I Tracked Every AI Hallucination for a Week — The Numbers Were Worse Than I Thought

Описание и применение

Разработчик провёл эксперимент, проверяя 200 выводов ИИ, обнаружив 18 % ошибок. Создал быстрый слой проверки, который проверяет цитаты, код, утечки и аргументы, снижая ручную работу.

Задача

AI agents frequently produce hallucinations—confidently wrong outputs, fabricated citations, tool argument errors, and prompt leaks—leading to wasted token budgets and debugging effort.

Результат

The verification layer runs in under 100 ms on CPU, is model-agnostic, and free, enabling developers to audit agent outputs daily and uncover hidden hallucinations.

Рабочий процесс

• Generate output with AI agent
• Pass output to verification layer
• Citation checker validates references
• Code validator checks syntax and logic
• Safety leak detector scans for prompt leaks
• Argument verifier ensures tool call parameters match schemas
• Coherence scorer compares output to original prompt
• If all checks pass, output is accepted; otherwise, flag for review

Доказательства

Out of 200 outputs across Claude, GPT, and DeepSeek: 36 were confidently wrong (18%)

Источник

All runs in under 100ms on CPU. Model-agnostic. Free.

Источник