat AI as a complement to, not a replacement for, deliberate practice.
Workflows, mentorship, and the future of “agentic” coding
The cultural impact inside Anthropic may be as significant as the experimental data. An internal account titled Anthropic Engineers Reveal describes senior staff who worry that pair programming and code review are being replaced by quiet sessions of prompting. A new internal study from Anthropic cited there notes that some engineers are anxious about long term job relevance as more of the “thinking work” is delegated to tools like Claude Code. In that same ecosystem, a detailed blog on How Anthropic engineers are winning and losing at the Same Time argues that Anthropic’s groundbreaking internal research reveals a future in which developers risk over trusting AI generated code and losing the tacit knowledge that once came from debugging production outages at 3 a.m.
Those worries are sharpened by the company’s own ambitions for “agentic” coding. In a 2026 Agentic Coding, Anthropic staff describe how standard coding assistance in 24 and 25 was mostly passive, where you type and it suggests, or you give it a prompt and it returns a snippet. The next wave, they argue, will involve agents that can plan, modify, and deploy entire systems with minimal human intervention. Anthropic’s CEO has already suggested that by the end of 2026, AI models will be capable of writing software largely on their own, as captured in a short video that frames the shift as moving from humans writing code to humans specifying intent and letting machines execute at scale. In that context, Anthropic’s own paper on How AI assistance learning reads less like a narrow academic result and more like an early warning label for an industry sprinting toward automation.
More from Morning Overview