Usage is not the bottleneck - review is.
What it is about
AI has arrived in everyday development. According to the Stack Overflow Developer Survey 2025 (over 49,000 respondents), 84% use AI tools or plan to. Only 31% use agents, and 46% distrust the accuracy of the results.
What the studies show
- “Almost right” is the main problem: 66% name solutions that are almost right, but not quite, as their biggest frustration. 44.1% do not plan to use AI for testing, 58.7% not for committing and reviewing code.
- Issues stay in the code: Across 302,600 AI commits in 6,299 repositories, 22.7% of the introduced issues still survive at the latest version (Liu et al., 2026, preprint). Most are code smells; there is no comparison with purely human code.
- Filters make it usable: At Meta, 73% of AI test improvements were accepted after human review. Before filtering, 75% of the tests built, 57% passed reliably and 25% increased coverage (Meta, 2024).
The surveys are self-reported, and the Meta studies concern an internal Meta tool. Together they still show the same pattern: the value appears only after review.
What we make of it
AI output in our projects first passes automated checks - build, tests, static analysis - and then human review. Only what survives the filters costs review time. How this works is described in From the artefact to the tool and Agents that write tests.
The studies in detail
Selection: only studies with a disclosed method - experiments, official statistics, legal texts, independent institutes. Vendor surveys about their own product and consultancy forecasts are not included here.
- 84%
Stack Overflow · 2025 · Surveysurvey, self-reported
84% of respondents in the Stack Overflow Developer Survey 2025 use or plan to use AI tools in development (2024: 76%).
Base: Over 49,000 responses from 177 countries; AI-usage question about 33,700 respondents
Limits: Self-reported; self-selected online sample
Stack Overflow Developer Survey 2025 - AI ↗We use it in: Agentic development in everyday enterprise work · From the artefact to the tool: controlled autonomy in software development · Agentic AI: 85% on the slide, 36% at best
- 31%
Stack Overflow · 2025 · Surveysurvey, self-reported
According to Stack Overflow 2025, only 31% of developers currently use AI agents; 17% plan to and 38% do not plan to.
Base: Over 49,000 responses from 177 countries
Limits: Self-reported; mid-2025 snapshot
Stack Overflow's 2025 Developer Survey Reveals Trust in AI at an All Time Low ↗We use it in: Agentic development in everyday enterprise work · From the artefact to the tool: controlled autonomy in software development
- 46%
Stack Overflow · 2025 · Surveysurvey, self-reported
In the Stack Overflow 2025 survey, 46% of developers distrust the accuracy of AI tools and only 33% trust it.
Base: Over 49,000 respondents, 177 countries (sentiment question about 33,400)
Limits: Self-reported
Stack Overflow Developer Survey 2025 - AI ↗We use it in: Agentic development in everyday enterprise work · Agents that write tests - how we approach it · From the artefact to the tool: controlled autonomy in software development · Agentic AI: 85% on the slide, 36% at best
- 66%
Stack Overflow · 2025 · Surveysurvey, self-reported
66% of developers name AI solutions that are almost right, but not quite, as their top frustration; 45% say debugging AI-generated code takes more time.
Base: Stack Overflow Developer Survey 2025, over 49,000 respondents
Limits: Quote is a chart label with its value, not running text; self-reported
Stack Overflow Developer Survey 2025 - AI ↗We use it in: Agents that write tests - how we approach it · Agentic AI: 85% on the slide, 36% at best
- 44,1%
Stack Overflow · 2025 · Surveysurvey, self-reported
44.1% of developers do not plan to use AI for testing code; for committing and reviewing code the figure is 58.7%.
Base: Stack Overflow Developer Survey 2025, all respondents to the workflow question
Limits: Chart values, not running text; self-reported
Stack Overflow Developer Survey 2025 - AI ↗We use it in: Agents that write tests - how we approach it
- 22,7%
arXiv (Liu et al.) · 2026 · Studyindependent
22.7% of the issues AI coding assistants introduce into public GitHub projects still survive at the latest version (302,600 AI commits, 6,299 repositories).
Base: 302,600 verified AI commits in 6,299 GitHub repositories (Copilot, Claude, Cursor, Gemini, Devin)
Limits: Preprint; no human-only baseline; static-analysis findings incl. false positives, mostly code smells
Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild ↗We use it in: Agentic AI: 85% on the slide, 36% at best · Agents that write tests - how we approach it · From the artefact to the tool: controlled autonomy in software development
- 73%
arXiv (Meta) · 2024 · Studypublisher's own internal tool
At Meta, engineers accepted 73% of TestGen-LLM's test improvements for production after human review; 11.5% of classes it was applied to were improved.
Base: Instagram and Facebook test-a-thons
Limits: Single company; improves existing human-written tests
Automated Unit Test Improvement using Large Language Models at Meta ↗We use it in: Agents that write tests - how we approach it
- 25%
arXiv (Meta) · 2024 · Studypublisher's own internal tool
Of LLM-generated test cases at Meta, 75% built correctly, 57% passed reliably and 25% increased coverage, which is why automated filters are needed.
Base: Instagram Reels and Stories evaluation
Limits: Single company; specific product surfaces
Automated Unit Test Improvement using Large Language Models at Meta ↗We use it in: Agents that write tests - how we approach it
- 73%
arXiv (Meta) · 2025 · Studypublisher's own internal tool
Meta's mutation-guided LLM test generator ran on 10,795 Kotlin classes; engineers accepted 73% of its tests and judged 36% privacy-relevant.
Base: 10,795 Android Kotlin classes in 7 Meta platforms; 571 test cases
Limits: Single company; compliance and privacy focus; preprint
Mutation-Guided LLM-based Test Generation at Meta ↗We use it in: Agents that write tests - how we approach it
