Skip to content
Tech

Why do my AI-generated tests pass on the bugs they were meant to catch?

83

Возможность

AI coding assistants generate unit tests by reading the implementation, so assertions reflect what the code currently does rather than what it should do. A test written against a buggy function learns the bug as a fixture, not a violation. Developers shipping with 90% AI-generated coverage see mutation scores stay flat and real regressions slip through unchanged. The core issue is the oracle problem: there is no ground truth for correct behavior unless you supply it, and AI assistants have no access to the requirements or intent behind the code. Research from July 2025 confirms that LLM-generated tests frequently pass on the exact buggy code they were written against, failing only when the bug is fixed.

Почему это важно

A test suite that passes on broken code is worse than no test suite, because it turns the primary quality signal into a false guarantee.

Как я оцениваю возможность

Оценка возможности отражает моё личное суждение, а не точное измерение: насколько проблема болезненна, как часто она встречается и насколько мало существует решений сегодня. Чем выше оценка, тем более достойной реализации я считаю эту задачу.

Серьёзность8/10

Насколько серьёзные проблемы это создаёт.

Частота9/10

Как часто люди с этим сталкиваются.

Пробелы8/10

Насколько мало хороших инструментов для этого существует сегодня.

Ещё задачи, достойные решения

Tech

Почему программное обеспечение, от которого мы зависим больше всего, так неудобно использовать?

Tech

Почему я по-прежнему не владею ни одними из данных, которые генерирую?

Tech

This is a fundamental limitation of how most data deletion works today, and it comes down to a few core problems: **Deletion is an absence, not a presence** A receipt is a positive artifact -- proof that something *happened*. Proving a negative ("this data no longer exists anywhere") is cryptographically and architecturally much harder than proving a positive ("this file was signed at time T"). **No standard audit trail** Most systems are not built to emit a verifiable, tamper-proof log of what was deleted, when, and from which storage layer. Even if a company tells you "your data was deleted," that statement lives in their system, which they control. **Backups and replication** Data often lives in dozens of places: primary DBs, read replicas, backups, logs, caches, CDN edge nodes, analytics pipelines. A deletion receipt would need to account for all of them -- and companies rarely track this end-to-end. **No legal requirement to provide one** GDPR Article 17 gives you the right to erasure, and Article 19 requires notification to recipients -- but neither mandates a cryptographically verifiable receipt. Companies only have to confirm deletion, not *prove* it in a way you can independently verify. **Technical proof would require architectural changes** A trustworthy receipt would need something like: a Merkle-tree deletion proof, a signed timestamp from an independent witness, and a commitment scheme proving the original data existed before deletion. Almost no consumer product is built this way. The gap here is a policy and incentive problem as much as a technical one -- companies have little motivation to build verifiable deletion infrastructure when regulators don't require it.

Tech

Почему я не могу знать, соответствует ли то, что запущено, тому, что задекларировано в моём SBOM?

Tech

Почему каждая цепочка происхождения C2PA разрывается в тот момент, когда контент попадает в социальные сети?

Tech

Почему я могу перенести свои публикации, но не профиль, который платформы создали, наблюдая за мной?