SmileShark Unveils AI 'Hallucination Control' Method, Slashing Errors and Review Time

SmileShark is tackling one of the most persistent challenges in artificial intelligence: the 'hallucination' problem, where AI generates factually incorrect information. In a presentation at the AWS Summit Seoul 2026, Solution Architect Choi Byung-joo detailed a successful case study with Maekyung AX, a digital division of Maekyung Media Group. The project focused on applying an AI model to evaluate and verify content generated by another AI model, a method described as 'LLM-as-a-Judge.'

The results are striking. The number of factual errors in AI-generated content, which previously averaged 15 to 20 per month for Maekyung AX's AI podcast, has been reduced to less than one per month. Crucially, the time spent on human review and verification has been cut by over 90%. This dramatic improvement underscores the potential for AI to not only create content but also to ensure its accuracy and reliability.

Choi emphasized that a quality evaluation system is essential for AI services during their operational phase. He noted that it's not just about the performance of the AI model itself, but also about establishing robust mechanisms to catch and correct errors. This approach is vital for building trust in AI-generated content, especially in a media environment where factual accuracy is paramount. SmileShark's innovation offers a practical solution for businesses looking to leverage generative AI without compromising on truthfulness.

“

A quality evaluation system is essential for AI services during their operational phase.