AI Coding Tradeoffs

Introduction

AI coding assistants are often presented as early proof that intelligence can become abundant: software tools that help ordinary people write code, speed up engineering work, and spread technical capability far beyond elite specialists. In some settings, that story is real. Developers using tools such as GitHub Copilot, Cursor, Claude, and ChatGPT can complete certain coding tasks much faster, especially when starting from scratch, learning unfamiliar frameworks, or automating repetitive work. arXiv [The GitHub Blog]github.blogThe GitHub BlogResearch: Quantifying GitHub Copilot's impact in…May 13, 2024 — We found that our AI pair programmer helps developers c…Published: May 13, 2024

But software engineering has also become one of the clearest demonstrations that AI assistance does not automatically translate into higher productivity. Recent studies of experienced developers working on mature codebases have found that AI tools can slow them down rather than speed them up. The reason is not simply that the models make mistakes. It is that expert work often depends on tacit knowledge, architectural judgement, maintenance discipline, and understanding years of accumulated decisions inside complex systems. In those environments, generating code is only a small part of the job.

That tension matters far beyond programming. Coding is unusually measurable: outputs can be tested, reviewed, benchmarked, and deployed. If even software engineering shows mixed results, it suggests a broader lesson for the larger question of AI abundance. Intelligence may become cheaper and more available, but the hardest parts of expert work may remain stubbornly human for longer than enthusiasts expect.

Where coding assistants have genuinely sped up work

The optimistic case for AI coding tools is not imaginary. Many developers report substantial gains in specific kinds of tasks, and some controlled studies have shown large improvements in completion speed for constrained programming exercises.

One widely cited 2023 GitHub study found developers using Copilot completed a programming task roughly 56% faster than a control group. [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023 Other enterprise studies reported shorter cycle times, faster onboarding, and improved developer satisfaction. [Harness.io]harness.iothe impact of github copilot on developer productivity a case studyThe Impact of Github Copilot on Developer Productivity12 Sept 2023 — The study demonstrated a reduction in cycle time by an average of 3… [The GitHub Blog]github.blogThe GitHub BlogResearch: Quantifying GitHub Copilot's impact in…May 13, 2024 — We found that our AI pair programmer helps developers c…Published: May 13, 2024

The strongest gains usually appear in a few recurring situations:

These are genuine capability multipliers. A junior developer who previously struggled to assemble a basic web application can now generate a workable draft in minutes. A small startup can prototype products faster with fewer engineers. Non-specialists can automate tasks that once required hiring a programmer.

That matters for the broader “intelligence abundance” argument. Software development has long been bottlenecked by scarce expertise. If AI lowers the skill threshold for useful programming, more people can participate in building digital systems. In principle, that could spread innovation more widely across economies and institutions.

Yet the same evidence also reveals an important limit: generating plausible code is easier than maintaining reliable systems over time.

Why mature codebases can reverse the gains

The clearest challenge to the simple “AI makes programmers faster” narrative came from a 2025 randomised controlled trial by the research organisation METR. The study examined experienced open-source developers working on repositories they already knew well. Instead of speeding them up, AI tools increased task completion time by around 19%. [metr.org]metr.org2025 07 10 early 2025 ai experienced os dev studyMeasuring the Impact of Early-2025 AI on Experienced…10 Jul 2025 — We conduct a randomized controlled trial (RCT) to understand how ea… [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023

The result surprised even the participants. Before the study, developers expected AI to reduce completion time substantially. Afterward, many still believed the tools had helped, despite the measured slowdown. [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023

This gap between perceived productivity and actual productivity is one of the most important findings in the entire AI coding debate.

Several mechanisms help explain the slowdown.

Experts already know the terrain

The developers in the METR study were not beginners facing unfamiliar code. They were long-term maintainers with years of accumulated knowledge about their repositories. In that context, AI often duplicated knowledge they already possessed while adding review overhead. [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023

An expert maintainer may already know:

Large language models usually do not possess this hidden context unless it is painstakingly reconstructed inside prompts or retrieved from documentation. Even then, the model may miss organisational norms or historical reasoning embedded in years of prior discussion.

