AI 编程工具真实增效数据：研究综述与选型建议

Ask two engineering leaders whether AI coding assistants make their teams faster, and you will often get two opposite answers. Both can point to credible research to back their position. That contradiction is not a sign that the studies are wrong - it is the single most important thing to understand before you spend another euro on AI tooling.

Below is what the most cited independent and vendor studies actually found, attributed source by source. None of these numbers are DevPrism's, and none of them are yours. That is precisely the point.

The optimistic case: controlled experiments show real speed-ups

In a randomised controlled trial run by GitHub and Microsoft Research, 95 developers were asked to write an HTTP server in JavaScript. The group with GitHub Copilot finished 55% faster (1 hour 11 minutes versus 2 hours 41 minutes), and 78% completed the task versus 70% without the tool (statistically significant, p = .0017). The same research, using the SPACE framework, reported that 73% of developers said they stayed in flow and 87% preserved mental energy on repetitive tasks (GitHub, 2022).

This is a real, well-designed result. It is also a narrow one: a greenfield task, a single language, no legacy codebase, no review cycle, no production incident three weeks later. It tells you AI can accelerate a well-scoped task. It does not tell you what happens to your delivery system as a whole.

The cautionary case: experienced developers got slower

In 2025, the non-profit research group METR ran a randomised controlled trial with experienced open-source developers working on their own large, mature repositories. The result was the opposite of the headline narrative: developers were 19% slower when allowed to use AI tools.

The most striking finding was not the slowdown itself - it was the perception gap. Developers expected AI to make them 24% faster, and even after the experiment they still believed it had made them roughly 20% faster. They were wrong about the direction of the effect, on their own code (METR, 2025).

The lesson is uncomfortable and unavoidable: how fast you feel is not how fast you are. Self-reported productivity, and even seasoned intuition, can point the wrong way. Only measurement settles it.

The systemic view: AI is an amplifier, not a guarantee

Google's DORA programme - a decade of research on software delivery performance - reached a conclusion that reconciles the two camps above. In its 2024 report, DORA found that AI adoption increased individual productivity, flow and satisfaction, yet was associated with a decline in delivery throughput and stability at the team level (DORA, 2024).

In 2025, DORA sharpened the framing: AI is an amplifier. It magnifies the strengths of teams that already have strong practices, and magnifies the dysfunction of teams that do not. The largest returns came from the surrounding organisational system - clear processes, fast feedback loops, healthy codebases - not from the tool itself (DORA, 2025).

In other words, there is no universal "+55%" or "−19%." There is only what AI does to your system, given your current practices. Two organisations deploying the identical tool will get different outcomes - and both should expect to measure them, not assume them.

The hidden cost: quality and the bill that follows

Speed is only half of the equation. An independent analysis by GitClear of 211 million changed lines of code across large public repositories found worrying quality trends as AI assistance spread: code churn (lines reverted or rewritten within two weeks) is on track to roughly double versus the pre-AI 2021 baseline; the share of refactored lines fell from ~25% to under 10% between 2021 and 2024; and copy-pasted code rose while duplicate code blocks increased several-fold (GitClear, 2025).

This matters beyond aesthetics, because quality now has a direct, measurable cost. A study by Sonar found that a cleaner codebase measurably lowers what AI agents cost to run - fewer tokens for the same task and far fewer file revisits, with no drop in task completion (Sonar, 2026). Messy code does not just slow humans down; it makes your AI spend go up.

(Both GitClear and Sonar have a commercial interest in code quality, so treat the magnitudes as directional. The direction, however, is consistent with DORA's independent findings.)

The financial reality: the J-curve

So if the speed gains are real but conditional, and the quality risks are real too, what does that mean for ROI? DORA's ROI of AI-assisted Software Development report is blunt: initial gains in coding speed do not automatically translate to the bottom line.

The report describes a J-curve: teams should expect a productivity dip - a "tuition cost" - while they learn to work with AI, before durable returns appear. Crucially, it argues the largest returns come not from cutting headcount but from reclaiming engineering capacity by reducing unnecessary rework, and that leaders need to translate technical metrics into financial outcomes to defend the budget (DORA / Google Cloud, 2025).

Reducing rework. Connecting velocity to cost. Proving impact with data rather than anecdote. That is an engineering-measurement problem, not a licensing decision.

The only conclusion that holds: measure your own data

Line the studies up and the pattern is clear:

No two of these numbers will be true for your organisation simultaneously, and you cannot pick the flattering one and assume it applies. The only figure that describes your teams is the one you compute from your data.

That is the entire premise of DevPrism. We correlate three axes that every study above treats in isolation:

• Velocity - DORA delivery metrics, per team and per period.
• Quality - code smells, churn, coverage, duplication, from your SonarQube/Codacy signal.
• Cost - AI adoption and spend, per developer, across every assistant you use.

Put together, that lets you answer the question the research cannot answer for you: not "does AI make developers faster in general?" but "is it making my teams faster, without quietly raising my defect rate or my token bill - and is the ROI showing up yet, or are we still on the dip?"

The studies give you the right questions. Your data gives you the answer.

Stop guessing whether AI is paying off. Try DevPrism for free - correlate velocity, quality and AI cost across your teams from the Starter plan.