An analyst compares an extracted answer with the original PDF. The extraction contains a longer sentence than the one visible on the page.
It looks like the system has added something.
While investigating an extraction discrepancy, we found that the additional text was already stored in the PDF’s form field.
No language model was involved in that extraction. Yet the extracted answer and the visible source did not match.
What exactly should the reviewer verify?
What the PDF stores and what it shows
A PDF form field has a stored value and a visual representation on the page. If the text extends beyond the field’s visible area, a tool may retrieve the full value even though the reviewer sees only part of it.
This simplified example illustrates the observed pattern; it does not reproduce the source document:
Reviewer sees
“Reviewed annually…”
PDF field stores
“Reviewed annually and after material changes.”
The additional words introduce a condition that matters. Before accepting them, the reviewer needs to understand why they are absent from the visible page.
A page reference may not be enough
The reviewer can be looking at the right document, on the right page, and still be unable to confirm the full answer.
Additional text should not silently become accepted evidence simply because a tool can retrieve it. Equally, a mismatch should not automatically be labelled an AI hallucination. The discrepancy needs to be understood before either version is treated as authoritative.
Traceability should explain which representation of the source produced the answer.
That gives a source reference a practical purpose: helping someone establish how the extracted answer relates to the document they are reviewing.
Make the discrepancy reviewable
For MultiplAI, the design lesson is specific: show what was extracted, identify whether it came from a stored form value or the visible page, and make any difference clear.
The system should leave the choice of which version to accept open to review. Depending on the discrepancy, further inspection or clarification from the document’s provider may be necessary.
When an extracted answer differs from what the reviewer can see, the system should help explain the difference—not leave the reviewer to guess.