Skip to content
Codrlabs Open

AI assistance

AI coding tools are part of how software is written now, and Codrlabs Open works with them rather than around them. This section sets out how to use them so they make you a better developer instead of standing in for one.

The principle

Section titled “The principle”

Offload the mechanical. Never offload the judgment.

Lars Faye draws the line well in AI Coding will Prevent Expertise: cognitive offloading delegates the mechanical or tedious, while cognitive debt abdicates your judgment and decisions. Offloading makes you faster. Debt makes you dependent, and it compounds.

For experienced developers the risk is mostly about quality and security. For people still building their skills it is sharper: the friction AI removes is the same friction that builds expertise. Faye’s “expert novice” is the result — a developer expected to steer tools that assume a judgment they have not yet had the chance to develop.

What the research shows

Section titled “What the research shows”
Finding Study
Developers learning a new library with AI scored 50% on a comprehension quiz, against 67% for those coding by hand. The biggest gap was on debugging, and the AI group was not significantly faster. Anthropic, 2026 — randomised trial, 52 developers
Students practising maths with unrestricted GPT-4 did 48% better in practice but 17% worse on the exam without it. A tutor version that gave hints instead of answers caused no significant exam harm. Bastani et al., PNAS 2025 — about 1,000 students
Novices already progressing used AI to write code they had planned and ignored bad suggestions. Struggling novices had their difficulties compounded and “finished with an illusion of competence”. Prather et al., ICER 2024 — 21 lab sessions
Experienced open-source developers took 19% longer with AI on their own repositories, while believing it had made them 20% faster. METR, 2025 — randomised trial, 16 developers, 246 tasks
Participants with an AI assistant wrote less secure code and were more confident it was secure. Perry et al., CCS 2023
At least 5.2% (commercial) and 21.7% (open-source) of packages suggested by code models did not exist. Spracklen et al., USENIX Security 2025
Higher confidence in AI was associated with less critical thinking; higher self-confidence with more. Lee et al., CHI 2025 — 319 knowledge workers

Reading this honestly

Section titled “Reading this honestly”

These studies are small or specific: 16 developers in METR’s setting, 52 in Anthropic’s, one school in Bastani’s. Anthropic measured comprehension straight after the task, not months later. The tools change quickly. No single result settles anything.

What holds across very different designs is the pattern:

  • The benefit tracks existing expertise. People who already know what good looks like use AI to go faster. People who don’t are the most likely to be misled, and the least likely to notice.
  • Learning suffers when AI does the thinking, and holds up when AI is made to teach instead — hints rather than answers, questions rather than code.
  • Confidence rises faster than quality. In security, in speed and in learning, people using AI rated their work higher than it measured.

In this section

Section titled “In this section”

Sources

Section titled “Sources”