Plausible but Wrong

AI-generated code has a unique failure mode: it looks right. It compiles. It passes linting. The variable names are reasonable. The structure follows conventions. A quick scan suggests it does what you asked.

But the logic is subtly wrong in ways that are hard to spot precisely because everything else looks so polished.

The Pattern

Human-written bugs tend to be obvious - typos, off-by-one errors, missing null checks. You develop an eye for them over years of code review. AI-generated bugs are different. The code reads like it was written by a competent developer who misunderstood the requirements.

Common examples: an API client that handles the happy path perfectly but silently swallows errors instead of propagating them. A sorting function that works for the test cases but breaks on edge cases the AI never considered. A database query that returns correct results for small datasets but has O(n^2) performance that only shows up at scale.

How Review Habits Need to Change

Traditional code review asks "does this code look correct?" With AI-generated code, you need to ask "what would happen if this code is wrong in a way I can't see?"

Concrete steps that help:

• Test the unhappy paths first. AI code almost always handles the happy path correctly. The bugs hide in error handling, edge cases, and boundary conditions.
• Check assumptions, not syntax. The syntax will be perfect. Question whether the approach itself makes sense for your specific use case.
• Run it with adversarial inputs. Feed it empty strings, null values, extremely large numbers, and concurrent access. AI-generated code rarely accounts for these unless explicitly prompted.
• Diff against your mental model. If you expected the solution to use approach A and the AI used approach B, investigate why. Sometimes B is better. Often B is plausible but subtly inappropriate.

The Desktop Agent Advantage

A desktop agent that can run your code, observe the results, and iterate on failures catches these bugs during development rather than in production. The agent becomes its own reviewer, testing edge cases automatically before presenting the final result.

Fazm is an open source macOS AI agent. Open source on GitHub.