دانلود کتاب ارزیابی‌های قابل‌اطمینان برای مدل‌های زبانی بزرگ (LLM) و عامل‌های هوش مصنوعی؛ چارچوب‌های ارزیابی سرتاسری برای مدل‌های زبانی بزرگ و عامل‌های هوش مصنوعی خودمختار.

In February 2024, the British Columbia Civil Resolution Tribunal ruled that Air Canada was liable for negligent misrepresentation after its website chatbot gave a passenger false information about bereavement fares, and it ordered the airline to pay compensation (Moffatt v. Air Canada 2024). The decision, and the scrutiny that followed, shows that when AI touches customer facing _lows, wrong answers can create real financial, legal, and trust risks. These high-profile missteps are not isolated. In 2024, Google’s Gemini (previously Bard) had to abruptly pause its image generation feature after it produced “inaccurate or even offensive” images; executives publicly called the outputs “embarrassing and wrong” and apologized (Raghavan 2024). Together, cases like Air Canada and Gemini surface the same lesson: no matter how impressive a Large Language Model (LLM) appears in demos, if it is not rigorously evaluated and tested under real operating conditions, it can break a product, or shatter public trust in it. In Fig. 1.1, the AI deployment risk curve shows that risk increases exponentially as systems move from controlled testing to real-world deployment. The “last mile” to customers is where most failures occur.