As a result, experts often spend time checking, correcting, constraining, or rewriting AI output that superficially appears correct.

Review becomes the bottleneck

Programming is not mainly typing. In mature software systems, much of the work involves validation, coordination, debugging, and maintenance.

The METR study found developers spent more time reviewing and cleaning AI-generated code than expected. Only a minority of suggestions were accepted directly. [IBM]ibm.comreduce technical debtIBMReducing technical debt in 2026Although AI automation can make tech debt worse, AI-powered code assistance promises a powerful solutio…

This reflects a broader pattern increasingly discussed inside engineering teams:

That shift resembles what happened in some other knowledge industries. Machine translation, for example, reduced some forms of translation labour while increasing “post-editing” work: humans correcting fluent but unreliable drafts. Similar dynamics now appear in programming. [Time]time.comHow AI Is Changing White-Collar WorkFreelance translator Julian Pintat, like many in linguistically intensive professions, has seen his expertise undervalued as AI translati…

The danger is not merely bugs. It is subtle degradation.

Plausible code can hide long-term problems

A recurring criticism of AI-generated code is that it often looks convincing while quietly introducing technical debt: hidden complexity, duplicated logic, poor abstractions, or fragile architecture.

A 2026 large-scale empirical study analysing hundreds of thousands of AI-authored commits across thousands of repositories found that AI-generated code frequently introduced maintainability and quality problems, many of which persisted over time. [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023

Another study examining open-source development after GitHub Copilot adoption found that increased productivity among peripheral contributors coincided with greater review and maintenance burdens on core maintainers. Experienced developers reviewed substantially more code while their own original coding productivity fell. [arXiv]arxiv.orgarXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re…Published: February 13, 2023

This is a crucial distinction:

• AI may increase the volume of code.
• But more code is not necessarily better software.

In large systems, complexity compounds. Every additional shortcut, workaround, or unclear abstraction creates future costs. Experienced engineers are often paid not to maximise code output, but to minimise future fragility.

That is one reason senior developers may become more sceptical as projects mature.

The paradox of “feeling faster”

One striking result from several studies is that developers often report enjoying AI-assisted programming even when measured productivity improvements are weak or negative. [Let's Data Science]letsdatascience.comLet's Data ScienceAI Coding Tools Made Developers 19% Slower: METR StudyA METR randomized controlled trial found AI coding tools made exp…

AI tools reduce some unpleasant aspects of programming:

The experience can feel smoother and cognitively lighter.

That matters because subjective experience affects adoption. Developers may continue using tools that make work more pleasant even if objective gains are modest. This helps explain why AI coding adoption remains high despite falling trust in generated outputs. Stack Overflow’s developer reporting has shown growing use of AI tools alongside declining trust in their reliability. [stackoverflow.blog]stackoverflow.blogMind the gap: Closing the AI trust gap for developers18 Feb 2026 — But their trust in those tools dropped sharply: Only 29% of 2025 respo…

This tension reveals something broader about human-AI collaboration. People do not evaluate tools only by measurable productivity. They also care about effort, flow, frustration, confidence, and emotional fatigue.

The result may resemble earlier automation waves:

Instead, the nature of expert work shifted.

What this reveals about expert augmentation

The coding debate matters because software engineering is unusually favourable terrain for AI assistance.

Code is:

• digital,
• text-based,
• testable,
• and richly represented online.

If AI still struggles in high-context expert environments there, that suggests caution about claims that all knowledge work will rapidly become frictionless.

At the same time, the findings do not imply AI coding tools are useless. Instead, they point toward a more realistic model of augmentation.

AI helps most when context is thin and structure is standard

AI systems excel when:

• the task resembles common internet patterns,
• requirements are clear,
• architecture is simple,
• and verification is easy.

