Test — Verify AI Output Before It Ships

You have a coverage threshold. Most teams do. 80% on the line, green check on the PR, merge. That threshold was set in a world where a human wrote the code — someone who got tired near the edges, but who also understood the domain, remembered the null case from the last incident, and felt uneasy about the weird input before writing the test for it. Your 80% bar was tuned to that person's failure modes.

AI does not have those failure modes. It nails the happy path with unnerving fluency — the main flow is clean, idiomatic, well-named. Then it misses the boundary: the empty array, the off-by-one, the null from a code path it never saw, the unauthorized state, the second concurrent request. It misses these because it's completing the most probable continuation of your prompt, and boundaries are where the probable and the correct diverge.

So here is the uncomfortable arithmetic. AI-authored pull requests carry roughly 1.7 times the defects of human ones — 10.83 issues each versus 6.45 (CodeRabbit 2025). Bugs per developer are up about 54% year over year, against 9% the year before (Cortex 2026). And your gate — built for 2023 humans — waves it all through, because it was never calibrated for code that fails at the edges while looking flawless in the middle. That is a verification debt you are accruing at machine speed.

Test is two things at once. First, it is verification calibrated to how AI actually fails: a higher bar for AI-generated code, adversarial probing of the boundaries AI skips, and blocking gates that a green happy-path result cannot bypass.

Second, it is the loop that closes back into Curate. Every defect Test catches is a signal about what your context failed to teach the AI. Feed that signal upstream — into the AGENTS.md, the specs, the patterns — and the same class of error stops recurring at the source. Without that loop, you catch the same bug forever. With it, the method compounds.

Without Test, governance is theatre. You can curate context, refine every spec, and enforce every pattern, and still ship a defect that passed functional tests while looking perfect. Test is where the method meets production reality.

If your merge gate applies the same bar to AI code and human code — the overwhelmingly common case — you are letting roughly 70% more defects per pull request through on the AI-authored ones, and you cannot even see it, because the gate is green and the happy path is clean. AI-authored PRs carry 10.83 issues each against 6.45 for human ones (CodeRabbit 2025). Your dashboard says everything is fine. Production disagrees on a schedule you don't control.

And the volume is climbing. Bugs per developer are up 54% year over year (Cortex 2026). More AI code, more defects per unit of it, verified against a bar built for a slower, differently-failing author. That gap widens every sprint you don't recalibrate.

The teams that feel this most viscerally are your best engineers — reviewing AI PRs at 2x the old volume, sensing the boundaries aren't covered, with no gate that backs up their instinct. They are doing Test manually, in their heads, and burning out doing it.