Testing and Debugging with AI

A starting question

Your checkout page sometimes shows an empty cart after you add an item. An assistant says it is a “race condition” and writes a test. How do you know that diagnosis is right?

You reproduce the failure, inspect the evidence, and run a test that captures the intended behavior. The AI explanation is a hypothesis—a possible cause—not a measurement. GitHub warns that AI review and code suggestions can be wrong even when they sound convincing.Verify the particular claim against the running application.¹

Three useful test scopes

A unit test checks a small piece of logic, such as a function that calculates a total. An integration test checks pieces working together, such as a route and its data-access layer. An end-to-end test follows a user flow through the running application, such as adding an item and seeing it in the cart. Boundaries vary by project; these labels describe scope, not a ranking of importance. Browser tools such as Playwright support end-to-end checks.A browser check can reveal wiring problems that a small unit test never reaches.²

For the cart bug, a unit test might confirm the cart-total calculation, while an end-to-end test may expose that the UI never applied the latest response. One does not substitute for the other.

Use AI for suggestions, not verdicts

Give the assistant the observed behavior, expected behavior, relevant code, and a failing test or error message. Ask for possible causes and the next discriminating check. A useful suggestion might be, “Log the request ID and compare the order in which responses finish.” A poor suggestion is, “Rewrite the cart state layer” before the failure boundary is known.

When AI drafts a test, read its assertions. Does it check the behavior users need, or merely repeat the implementation? Would the test fail if the bug returned? Run it and deliberately inspect a failing version where practical. Automated checks can also run in CI, a system that builds and tests proposed changes before integration; GitHub’s testing documentation describes that role.CI reports the checks that ran, not all possible behavior.³

When the error explanation sounds certain

An error message often describes where a failure surfaced, not where it began. If an API request fails, check whether the browser sent it, whether the server received it, and what response returned. AI may help translate the stack trace, but your logs, network panel, and reproducible steps are stronger evidence.

Recap and try it

Use AI to generate hypotheses and test ideas. Keep the feedback loop grounded in observed behavior and checks that would catch the original bug.