---
title: Anthropic Study Finds AI Coding Assistants Speed Up Work but Reduce Skills
description: Anthropic research finds AI coding tools cut skill mastery by 17% and junior developer hiring fell 20%. The hidden cost of AI-assisted programming.
author: Darie Nani (Editor-in-Chief)
date: 2026-01-30T14:59:23.000Z
updated: 2026-05-05T19:29:40.265Z
canonical: https://www.sovereignmagazine.com/article/anthropic-study-finds-ai-coding-assistants-speed-up-work-but-reduce-skills
image: https://cdn.nanimediahouse.com/270373.jpeg
categories: Education
content_type: Analysis
region: United States
publication: Sovereign Magazine
about:
  - type: Organization
    name: Anthropic
---

Employment for software developers aged 22 to 25 has fallen nearly 20 percent since late 2022, according to [Stanford University payroll data](https://digitaleconomy.stanford.edu/wp-content/uploads/2025/08/Canaries_BrynjolfssonChandarChen.pdf) tracking millions of workers. The decline coincides precisely with the mainstream adoption of [AI coding tools](https://www.sovereignmagazine.com/article/vibe-coding-killing-saas-tools). Meanwhile, employment for developers over 26 has held steady or grown.

The job market shift reflects a calculation playing out across the technology sector. [Companies that once hired ten developers](https://www.sovereignmagazine.com/article/laying-off-humans-but-pouring-billions-of-dollars-into-ai-s-future) now believe they need two experienced engineers and an AI assistant. Where junior developers used to spend months learning through debugging and writing repetitive code, AI tools now handle those tasks in seconds. But [new research from Anthropic](https://www.anthropic.com/research/AI-assistance-coding-skills) suggests this productivity gain comes with a hidden cost that could reshape software engineering for a decade.

Anthropic researchers Judy Hanwen Shen and Alex Tamkin designed a randomised controlled trial to measure what happens when software developers learn new skills with AI assistance. They recruited 52 engineers, mostly junior, who regularly used Python but were unfamiliar with Trio, a library for asynchronous programming. Each participant worked through coding tasks with either AI assistance or traditional methods, then took a quiz on the concepts they had just used.

Developers who used AI scored 17 percent lower on the quiz than those who coded manually, a gap equivalent to nearly two letter grades. On debugging questions specifically, the difference was even more pronounced. The AI group completed tasks roughly two minutes faster on average, but the time saving did not reach statistical significance.

The study reveals a tension at the heart of [AI adoption in software development](https://www.sovereignmagazine.com/article/cielara-code-ai-coding-localization). Tools designed to boost productivity may be preventing developers from building the expertise needed to validate AI-generated code. Companies racing to automate coding work are creating a skills gap that will only become visible when something goes wrong. The pattern [mirrors concerns emerging in other professional industries](https://www.sovereignmagazine.com/article/ai-disrupts-newsrooms-as-journalists-voice-deepening-concerns) where AI adoption is accelerating.

## The Way Developers Use AI Determines What They Learn

Anthropic researchers found that not all AI assistance produces the same learning outcomes. They identified seven distinct interaction patterns among participants, with quiz scores ranging from below 40 percent to above 65 percent depending on how developers engaged with the AI.

Low-scoring patterns shared a common trait: heavy cognitive offloading. Some participants delegated all code writing to AI from the start. Others began with independent work but progressively relied on AI for debugging and verification rather than developing their own understanding. These developers scored poorly because they never grappled with the underlying concepts.

High-scoring participants used AI differently. Some generated code first, then asked follow-up questions to understand what they had created. Others crafted hybrid queries requesting both code and explanations. A third group used AI only for conceptual questions, resolving errors independently. These approaches took more time but produced better comprehension.

The distinction matters because software development is moving towards a model where humans provide oversight of AI-generated code. Developers who cannot debug or understand code will struggle to catch errors, assess design patterns, or determine when AI suggestions are inappropriate. The study suggests current tools may be training a generation of engineers who can produce code but cannot evaluate it.

## Experienced Developers Face Different Challenges

The Anthropic study focused on developers learning new skills, but [research from METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) examined how AI affects experienced engineers working on familiar codebases. The results complicate the productivity narrative.

METR recruited 16 experienced developers from large open-source repositories, each with an average of five years contributing to projects with over one million lines of code. Developers worked on 246 real issues from their repositories, randomly assigned to allow or prohibit AI assistance.

