Many AI tools now add citations to what they generate. That's progress, but a citation is a claim about evidence, not evidence itself. For training people rely on, provenance has to hold up.
The limits of plausible evidence
A model can cite the right document and the wrong paragraph, cite a passage that only partly supports the claim, or cite a version that has since changed. All three look fine at a glance.
Four properties of trustworthy provenance
- Specific: points to a passage, not just a file.
- Verified: checked that the passage actually supports the claim.
- Versioned: tied to the source version that was approved.
- Reusable: stored so change analysis and audits can use it later.
From citations to checks
Treat every link as something to test. Unsupported claims should be flagged before review, not discovered after release. Ambiguity in the source should be preserved and surfaced, not smoothed over.
A realistic path
Start by instructing the model to use only the source and flag gaps, which is what Coursecast does today, then add claim-level linking and automated checks. That's the evidence graph on our roadmap.