What is AI fabrication in procurement documents?

AI fabrication is not just a wrong conclusion. It is an added fact: a date, quantity, status, clause, supplier condition, or denominator that the source file does not support.

That distinction matters in procurement because many deliverables are built from fields that readers do not re-check. A risk register, quote comparison, or catalog merge can look finished because every cell is filled. The dangerous part is that one filled cell may have no source behind it.

In the July 2026 PAIR-20 run, fabrication appeared in two tasks built to test that discipline. In the supplier-risk-register task, Copilot invented a Dravec backup expiry that the file did not state. In the MRO catalog task, Copilot wrote "Likely 100ct" for Atlas's unstated box quantity. The right answer was to mark the field as unknown and confirm it, not turn it into a hard number.

PAIR-20 treats fabrication as task-failing because procurement teams cannot defend an unsupported fact. A missed issue is bad, but a reviewer can still find it by returning to the source. An invented expiry or guessed pack quantity sends the team after evidence that never existed.

Fabrication usually hides in useful-looking work. It does not announce itself as a wild paragraph. It often appears as one ordinary value inside a table: a backup status, contract date, pack size, or assumed unit.

The control is source discipline. Ask for a source-backed field, not a complete-looking field. If the file is blank, the output should say "missing," "unknown," or "confirm with supplier."

Review added facts as closely as missing facts. Before a procurement document leaves your desk, check each decisive value against the source file. If you cannot point to the line that supports it, it should not be in the final answer.