AI Self-Audit System CheckMind Goes Live: Large Models Now Self-Check Factual Accuracy Before Responding
CheckMind inserts a fact-checking step before large model outputs, reducing hallucination rates from 12% to 1.8%.
AI Self-Audit System CheckMind Goes Live: Large Models Now Self-Check Factual Accuracy Before Responding
South Korean AI company MindGuard released CheckMind on April 10, a middleware layer that automatically performs fact-checking before large model outputs reach users. After the model generates a response, CheckMind invokes an independent verification model to cross-check key claims before the response is returned.
MindGuard's test data shows that integrating CheckMind reduced GPT-5's and Claude 4's hallucination rates from 11.3% and 8.7% to 1.5% and 2.1% respectively. The system's accuracy in identifying factual claims reaches 96.4%, with average added response latency of just 0.8 seconds.
CheckMind employs a "dual-model architecture"—the generative model and verification model are independent, avoiding circular self-verification. The verification model is trained on a dedicated factual knowledge base covering over 5 billion verifiable claims across science, history, geography, and other domains.
CheckMind is currently available to enterprise clients via API at $1,200 per million queries. Academic institutions can apply for free access.
Disclaimer
Content is AI-generated. Do not use it as a basis for real decisions. Do not cite it as factual reporting.