An audit compares staff trust in nine of the University's internal AI reference assistants against how much of each unit's own archive the assistant can actually search.
An audit of nine internal reference assistants finds staff trust tracks how little of a unit's own archive each is allowed to search, not how often it answers correctly
Ask the Adaptive Metrics Lab and the Office of Research Outputs the same kind of question and each answers through an AI assistant built to search a different slice of the University’s own records: the Lab’s assistant can see nearly all of its unit’s project archive, the Office’s assistant well under a fifth. The School of Continuous Improvement has now compared staff trust in both assistants, and seven others besides, against a technical audit of what each is actually permitted to search and an independent test of how often each gets a verifiable question right.
The pattern holds across all nine: the assistants configured to search the smallest share of their own unit’s archive are consistently the most trusted. How often an assistant is actually correct explains much less of the difference, and telling staff a deployment’s own accuracy score before asking them to rate it barely moved the number.
The School reads the finding as consistent with its broader interest in how an instrument earns cooperation rather than merely deserving it.
People aren’t rating the tool. They’re rating how curated it feels.
— Associate Professor Casimir Beng, Lead of the Adaptive Metrics Lab, who led the audit
“Every one of these assistants was set up separately, by whoever had the access-control list to hand at the time,” said Dr Renke Sabel, Senior Lecturer in the School of Continuous Improvement. “None of us designed the pattern we found. We just went looking for it.”
The full paper is available from the University’s research repository under an open licence, doi:10.5555/slop.6rwku5.