That makes them powerful for:

• tutorials,
• prototypes,

These are still economically important capabilities. They may broaden access to software creation dramatically.

Human expertise matters more as systems become valuable

The more critical a system becomes, the more software work depends on:

• institutional memory,
• reliability,
• security,
• interoperability,
• and long-term maintainability.

Those are not simply coding problems. They are governance and judgement problems.

A model can generate a function quickly. It is much harder for it to understand why a decade-old workaround exists, which legal constraints shaped a system, or which hidden dependencies would collapse after a “clean-up”.

This distinction may generalise beyond software engineering into medicine, law, infrastructure, science, and public administration. AI may automate parts of expert labour while simultaneously increasing demand for high-trust human oversight.

Coding tradeoffs and the larger AI bloom question

The mixed record of AI coding assistants complicates simplistic visions of an immediate “intelligence explosion”. Productivity gains are real, but they are uneven, context-dependent, and often offset by new forms of friction.

That does not destroy the larger AI bloom thesis. In fact, software engineering may still be an early example of cognitive abundance emerging in messy, partial form.

Several long-term possibilities remain plausible:

• AI may continue improving rapidly at repository-scale reasoning.
• Better retrieval systems may supply more organisational context.
• Verification tools may reduce review costs.
• Teams may redesign workflows around AI collaboration rather than bolting assistants onto old processes.
• Software architectures themselves may evolve to become more machine-readable and machine-maintainable.

But the evidence so far suggests that abundance is not the same thing as effortless automation.

The strongest interpretation of current coding evidence is not “AI replaces programmers” or “AI is hype”. It is that expert work contains layers that are invisible from the outside. Writing code is only one layer. Understanding systems, anticipating failures, preserving coherence, and maintaining reliability across years of change are often the scarcer skills.

That distinction matters for the wider future of AI and human flourishing. If advanced AI systems eventually make intelligence far cheaper and more widely available, the benefits may still depend heavily on human institutions, incentives, review structures, and stewardship. Cheap generation alone is not enough. Civilisation-scale flourishing depends on whether societies can convert abundant outputs into trustworthy, maintainable, and genuinely useful systems.

Endnotes

1. Source: arxiv.org
   Link: https://arxiv.org/abs/2302.06590
   Source snippet

   arXiv[2302.06590] The Impact of AI on Developer ProductivityFebruary 13, 2023 — by S Peng · 2023 · Cited by 1130 — This paper presents re...

   Published: February 13, 2023

2. Source: github.blog
   Link: https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-in-the-enterprise-with-accenture/
   Source snippet

   The GitHub BlogResearch: Quantifying GitHub Copilot's impact in...May 13, 2024 — We found that our AI pair programmer helps developers c...

   Published: May 13, 2024

3. Source: github.blog
   Title: The Git Hub Blog Does Git Hub Copilot improve code quality?
   Link: https://github.blog/news-insights/research/does-github-copilot-improve-code-quality-heres-what-the-data-says/
   Source snippet

   Here's what...Nov 18, 2024 — In that time, GitHub Copilot has helped developers code up to 55% faster. Prior research also showed that 8...

4. Source: harness.io
   Title: the impact of github copilot on developer productivity a case study
   Link: https://www.harness.io/blog/the-impact-of-github-copilot-on-developer-productivity-a-case-study
   Source snippet

   The Impact of Github Copilot on Developer Productivity12 Sept 2023 — The study demonstrated a reduction in cycle time by an average of 3...

5. Source: arxiv.org
   Link: https://arxiv.org/pdf/2509.20353
   Source snippet

   arXivDeveloper Productivity With and Without GitHub Copilotby V Stray · 2025 · Cited by 6 — The results showed a 26.08% increase in compl...

6. Source: metr.org
   Title: 2025 07 10 early 2025 ai experienced os dev study
   Link: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
   Source snippet

   Measuring the Impact of Early-2025 AI on Experienced...10 Jul 2025 — We conduct a randomized controlled trial (RCT) to understand how ea...

