Skip to content
Codrlabs Open

As a working developer

Most of the risk in AI-assisted development is not that the tools are bad. It is that they are almost right, confidently — and that the feeling of speed arrives whether or not the speed does.

You are probably not as fast as you feel

Section titled “You are probably not as fast as you feel”

In a 2025 randomised controlled trial, METR had 16 experienced open-source developers work through 246 real issues on repositories they had contributed to for years, with AI randomly allowed or disallowed per task. With AI allowed, they took 19% longer. Before starting, they expected AI to make them 24% faster; after finishing, they still believed it had made them 20% faster.

The authors are careful: this is one setting — experienced developers, large familiar codebases, early-2025 tools — and it does not show that AI slows most developers down. What it does show is that the sense of speed is not evidence of speed.

  • Judge AI on outcomes you can measure — time to a merged, correct change — not on how the session felt.
  • Notice when prompting, waiting and repairing “almost right” output costs more than writing it yourself.

Treat generated code as untrusted input

Section titled “Treat generated code as untrusted input”

Perry et al. (CCS 2023) found that participants with an AI assistant wrote significantly less secure code than those without, and were more likely to believe it was secure. The participants who trusted the assistant less and worked harder on their prompts produced fewer vulnerabilities.

  • Review AI output with the suspicion you would give a stranger’s pull request — more where it touches authentication, input handling, file paths, queries or anything else security-relevant.
  • Confidence is not a signal. The finding is that AI raises confidence and lowers security at the same time.

Check that every dependency is real

Section titled “Check that every dependency is real”

Code-generating models regularly suggest packages that do not exist. Across 576,000 generated samples from 16 models, Spracklen et al. (USENIX Security 2025) found that at least 5.2% of suggested packages were hallucinated for commercial models, and 21.7% for open-source models — 205,474 unique fake package names. An attacker who registers one of those names gets their code installed by anyone who trusts the suggestion; the authors call it a new form of package confusion attack.

  • Before installing anything an assistant suggests, confirm the package exists, is the one you meant, and is maintained.
  • Never let an agent add dependencies without review.

Keep the codebase maintainable

Section titled “Keep the codebase maintainable”

GitClear analysed 211 million changed lines from 2020 to 2024. The share of lines classified as copy/pasted rose from 8.3% to 12.3%, while “moved” lines — a marker of refactoring — fell from 25% of changed lines in 2021 to under 10% in 2024, and copy/pasted code exceeded moved code for the first time in the data. GitClear presents this as coinciding with the rise of AI assistants, not as proof that they caused it.

  • Prefer reusing and refactoring existing code over a fresh generated block that duplicates it.
  • In review, look specifically for duplication.

Keep your own judgment switched on

Section titled “Keep your own judgment switched on”

In a study of 319 knowledge workers, Lee et al. (CHI 2025) found that higher confidence in AI was associated with less critical thinking, while higher confidence in one’s own ability went with more. The thinking does not disappear; it shifts toward verifying information, integrating responses and stewarding the task.

That shift is now part of the job. In Stack Overflow’s 2025 survey, 66% of developers named AI answers that are “almost right, but not quite” as their top frustration, 45.2% said debugging AI-generated code takes longer, and 45.7% distrusted the accuracy of AI output, against 32.7% who trusted it.

AI amplifies what is already there

Section titled “AI amplifies what is already there”

Google’s DORA research puts it plainly: “AI’s primary role is as an amplifier, magnifying an organization’s existing strengths and weaknesses.” Good tests, clear architecture and real review make AI more useful; without them, AI makes things worse faster. The practices matter more with AI, not less.

Rules of thumb

Section titled “Rules of thumb”
  • Own every line. If it is in your diff, you can explain how it works, why it is correct, and what it does on bad input.
  • Offload the mechanical, not the decisions. Boilerplate and repetitive transformations are fair game. Design, debugging and correctness calls stay with you.
  • Keep diffs small. Large generated changes are where “almost right” hides.
  • Tests before trust. Run it, try the edge cases, then believe it.
  • Never paste secrets or private material into an AI tool — credentials, tokens, private repository content, or anything that must stay out of a public repository.

Sources

Section titled “Sources”