Use AI for Tests — AI-Assisted Coding, Part 6

Contributor
6 days ago
3 min read

Updated: 4 hours ago

AI-Assisted Coding · Part 6

Generating tests with AI saves time. But generated tests can be fake-rigorous. Review carefully.

Step 1: Generate Tests for Existing Code (10 min)


"Write Jest tests for this function:
[paste function]

Cover:
- Happy path
- Empty input
- Boundary values
- Error cases"

AI returns tests. Review them.

Often: produces decent coverage in minutes.

Step 2: Verify the Tests Are Real (10 min)

Look for fake tests:


test("should work", () => {
  expect(true).toBe(true);
});

Or:


test("calls myFunc", () => {
  myFunc();
  // no assertion!
});

Read every assertion. Does it actually verify behavior?

Step 3: Watch for Tautological Tests (10 min)


test("formatPrice returns formatted price", () => {
  const result = formatPrice(100);
  expect(result).toBe(formatPrice(100));   // ← tests itself
});

Or:


test("...", () => {
  expect(complexFunction(input)).toBe(complexFunction(input));
});

Self-referential. Useless.

AI sometimes does this when it doesn't know the expected output.

Step 4: TDD-Style With AI (15 min)

Reverse:


"Generate Jest tests for a function called `parseDate` that:
- Takes a string
- Returns a Date
- Throws on invalid input

Output only the tests."

AI writes tests first. You then implement to make them pass.

Powerful: tests define the spec; you implement against it.

Step 5: Edge Cases Especially (10 min)


"Generate edge case tests for this function. What inputs might
break it?"

AI often spots cases you missed:

Very large inputs
Empty
Concurrency
Unicode
Boundary values

Trust but verify. Some are real edges; some are nonsense.

Step 6: Avoid Over-Mocking (15 min)

AI loves to mock everything:


jest.mock("../db");
jest.mock("../email");
jest.mock("../analytics");

Result: tests that pass when nothing works.

Mock only at:

True boundary (network, time, randomness)
External services

Internal code: use real implementations when possible.

Step 7: Integration Tests Help More (10 min)

For features: integration tests (real DB, real HTTP) > 100 unit tests with mocks.


"Generate Cypress E2E tests for this user signup flow."

AI writes the click-through. Real browser; real DB. Verifies the actual experience.

Step 8: Test Data Generation (10 min)


"Generate 20 realistic User objects for testing. Vary names,
countries, ages. Include edge cases (very long names, accents,
emojis)."

AI creates fixtures. Useful.

Better than:


const user = { name: "test", age: 1 };

That tells you the code works for test. Not for real users.

Step 9: Coverage Reports as Guide (10 min)


npm test -- --coverage

Find untested lines. Feed them to AI:


"This function is untested. Write tests."

Targeted gap-filling.

But: 100% coverage ≠ good tests. Quality matters more than line count.

Step 10: Trust Your Tests (10 min)

After AI generates:

Run them all
Look at failure messages (good error messages?)
Run them against a buggy version (do they catch it?)
Mutation testing if available

If AI's tests pass against broken code: they're not real tests.

What You Just Did

AI for tests: generate, verify real, no tautology, TDD-style, edge cases, no over-mock, integration, fixtures, coverage gaps, trust. Good test generation.

Common Failure Modes

Accept fake tests. "Tests pass" but verify nothing.

Over-mock. Tests don't reflect reality.

100% coverage with garbage tests. False security.

No mutation / negative testing. Don't know if tests catch bugs.

Generate; never read. Garbage tests committed.

Continue the AI-Assisted Coding path

Previous — Part 5: Use AI for Refactoring
Next — Part 7: Use AI for Debugging

Part of the AI-Assisted Coding learning path.

ShiftQuality