7. Source: arxiv.org
   Link: https://arxiv.org/abs/2507.09089
   Source snippet

   arXivMeasuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity...

8. Source: ibm.com
   Title: reduce technical debt
   Link: https://www.ibm.com/think/insights/reduce-technical-debt
   Source snippet

   IBMReducing technical debt in 2026Although AI automation can make tech debt worse, AI-powered code assistance promises a powerful solutio...

9. Source: time.com
   Title: How AI Is Changing White-Collar Work
   Link: https://time.com/7322933/ai-changing-white-collar-work/
   Source snippet

   Freelance translator Julian Pintat, like many in linguistically intensive professions, has seen his expertise undervalued as AI translati...

10. Source: arxiv.org
    Link: https://arxiv.org/html/2603.28592v2
    Source snippet

    arXivA Large-Scale Empirical Study of AI-Generated Code in the...Apr 26, 2026 — In this paper, we conduct a large-scale empirical study...

11. Source: arxiv.org
    Link: https://arxiv.org/abs/2603.28592

12. Source: arxiv.org
    Link: https://arxiv.org/abs/2510.10165
    Source snippet

    arXivAI-assisted Programming May Decrease the Productivity of Experienced Developers by Increasing Maintenance BurdenOctober 11, 2025...

    Published: October 11, 2025

13. Source: arxiv.org
    Link: https://arxiv.org/abs/2509.20353

14. Source: stackoverflow.blog
    Link: https://stackoverflow.blog/2026/02/18/closing-the-developer-ai-trust-gap/
    Source snippet

    Mind the gap: Closing the AI trust gap for developers18 Feb 2026 — But their trust in those tools dropped sharply: Only 29% of 2025 respo...

15. Source: github.com
    Title: Git Hub · Change is constant
    Link: https://github.com/
    Source snippet

    GitHub keeps you ahead. · GitHubFrom your first line of code to final deployment, GitHub provides AI and automation tools to help you bui...

16. Source: learn.github.com
    Link: https://learn.github.com/learning-pathways/github-copilot
    Source snippet

    Developers resolved programming tasks 55% faster with GitHub Copilot. As... 85% of developers felt more confident in code quality...

17. Source: arxiv.org
    Link: https://arxiv.org/pdf/2603.28592
    Source snippet

    A Large-Scale Empirical Study of AI-Generated Code in the...by Y Liu · 2026 · Cited by 2 — In this paper, we conduct a large-scale empir...

18. Source: arxiv.org
    Link: https://arxiv.org/pdf/2302.06590
    Source snippet

    We present a controlled trial of GitHub Copilot, an AI pair...Read more...

19. Source: github.blog
    Link: https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/
    Source snippet

    quantifying GitHub Copilot's impact on developer...7 Sept 2022 — In our research, we saw that GitHub Copilot supports faster completion...

20. Source: metr.org
    Title: 2026 02 24 uplift update
    Link: https://metr.org/blog/2026-02-24-uplift-update/
    Source snippet

    Our early 2025 study found the use of AI causes tasks to take 19% longer, with a confidence...Read more...

21. Source: stackoverflow.blog
    Title: developerweek 2026
    Link: https://stackoverflow.blog/2026/03/05/developerweek-2026/
    Source snippet

    Making AI tools that are actually goodMar 5, 2026 — AI coding tools without company context, for instance, will generate code that doesn'...

22. Source: letsdatascience.com
    Link: https://letsdatascience.com/blog/developers-thought-ai-made-them-faster-the-data-said-otherwise
    Source snippet

    Let's Data ScienceAI Coding Tools Made Developers 19% Slower: METR StudyA METR randomized controlled trial found AI coding tools made exp...

Additional References

1. Source: researchgate.net
   Link: https://www.researchgate.net/publication/381609417_The_impact_of_GitHub_Copilot_on_developer_productivity_from_a_software_engineering_body_of_knowledge_perspective
   Source snippet

   (PDF) The impact of GitHub Copilot on developer...14 Aug 2024 — A case study was conducted at a leading automotive organization investig...

