Check an AI answer against its evidence

Try a bounded example

The FActScore study separates text into individual factual claims and checks support against a specified source. Here is our fictional practice exercise, not a recorded model test. Notice: Saturday entry is free for residents; closing time is 18:00. Answer to check: everyone enters free on Saturday and the museum closes at 18:00. Separate day, eligibility and closing time. Expected result: Saturday and 18:00 are supported; free admission for everyone is not. The resident restriction must remain in the corrected answer.

Open the evidence

A source link is a starting point. Google explains that Gemini links can lead to material related to parts of a response; some responses have none. When available, Sources opens the relevant links. Our suggested check: open the page and locate the passage supporting your exact claim. If it only discusses a related subject, that link has not established support.

Produce and inspect a correction

For this exercise, provide the notice and the answer. Following Anthropic’s guidance, ask the assistant to use only the supplied material, identify supporting passages and acknowledge missing information. Use one claim, passage and result per row. Compare those rows yourself with the notice. Remove the unsupported universal admission claim instead of merely softening its wording. These techniques reduce hallucination risk; they do not guarantee a correct answer.