Can an AI answer be used as a source for a fact about Mike Millett?
No. A model states a claim with no publisher behind it, so there is nothing to check and nobody to correct. If the claim is true there is a real source, and that source is what gets cited.
In brief
A model can state something about Mike Millett that no page states, here or anywhere. That is a claim with no publisher.
Mike Millett founded Digilu in 1999 and still runs it. That sentence is published here, so a model repeating it is repeating this site.
The case that needs a rule is the sentence on no page at all.
What an AI answer is good for
An AI answer can start an investigation and can never finish one.
If the claim is true there is a real source, and that source is what appears on the page.
If the only possible source is Mike Millett himself, it is written as a first-party claim with the attribution inside the sentence.
If it is false, nothing is published.
What the answer does show is how the record is currently being understood. That makes it a reading, not a source.
Three ways an AI answer misleads
This site’s own sentence, coming back
A model states a fact this site published. It reads like corroboration and is a reflection: the confidence it lends was borrowed from this site.
Two true things joined by an inference
A title and a company overlap in time, and the answer joins them as cause. Both halves check out, the sentence does not, and the join is what no source confirms.
A claim attributed to a page that does not carry it
The answer names a publication, which feels like the end of the work. Go and read it: the claim is sometimes not on the page.
When one source looks like two
This site publishes a sentence. A model retrieves it. Somebody writing about the subject asks the model, and publishes the sentence in their own words.
That article now looks like an outside source, and citing it here would give one origin two footnotes.
Independent
Dated before the claim was published here, or naming a record that is: a filing, a printed program, an article from the time.
Downstream
Published after, carrying no source of its own, often tracking the wording here. It may be true. It is not a second confirmation.
Two questions, and a source has to pass one. Did it exist before the claim was published here? Does it name a record that did? A no to both means downstream.
What this cannot fix
Independence cannot be proved. There is no way to know whether an outside writer knew a fact already or read it here. A date is the only signal and dates get changed.
When a model states something wrong about a person there is no publisher to write to and no edit to confirm. All this page can do is make the correct version easy to find, and say so plainly.
Related
Where a claim goes when the investigation finds nothing: when the only source is the person.
For the scheduled checks, see how the record gets checked.
For sources that exist and go away, see when the proof lives on somebody else’s site.
Common questions about AI answers as sources
Can an AI answer be used as a source for a fact about a person?
No. A model states a claim with no publisher behind it, so there is nothing to check and nobody to correct. An AI answer can start an investigation and never ends one: if the claim is true there is a real source, and that source is what gets cited. A named publication is a place to look, because it sometimes does not carry the claim.
What is circular sourcing, and how do you test for it?
A site publishes a fact, a model retrieves it, somebody writes it up with the model’s help, and the original site then cites that write-up as independent confirmation. Two citations, one origin. The test is two questions: did the source exist before the claim was published here, and does it name a record of its own that did? A no to both means downstream, not independent.
What is an AI answer about a person good evidence of?
Of how the record is currently being understood, which makes it a reading rather than a source. Record the date, the exact question, which system answered and the full reply text. It cannot be reproduced later: the same question returns different words within the hour and versions are retired, so the stored text is the entire record.