Orientation
- explain unfamiliar vocabulary;
- suggest search terms and variant language;
- identify kinds of repositories to investigate;
- turn a broad topic into narrower questions;
- summarise a source you have actually supplied.
MODULE 7 · AI & EPISTEMIC DISTANCE · 45–70 MINUTES
Primary, secondary and tertiary sources describe relationships between evidence, interpretation and synthesis. Generative AI complicates that model because one answer may compress patterns from many layers while hiding the provenance chain behind individual sentences.
Traditional tertiary source: usually a stable synthesis or reference work whose editorial form and references can often be inspected.
Generative AI output: a newly produced synthesis that may combine learned patterns, retrieved material, user-provided context and inference without preserving a transparent source trail for every claim.
So do not teach “AI = tertiary source”. Treat AI as a mediation layer whose usefulness depends on what job you ask it to do and whether you can trace important claims back to evidence.
If your question is “What did this AI system say when asked about the Dawn Raids on this date?”, then a saved transcript of the interaction can be primary evidence of that model output.
That does not make the claims inside the answer primary evidence that the historical events happened as described. For those claims, you still need the underlying records, testimony, data or scholarship.
This distinction matters for research on AI itself. If you study how a model frames an issue, record enough provenance to make the interaction inspectable: system/model if known, date, exact prompt, relevant supplied context, and the output you actually received. Model behaviour can vary across versions, prompts and runs.
Epistemic distance here means the distance between a claim and the inspectable evidence that could support it. Distance is not automatically bad — synthesis is useful — but hidden distance creates risk.
A confident sentence does not tell you where the claim originated or whether the model can recover the real source.
An invented quotation does not become real because it sounds historically plausible.
Several AI systems or websites may repeat the same upstream material or one another's framing.
A model can summarise disagreement, but fluency is not authority to decide which interpretation the evidence best supports.
This paragraph is not presented as a real AI output and contains no claims you should trust merely because they appear here.
“During a major 1970s immigration enforcement campaign, police targeted Pacific communities because officials believed overstaying was concentrated there. Contemporary newspapers largely supported the policy, while later historians have shown that enforcement was racially disproportionate. One minister privately admitted that the campaign was designed mainly to reassure anxious voters.”
Your task is not to decide whether the paragraph “sounds right”. Break it into separate claims.
Now use the Trace the Claim Home tool to write what evidence would be required before each claim entered a serious argument.
If an AI supplies a citation, there are at least four possibilities:
The correct response is the same in every case: open the source yourself. A citation is a route to checking, not proof by formatting.
These prompts reduce some failure modes; they do not remove your responsibility to check the evidence.
Search engines, databases, recommendation systems and generative models increasingly mediate which evidence becomes visible. This creates a methodological question: if your research process depends on ranked or generated outputs, how should you document the role of the system in shaping your evidence base?
Extension: write a short methods note explaining how using an AI assistant could alter (a) what sources you discover, (b) which interpretations seem prominent, and (c) how easily you notice minority or poorly digitised perspectives.
Explain this distinction without using the words “AI is bad”:
“AI can be useful for research while still being a poor final evidentiary authority.”
Your answer should mention provenance, mediation and at least one legitimate use. Then give one example of a research question for which an AI transcript itself could be primary evidence.