iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI coding assistants are now widely used among professional developers, and several studies find measurable speed-ups on defined tasks. But the evidence does not show that every developer becomes faster, and it does not establish that AI has already caused broad job losses in software. What is changing most clearly is the mix of work: more time spent understanding, checking and steering generated code, and more uncertainty about which skills will matter most. The studies answer different questions, so this article separates four of them: task-level productivity, reported experience, organizational effects and employment.
Is AI going to replace software developers?
Not on the evidence available so far. The strongest labor-market signals are narrower than the headlines: a slowdown in how fast coder employment is growing in one U.S. study, and a warning about younger workers in a cross-sector review. Neither shows that AI has already eliminated a measurable number of developer jobs.
What the Federal Reserve paper found
The Board of Governors of the Federal Reserve System’s March 2026 working paper, “AI and Coder Employment: Compiling the Evidence,” links occupational data to labor-market data. It finds that coder employment kept growing, but much more slowly than before 2022. The authors tie the change to an occupation-specific shift around the arrival of ChatGPT, yet they report that their industry-level controls do not explain the slowdown. The paper is preliminary and was circulated to invite discussion. It documents a pattern that coincides with ChatGPT’s arrival, not a settled causal count of developer jobs lost to AI.
What the ILO review found
The International Labour Organization’s June 1, 2026 review, “The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence,” draws on experiments, firm data, platform studies and worker surveys from several countries. It concludes that large-scale displacement remains limited in the evidence it reviewed. It identifies risks, including fewer opportunities for younger workers and changes to work organization and job quality. The review covers all sectors, so it is context for developers rather than a developer-specific forecast.
#1 Best Overall
Reading the employment signals carefully
- Slower growth is not the same as job loss. The Federal Reserve paper describes employment that kept rising, just more slowly.
- Cross-sector findings do not transfer automatically to software, because the ILO review covers the whole economy.
- A risk to new entrants is a warning about how opportunities are distributed, not evidence that current developers are being displaced.
Vendors make their own case. In GitHub’s 2024 survey write-up, COO Kyle Daigle wrote: “AI doesn’t replace human jobs—it frees up time for human creativity.” That is an executive’s view rather than a finding, and it does not settle the employment question in either direction.
Does AI actually make developers faster?
Sometimes, on some tasks, in some settings. The studies below measure different things: time on a single task, completed work inside companies, and self-reported use. Their results should not be averaged or treated as interchangeable.
Rank #2
A controlled task: building an HTTP server
In a February 2023 Microsoft Research experiment, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” recruited developers implemented a JavaScript HTTP server as quickly as they could. The group with access to GitHub Copilot finished 55.8% faster than the control group. That is a precise result for one defined task and one set of participants. It does not mean developers are 55.8% more productive. The study did not measure debugging, design decisions or maintenance of a large codebase.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Field experiments across three companies
A later Microsoft Research study, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers,” appeared on Microsoft Research in June 2025 and was published online in Management Science on February 27, 2026. The authors (Cui, Demirer, Jaffe, Musolff, Peng and Salz) randomized access to an AI coding assistant in ordinary work at Microsoft, Accenture and an anonymous Fortune 100 company. Combining the three experiments, which cover 4,867 developers, they estimate a 26.08% increase in completed tasks, with a standard error of 10.3%.
Rank #3
Three qualifications belong with that figure. The individual experiments were noisy, and the authors report that outcomes varied across them. The gains were not uniform: less experienced developers had higher adoption rates and greater productivity gains. And the standard error is large relative to the estimate. If the 10.3% is read as percentage points and a conventional 95% range is applied, the figures are compatible with gains from roughly 6% to 46%. That range is a rough calculation from the reported numbers, not a range the authors state. The estimate is distinguishable from zero in these settings, but its size is uncertain, and it does not give a precise figure for any particular team.
