AI is used inside Xobin for interview scoring, response summarisation, and candidate recommendations. Each of those surfaces has a different failure mode: a scoring model can drift, a summariser can distort, and a recommender can amplify patterns already present in historical hiring data. We treat each surface as a separate audit target rather than a single "AI" black box, and document the specific risks that apply.
Every AI-scored interview is auditable on three properties: agreement with blind human raters, stability over time, and demographic parity across the candidate groups defined for the deployment. Audits run on rolling candidate samples so the system cannot silently overfit to a fixed benchmark, and inter-rater agreement between the AI and human raters is reported with the same statistics we would use to compare two humans.
An AI score is never the only signal a hiring team sees, and it can always be overridden. When the model's score distribution shifts or its agreement with human raters slips, a re-validation gate is triggered before it continues in production and affected customers are notified as part of that gate — not after. Transparent AI reporting is the point: customers can see what the model did, on which sample, and with what confidence.
Every AI interview Xobin runs is held to the audit standard described on this page. Teams who want the same rigour in their own hiring can use it directly through Xobin.
The methodology described here is what runs inside every Xobin assessment and AI interview — so hiring teams can make talent decisions on evidence, not instinct. See Xobin.