Is an AI answer supported by its source? Each sentence of the answer is checked against the source with hallucination-guard-small, a multilingual groundedness classifier, running entirely in your browser. The first run downloads about 270 MB, which is then cached.
Source
AI answer
Green = supported by the source, red = not supported (added, changed or contradicted). It checks support by the given source only, not world knowledge. See the model card for measured accuracy. This demo runs the small model, which misses some paraphrases (it can mark "weekdays" as unsupported when the source says "Monday to Friday"). The base model is more accurate, and MiniCheck and HHEM are stronger on English claim-level benchmarks.