Can AI tell true duplicates from false twins in item data?

A false duplicate is an item that looks matchable in a catalog merge but is not the same buying item. The description may be close, while the grade, thickness, material, clearance class, compliance rating, part-number suffix, or pack quantity changes the answer.

PAIR-20's July 2026 T06 task tested that problem with three synthetic MRO distributor catalogs. The job was to build a merged per-piece comparison, list false duplicates, and recommend a consolidation path. The task brief told the platforms to match on function, dimensions, material, and grade, not description keywords alone.

The nitrile-glove row was the important test. The products differed on thickness and pack quantity, and Atlas left its box quantity unstated. ChatGPT, Claude, and Gemini treated the gloves as not directly comparable and flagged the missing Atlas box count as a field to confirm before pricing.

Copilot broke on the blank field in the naive run. It wrote "Likely 100ct" for Atlas's unknown quantity. The case file said Atlas's box quantity varies by item and must be confirmed, not assumed. Under the PAIR rubric, a fabricated data point fails the task because the reader cannot see that the table contains a guess.

A true duplicate needs more than a similar description. In MRO data, the match has to survive the spec fields that change use, safety, compliance, and unit price. AI can help surface likely matches, but the false-duplicate list is the part to review first.

Blank fields are procurement facts too. "Confirm with supplier" is a useful output. A guessed pack size is not; it becomes the denominator for the per-piece price and can change the cheapest-source answer.

Use AI for the first pass, then inspect the dangerous rows. Start with the merged table, but verify rows with missing pack quantities, near-match descriptions, and any spec difference that would make two items unsafe to substitute.