Side by side
The table compares what each source actually measured, which determines what claim it can support.
| Source | Task or setting | What was measured | Main limit |
|---|---|---|---|
| Microsoft Research, February 2023 | Recruited developers; one JavaScript HTTP server task | Time to complete the task | One task and one participant group |
| Microsoft Research (June 2025); Management Science (online February 27, 2026) | Randomized access in ordinary work at Microsoft, Accenture and an anonymous Fortune 100 company; three experiments | Completed tasks | Individual experiments were noisy and results varied across them |
| GitHub with Wakefield Research, survey published August 20, 2024 (updated April 15, 2025) | 2,000 non-student, non-manager respondents at companies with 1,000 or more employees; 500 each in the U.S., Brazil, Germany and India | Reported use and perceptions | Self-reported; frequency of use was not asked; not a code-quality audit |
| Google DORA, 2025 | More than 100 hours of qualitative data and survey responses from nearly 5,000 technology professionals | Organizational patterns and conditions | Describes amplification effects; not a single productivity figure |
What developers report in everyday use
Survey respondents describe where the tools help most. Across the four surveyed markets, 60–71% said AI tools made it easy to adopt a new programming language or understand an existing codebase, and more than 98% said their organizations had experimented with AI tools for test generation. These are perceptions, not measured outcomes. They describe experience with the tools, not verified output.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsWhat does AI change about writing code?
The clearest changes concern how developers spend their attention, meaning reading, checking and explaining code, more than the act of typing it. Two patterns recur across the sources.
Best Value
Understanding and debugging existing code
Anthropic’s December 2, 2025 study, “How AI is transforming work at Anthropic,” surveyed 132 of its engineers and researchers, interviewed 53 people and examined Claude Code usage. Respondents described using Claude for debugging and for understanding code, among other tasks, and reported changes to their productivity and to the breadth of their work. The authors are explicit about the limit: Anthropic’s staff had early access to advanced tools and work in a relatively stable field, so these findings do not represent all developers. Read it as an account of one group with unusual access, not a census of the profession.
Oversight is part of the workflow
DORA’s 2025 report, based on more than 100 hours of qualitative data and survey responses from nearly 5,000 technology professionals, frames AI as an amplifier: it can magnify strong organizational practices and existing dysfunctions alike. The tool’s effect therefore depends on the system around it. GitHub’s survey notes that generated code and tests still require human review, and Anthropic’s respondents raised concerns about supervising model output. The following checks follow from those concerns. They are practical suggestions, not procedures tested in the studies.
- Confirm that generated tests assert the required behavior, not just that they pass.
- Read generated code for error handling, security-sensitive logic and newly added dependencies before merging.
- Plan review capacity explicitly, since faster drafting moves work to whoever checks it.
- Track rework after merge, not only the time to a first draft.
Will learning to code still matter?
The evidence does not support a simple yes or no. It points to a narrower question: which skills gain value when a tool can draft code quickly? Three concerns come up in the sources, and none is a proven outcome.
- Skill maintenance. Anthropic’s respondents worried about keeping their technical expertise sharp. Whether that concern becomes lasting skill loss over years is not measured in any of the studies reviewed here.
- Entry-level paths. The ILO review names reduced opportunities for younger workers as a risk. The studies do not yet show how that plays out for people learning to program.
- Collaboration and job security. The Anthropic study also records concern about collaboration and job security. These are reported concerns, not measured effects.
This is an inference from the evidence above rather than a finding of any single study: as drafting gets cheaper, reading unfamiliar code, judging whether output is correct, writing tests and describing a problem precisely carry more of the weight.
How to evaluate AI claims for your own team
If you are deciding whether an AI coding tool is worth adopting, the studies suggest a sequence of questions. Each one filters out a common overreach.
Quick Recap
- Name the outcome. Time on a task, completed work, self-reported use and employment are different measures. A gain on one does not imply a gain on another.
- Match the setting. A timed exercise with recruited participants says little about a mature codebase with legacy constraints.
- Split results by experience. The gains reported above were not uniform, so a team average can hide different results for junior and senior developers.
- Measure after review. Include testing, review and rework in the measurement window, because faster drafting can shift work downstream.
- Set a longer horizon. Results from the first few weeks do not show maintenance costs or career effects.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
