I probed my own AI‑writing linter 16 ways. The first probe lied to me.

Описание и применение

Автор исследовал AI‑сканер текста, обнаружив, что он дублирует слова, не игнорирует код и читает HTML‑комментарии, что привело к неверным выводам о плотности слов. Переместив список запрещённых слов и пересмотрев интерпретацию баллов, он исправил ошибки.

Задача

Relying on an AI‑writing linter without understanding its scoring logic led to incorrect assumptions about word usage and content filtering.

Результат

The linter deduplicates flagged words, does not strip code, and reads HTML comments, contradicting initial density‑based assumptions. Adjusting the word list location and score interpretation resolved false positives.

Рабочий процесс

• Write a test string to a file
• Run the scanner script and capture output
• Parse score and flag count from output
• Repeat with varying word counts and contexts
• Analyze results to understand scoring logic
• Move word list to external file
• Adjust how scores are interpreted

Доказательства

The detector flags a specific set of filler words. leverage is on it, so it is my test particle.

Источник

It deduplicates. Five occurrences cost exactly what one costs: score 3 either way.

Источник

It does not strip code. The word inside a fenced block scored the same as the word in plain prose.

Источник