I Built a YouTube Credibility Scanner. Its Worst Bug Was Accusing Honest Creators

Описание и применение

Разработчик создал инструмент оценки доверия YouTube‑видео, но изначальная система распознавания аффилиатных ссылок ошибочно помечала собственные ссылки создателей как фишинг. После разделения сигналов на жёсткие и мягкие, точность улучшилась, но остаются ложные срабатывания и проблемы с оценкой дизлайков.

Задача

The initial affiliate detector flagged every outbound link in a YouTube description as a funnel, incorrectly accusing creators who linked to their own products.

Результат

After the fix, 15 of 20 new AI‑related videos were flagged as "Be Careful" and 5 were deemed "Likely Legit", matching expected patterns of affiliate funnels versus legitimate product sales.

Рабочий процесс

• 1. Identify outbound links in the video description.
• 2. Classify each link as hard or soft signal based on patterns.
• 3. If hard signals present, add weighted hard signal score and a small boost from soft signals; otherwise add no score.
• 4. Aggregate per-comment sentiment scores instead of a single aggregate.
• 5. Apply a raised baseline to prevent premature flagging.
• 6. Output the final credibility report.

Доказательства

My first affiliate detector treated every outbound link in a description as one signal.

Источник

I split detection into two tiers: Hard signals: known affiliate network patterns, referral parameters, discount-code formats; Soft signals: shorteners, redirects, link density.

Источник

After the fix, I ran the scanner on 20 new “make money with AI” videos, mostly dropshipping, digital products, and AI side hustles. 15 came back “Be Careful”, 5 came back “Likely Legit”.

Источник