Where AI Recruitment Stops Helping | AI Learning Guide

Both limits come from the data, not from the method. Privacy first: health records are highly sensitive, and rules on using them for research differ by country and region. In practice a search often runs only inside one institution, or only after approval or consent, or only on de-identified data. Each of those constraints shrinks the pool the search can see. Then record quality: the search can only read what was written down, and records are often incomplete, inconsistent, or out of date. A diagnosis under an unusual abbreviation, a result stored as a scanned image, a medication that was stopped but never removed. Any of these can hide an eligible person or flag an ineligible one. So the honest claim is a faster search over the records that are available and readable, not a complete search over everyone who exists.

0:00 / 0:00

Auto-play

Privacy constrains the reach of the search

Health records are highly sensitive, and rules on using them for research differ between countries and regions. The practical effect is that a search often runs only within one institution, or only after approval or consent, or only on de-identified data. Every one of these constraints shrinks the pool the search can see, and a smaller pool yields a smaller shortlist.

Record quality sets the ceiling on accuracy

AI can only work with what is recorded. Diagnoses written in free text, results stored as images, and medications that were stopped but never removed all create errors in both directions: an eligible person can be missed, and an ineligible person can be flagged. The shortlist is only as good as the records behind it.

The realistic claim is a faster search over the records that are available and readable, not a complete search over everyone who exists. Human review is what catches the cases the records got wrong.