Imagine a researcher describing a rare method that made an experiment possible. Remove the method and an important part of the interview is lost. Leave it in and someone might find the lab that used it. A research prototype called AURA tries to weigh those two costs as it rewrites transcripts.
The researchers’ Figure 1 shows AURA’s separate checks for privacy and useful context. Image by Li, Wen and Li, CC BY 4.0, unchanged.
Its authors separate the tasks of finding identifying passages and rewriting them. The system checks candidate rewrites for two things. A search agent tries to identify the speaker, and another evaluation checks whether the text retains useful facts about what the person said.
The comparison measured both costs
On the authors' tests with real-user interview transcripts, the version that varied how much text it rewrote produced fewer successful identifications across three attacker models than other tested methods without a formal differential privacy guarantee. In a like-for-like comparison with an earlier language-model anonymizer, it retained more of the contextual facts used in the benchmark at comparable privacy.
That is a comparison among particular defenses, attackers and transcripts. It shows how to measure a trade-off that a count of removed names would leave out. A release decision still depends on how sensitive the interviews are and what an attacker could learn elsewhere.