AI code review is any system that uses machine learning to automatically analyze code — flagging bugs, style violations, security issues, or evaluating how well a developer understands what they've read. The term covers a wide range of tools with very different goals.
Most people mean one of two things when they say "AI code review": AI-powered static analysis tools, or code comprehension evaluation systems.
AI-powered static analysis tools — like GitHub Copilot suggestions, Amazon CodeGuru, or Sourcery — analyze your code as you write or submit it. They look for bugs, style problems, performance issues, and security vulnerabilities. The output is a list of suggestions: "this variable is never used," "this SQL query is vulnerable to injection," "this loop could be O(n²)."
Code comprehension evaluation systems are a separate category: they score how well a human understands a piece of code. Instead of analyzing the code itself, they analyze the human's explanation of the code. The output is a structured assessment: how accurate was the explanation, how complete, how clear, how insightful. CodeEvaluator is this second type — which is distinct from code review tooling, though both involve reading code carefully.
Static analysis tools examine source code using rules, program-analysis techniques, and language-specific models to detect potential issues without executing the program. They can catch a real class of bugs — especially the mechanical ones that show up repeatedly in codebases.
AI-powered review systems go further: they use machine learning or large language models to reason about code changes in context, flagging not just rule violations but likely design problems and missing considerations. The practical result is useful but imperfect — these tools miss the conceptual bugs that require understanding the business domain, the architectural intent, or the implicit contract between modules.
Comprehension evaluation is harder. The AI needs to assess not just whether a statement is true, but whether it demonstrates genuine understanding at the right level of depth.
A well-designed comprehension evaluator breaks the assessment into dimensions. At CodeEvaluator these are: Accuracy (is what you said factually correct?), Completeness (did you cover the important parts?), Clarity (would another developer understand your explanation?), and Insight (did you go beyond surface description — edge cases, complexity, design tradeoffs).
Each dimension is scored independently because they fail independently. You can be accurate but incomplete. You can be complete but unclear. You can be clear but shallow.
Static analysis AI misses anything that requires understanding intent. It cannot tell you that a function does the wrong thing if it does it consistently. It cannot detect architectural drift. It cannot catch a bug that only manifests under a specific business rule it has never seen documented.
Comprehension evaluation AI has a different limit: it can only assess what you wrote, not what you know. A developer might understand something deeply but express it poorly. A developer might write a fluent-sounding explanation that is subtly wrong. The AI scores the text, not the mental model behind it.
Both types are tools, not oracles. They are consistent, scalable, and available instantly. But they work best alongside human judgment, not instead of it.
Static analysis makes you aware of problems in code you wrote. Comprehension evaluation makes you aware of gaps in your understanding of code you read.
The feedback loop is different. Static analysis says: "here is a bug in your output." Comprehension evaluation says: "here is where your mental model of this code breaks down." The second kind of feedback is rarer and more valuable for skill development.
Most developers get the first kind constantly — linters, CI checks, code review comments all point at problems in code they wrote. Almost no one gets the second kind systematically. That gap is what AI comprehension evaluation is designed to close.