Experienced developers using AI took 19 percent longer to complete tasks than those working without it. The time cost came from reviewing AI-generated code, catching errors, and managing context switches. Participants estimated they were 20 percent faster with AI, suggesting developers themselves may not recognise the overhead AI introduces.

The finding does not mean AI is useless for experienced developers. The study examined complex, context-heavy work on familiar codebases where developers already possessed relevant skills. Different tasks and less experienced developers may see genuine productivity gains. But the research indicates AI acceleration is not uniform across all software development work.

## The Talent Pipeline Problem

Software companies are creating a structural problem that will take years to manifest. By eliminating entry-level positions and expecting AI to fill the gap, they are severing the pipeline that produces senior engineers.

Computer science graduates now face unemployment rates of 6.1 percent, higher than liberal arts majors. Stanford bioengineering professor Jan Liphardt [told the Los Angeles Times](https://www.latimes.com/business/technology/story/2025-12-19/stanford-computer-science-graduates-struggle-to-find-jobs-in-the-age-of-ai) that computer science graduates are struggling to find entry-level jobs with prominent tech companies. One recent graduate described a “very dreary mood on campus” as students compete for positions that have contracted by 40 percent compared to pre-2022 levels.

AWS chief executive Matt Garman articulated the long-term risk in remarks to staff. If companies have no talent pipeline and no junior people to mentor, they lose the source of future senior engineers and fresh ideas. The comparison to sports teams is apt. A roster of only veteran players cannot sustain itself. When those veterans retire, the organisation faces a leadership vacuum with no one trained to step up.

Industry observers call this “slow decay” – an ecosystem that stops training its replacements. The [Bureau of Labor Statistics projects](https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm) 15 percent growth in software jobs from 2024 to 2034, but that growth assumes companies hire junior developers who eventually become senior ones. Current trends suggest companies may find themselves competing for a shrinking pool of [experienced engineers](https://www.sovereignmagazine.com/article/cowork-is-a-big-disappointment-285-billion-wiped-because-of-this) while AI handles work that used to train new talent.

## What Companies Should Consider

Anthropic’s research suggests companies need deliberate policies around AI tool deployment, particularly for junior developers. The fastest interaction pattern in the study was complete AI delegation, where developers let AI write all code. This approach sacrificed learning entirely. [Managers under pressure to deliver quickly](https://www.sovereignmagazine.com/article/ai-influence-in-the-workplace-job-stratification-and-economic-impact) may inadvertently encourage this behaviour.

The study identified specific practices associated with better learning outcomes. Developers who asked conceptual questions, requested explanations alongside code generation, or used AI to verify their understanding rather than replace it showed stronger mastery. Companies could structure AI tool usage to encourage these approaches, particularly for developers building foundational skills.

Some organisations are already adapting. Anthropic notes that major language model services now offer learning modes specifically designed to foster understanding rather than just productivity. Claude Code includes learning and explanatory modes, while OpenAI offers ChatGPT Study Mode. These features recognise that different use cases require different AI behaviours.

Accenture and Anthropic [announced a partnership](https://newsroom.accenture.com/news/2025/accenture-and-anthropic-launch-multi-year-partnership-to-drive-enterprise-ai-innovation-and-value-across-industries) in December 2025 to help enterprises scale AI-powered software development with proper measurement and training frameworks. The collaboration trains approximately 30,000 Accenture professionals on Claude, creating what the companies describe as one of the largest ecosystems of Claude practitioners globally. The offering addresses [broader challenges businesses face](https://www.sovereignmagazine.com/article/when-machines-shop-preparing-your-business-for-ai-powered-buyers) as AI systems become more autonomous. The offering includes frameworks to quantify productivity gains, workflow redesign for AI-first development teams, and change management that keeps pace as AI evolves.

The structured approach reflects recognition that simply providing AI tools does not guarantee positive outcomes. Companies need systems to ensure efficiency gains do not come at the expense of skill development, particularly for early-career engineers who will provide oversight of AI systems in the future.

## The Skills That Matter Now

Software development is bifurcating into two tiers with vastly different prospects. Developers with AI expertise and machine learning capabilities command salaries reaching 235,000 dollars for senior roles, with elite positions paying up to 900,000 dollars. Entry-level developers face 50 percent fewer job openings than pre-pandemic levels and intense competition for remaining positions.

The market increasingly values developers who can work effectively with AI tools while maintaining strong fundamentals. Debugging skills matter more when AI generates code that may contain subtle errors. Architectural thinking becomes crucial when AI handles implementation details. Communication and problem decomposition grow more important as development work shifts towards guiding AI rather than writing every line manually.

