Variation Surfacing

A technique for helping users assess AI reliability by generating multiple responses from one or more AI models and systematically presenting the differences, agreements, and unique mentions across those responses. In the context of image descriptions for blind and low vision users, variation surfacing involves querying multiple multimodal language models, decomposing their responses into atomic facts, comparing those facts across responses, and presenting the results in formats that highlight where models agree, disagree, or provide unique information. Research shows that variation surfacing increases BLV users' ability to identify unreliable claims by nearly 5x compared to single descriptions and significantly reduces inappropriate over-trust in AI output.