In SourceCheckup the same answers of GPT-4o with web search give 75.7 percent supported statements and 38.4 percent fully supported responses on HealthSearchQA, and about 70 percent supported statements against 55 percent fully supported responses on the 800-question set; a response-level rate and a statement-level rate are different quantities and cannot be set side by side across studies.

The first parent contributes both units on one sample with intervals (74.0 to 77.2 and 26.7 to 49.3). The second contributes the same pair on the main set (approximately 30 percent of statements unsupported, 55 percent response-level support) and the summary that 50 to 90 percent of responses are not fully supported across seven models. The arithmetic behind it is that a response counts as supported only if every statement is, so the response rate falls with the number of statements per response; the paper does not print that number, so the step does not compute a prediction.

A dataset in which response-level full support is not lower than statement-level support although responses contain several statements, which would mean unsupported statements cluster in few responses. Query: distribution of unsupported statements per response in the SourceCheckup release.

Half of AI answers are not supported by their sources. Too coarse as a citation rate: in the same data three quarters of the individual statements are supported.