AI agents hallucinate prices: LLM grounding with live data (blog.apify.com)

Recent testing of AI agents, specifically GPT-6 Luna and Claude Opus 5.5, has revealed persistent issues with "hallucinations" in pricing information due to reliance on outdated data. In trials, both models provided accurate prices for only a small fraction of products when sourcing from memory, often defaulting to old launch prices. While integrating live data from the E-commerce Scraping Tool improved accuracy, errors persisted. The study highlights three main causes for these inaccuracies: LLMs recalling outdated prices, errors in fetching the correct web pages, and tools returning false confirmations of successful data retrieval. This research is crucial for the AI/ML community as it underscores the importance of grounding AI agents in real-time, reliable data to avoid misleading users. Implementing a robust grounding strategy, which includes clear data sourcing and effective error handling, can mitigate the potential for misinformation in pricing. The findings serve as a guide for developers, emphasizing the need for structured data extraction methods, like those provided by the E-commerce Scraping Tool, which standardizes output across multiple platforms and improves accountability in AI-generated responses. This ensures that AI agents can provide accurate and trustworthy information, enhancing user experience and trust in automated systems.