An AI company has developed a new model that can catch hallucinations produced by large language models and explain why they’re wrong. AI hallucinations are a significant issue, with off-the-shelf GPT-4 hallucinating 5-10% of the time, and open-source LLMs performing even worse. To address this, the company created an automated fact-checking pipeline that catches 99% of errors before they reach users. Aporia Guardrails, developed by AporiaAI, has outperformed NeMo, GPT-4o, and GPT-3.5 in AI hallucination detection and latency.