2. Source: reddit.com
   Link: https://www.reddit.com/r/programming/comments/1it1usc/how_ai_generated_code_accelerates_technical_debt/
   Source snippet

   How AI generated code accelerates technical debtAI-generated code—especially through trends like “vibe coding”—can quickly create technic...

3. Source: linkedin.com
   Link: https://www.linkedin.com/posts/thestevemitchell_measuring-the-impact-of-early-2025-ai-on-activity-7349786687230644225-OqTH
   Source snippet

   METR study: AI assistants slow down experienced...A study from METR shows that experienced developers working with AI assistants produce...

4. Source: medium.com
   Link: https://medium.com/%40adnanmasood/rethinking-developer-productivity-in-the-age-of-ai-metrics-that-actually-matter-61834691c76e
   Source snippet

   Rethinking Developer Productivity in the Age of AI: Metrics...In day-to-day usage, GitHub reports that Copilot users feel more productiv...

5. Source: medium.com
   Link: https://medium.com/%40justhamade/true-cost-of-ai-generated-code-f4362391790c
   Source snippet

   True Cost of AI-Generated CodeFurthermore, it explores the underlying cognitive science, explaining how AI-assisted programming shifts th...

6. Source: altersquare.medium.com
   Link: https://altersquare.medium.com/ai-coding-tools-in-2026-what-we-actually-use-across-20-client-projects-and-what-we-dont-84b3306ce71a
   Source snippet

   Coding Tools in 2026: What We Actually Use Across 20+...AI coding tools aren't experimental anymore. They're embedded in production work...

7. Source: researchgate.net
   Link: https://www.researchgate.net/publication/368473822_The_Impact_of_AI_on_Developer_Productivity_Evidence_from_GitHub_Copilot
   Source snippet

   (2023) reported in a controlled study that GitHub Copilot increased programmer productivity by approximately 55% on a specific task...R...

8. Source: reuters.com
   Title: ai slows down some experienced software developers study finds 2025 07 10
   Link: https://www.reuters.com/business/ai-slows-down-some-experienced-software-developers-study-finds-2025-07-10/
   Source snippet

   AI slows down some experienced software developers...10 Jul 2025 — Using cutting-edge artificial intelligence tools slowed down experien...

9. Source: youtube.com
   Link: https://www.youtube.com/watch?v=dZ11xA4sUFQ
   Source snippet

   Developer productivity on GitHub Copilot (w/ Eirini...How to use GitHub Copilot (the complete beginner's guide) · Funnel analytic...

10. Source: reddit.com
    Link: https://www.reddit.com/r/programming/comments/1m25iw2/metr_study_finds_ai_doesnt_make_devs_as/
    Source snippet

    • 10mo ago. AI slows down some experienced software developers, study finds. 746. 231. We asked 15,000 European devs about jobs...Read more...

Amazon book picks

Further Reading

Books and field guides related to AI Coding Tradeoffs. Use these as the next step if you want deeper reading beyond the article.

Book

Paradigms of Artificial Intelligence Programming

By Peter Norvig

Paradigms of AI Programming is the first text to teach advanced Common Lisp techniques in the context of building major AI systems. By re...

See on Amazon

Book

Programming Collective Intelligence

By Toby Segaran

First published 2007. Subjects: Social aspects, Internet programming, Information technology, Social aspects of Information technology, W...

See on Amazon

Book

AI Code Revolution with Vibe Coding

By Rajender Kumar

Build Software Through Conversation – The AI Coding Revolution Starts Here Transform from idea to working application in hours, not month...

See on Amazon

Book

Improving Software Development Productivity

By Randall W. Jensen

First published 2014. Subjects: Computer software, development, management, Computer software -- Development, Computer software, Developm...

See on Amazon

Browse more on Amazon

As an Amazon Associate I earn from qualifying purchases.