Luciano Nooijen, an engineer at video game infrastructure developer Companion Group, [found himself struggling](https://www.technologyreview.com/2025/12/15/1128352/rise-of-ai-coding-developers-2026/) with basic tasks when working on a side project without access to the AI tools he used daily. Tasks that once felt instinctive became manual and cumbersome. He compared the experience to athletes who still perform basic drills to maintain muscle memory. The only way to preserve coding instinct, he concluded, is regular practice of fundamental work that AI might otherwise handle.

The observation points to a potential paradox. As AI handles more routine coding, developers may need more deliberate practice of fundamentals to maintain competence. Companies optimising for short-term productivity may be creating long-term capability gaps.

## Research Limitations and Open Questions

Anthropic researchers acknowledge their study provides only a snapshot of early 2026 AI capabilities in one setting. The sample of 52 developers was relatively small, and the assessment measured comprehension shortly after the coding task. Whether immediate quiz performance predicts long-term skill development remains unknown.

The study examined learning a new library through self-guided practice, which differs from how developers might use AI for other tasks. Effects could vary for different types of work, more experienced developers, or longer-term usage patterns. The research does not address whether AI assistance differs from human assistance while learning, or how effects might change as developers gain fluency with AI tools.

Previous research has found mixed results on AI coding productivity. Some studies show significant gains while others find minimal impact or slowdowns. Anthropic’s own observational research found AI can reduce task completion time by up to 80 percent, which appears to contradict the current findings. The difference lies in what was measured: earlier work examined productivity on tasks where participants already had relevant skills, while this study focused on learning something new.

The research suggests AI may both accelerate productivity on well-developed skills and hinder acquisition of new ones, though more work is needed to understand this relationship. As AI capabilities evolve rapidly, the interaction between AI assistance and skill development will likely shift. The findings provide a baseline for tracking how these dynamics change.

## Further context

**Q: Do AI coding assistants actually make developers more productive?**
The research shows mixed results. While developers feel about 20 percent faster when using AI assistants, studies show experienced developers actually took 19 percent longer to complete complex tasks. The productivity gains depend heavily on the type of work and the developer’s experience level. For routine tasks with clear requirements, AI can help. For complex, context-heavy work on familiar codebases, the overhead of reviewing AI-generated code can slow developers down.

**Q: How does using AI coding assistants affect learning programming skills?**
Research from Anthropic found that developers who used AI assistance whilst learning new programming concepts scored 17 percent lower on comprehension tests compared to those who coded manually. This is equivalent to nearly two letter grades. The key issue is cognitive offloading: when developers delegate code writing to AI without engaging with the underlying concepts, they fail to build the deep understanding needed to debug errors and evaluate code quality.

**Q: Will AI replace junior developers?**
AI is not replacing junior developers outright, but it is fundamentally changing what entry-level positions look like. Employment for software developers aged 22 to 25 has fallen nearly 20 percent since late 2022, whilst positions for developers over 26 have remained stable. Companies now expect fewer developers who can work effectively with AI tools from day one. The traditional learning path of writing repetitive code and fixing bugs is disappearing, creating a talent pipeline problem for the industry.

**Q: What is the best way to use AI coding assistants without losing programming skills?**
The research identified several effective approaches. High-performing developers used AI to generate initial code but then asked follow-up questions to understand what was created. Others crafted queries that requested both code and explanations. Some used AI only for conceptual questions whilst resolving errors independently. The key is to use AI as a learning aid rather than a replacement for thinking. Regular practice of fundamental skills without AI assistance also helps maintain coding instinct.

**Q: How is AI changing the skills companies look for in developers?**
The market is bifurcating into two tiers. Developers with AI expertise and machine learning capabilities command salaries reaching 235,000 dollars for senior roles. Meanwhile, entry-level developers face 50 percent fewer job openings than pre-pandemic levels. Companies increasingly value developers who can work effectively with AI tools whilst maintaining strong fundamentals in debugging, architectural thinking, and problem decomposition. The ability to provide oversight of AI-generated code has become essential.

**Q: What should software engineering students do to prepare for an AI-driven job market?**
Focus on building strong fundamentals that AI cannot easily replicate: understanding system architecture, debugging complex issues, and making design trade-offs. Learn to use AI tools effectively as part of your workflow, but ensure you understand the code they generate. Seek projects that require you to evaluate and improve AI-generated code rather than just accepting it. Build expertise in areas where human judgment remains crucial, such as determining when AI suggestions are inappropriate or recognising subtle security vulnerabilities.