How does a person trace an AI claim to its source? (T1 Evidence Verification)
An AI summary can mix what the AI wrote, what it found, what someone said and what someone checked. T1 Evidence Verification asks a person to tell them apart and to check the claim a decision depends on.
The short answer
In AIRS-J/1, T1 Evidence Verification asks a person to sort an AI summary's claims into four kinds. The person then checks the claim the decision depends on against an available source before acting.
What T1 describes
Judgment is how well a person decides once AI has given the answer. AIRS-J/1, the Standard that AIR APAC publishes, describes it in nine dimensions. T1 Evidence Verification sits in the Test domain: "Interrogating evidence, including machine-generated evidence."
The Standard's words for T1 are: "Traces claims to source. Distinguishes generated, retrieved, asserted and verified content. Applies verification effort proportionate to consequence." In plain words, T1 traces a claim to its source and checks in proportion to what depends on it.
The four kinds of content
- Generated: the AI wrote it.
- Retrieved: the AI found it in a source.
- Asserted: someone said it.
- Verified: someone checked it against a source.
How much to check
The Standard does not ask a person to check every claim with the same effort. A published anchor describes that effort as heavy where the decision is irreversible or high in consequence, and light where it is neither. The threshold anchor reads: "Checks the load-bearing claim against an available source before acting, unprompted." A load-bearing claim is one that the decision depends on.
The opposite is also a mistake. In the Standard's words, "rejecting AI output wholesale without examining it is a failure to verify, not caution."
One fictional decision, worked through
A fictional example.
The situation
Priyanka Das is the procurement manager at a regional cold-storage operator, with its head office in Singapore and warehouses in Johor and Bangkok. A contract with a refrigeration maintenance supplier is due for renewal by the end of the month. The supplier's staff work inside the warehouses.
What the AI produced
She asks DowitcherAI to summarise the supplier's due-diligence file. One line of the summary reads: "The supplier passed its last fire-safety inspection. Taken from page 4 of the file. I am 90 percent sure."
The check
Priyanka sorts the summary's claims into the four kinds. Most of them matter little to this decision. The inspection claim is the load-bearing one, because the supplier's staff work inside the warehouses. She opens page 4. It holds the supplier's own statement that it passed. The inspection report is not in the file. The summary had shown the supplier's statement as if it were the inspection result. The claim is asserted, not verified.
The decision, and who was told
She does not sign yet. She asks the supplier for the inspection report. She tells the legal team and the Johor warehouse manager why the renewal is on hold. In the file, she notes which claims she verified and which she did not.
What happened after
The report arrives and supports the claim. The contract is signed a week later, before the renewal date. The team now marks each line of a due-diligence summary as checked or not checked.
What this does not claim
- Priyanka Das and her company are fictional. She has no level, band or confidence. Those come only from an Assessment.
- This example is not an assessment scenario and is not preparation for one. Reading it does not change anyone's chance of certifying.
AIR APAC publishes the standard for judgment in AI-enabled work, certifies people against it through its official instrument, and stands behind every result.