AI text interview

Video interview scoring had a bad decade and deserved it. Systems that rated applicants on facial expression and vocal tone were sold widely, then withdrawn or restricted after regulators and researchers looked closely. The core problem was never the accuracy of the model. It was that appearance and speech carry disability, ethnicity, age and neurotype, and a model trained on past hiring outcomes learns to reward whatever the past rewarded.

Text sidesteps the whole category. What is being read is what the talent wrote about the work. A talent with a stammer, a strong accent, a facial difference, or a webcam pointed at a shared kitchen is on identical footing with everyone else, and no adjustment or accommodation request is needed to get there.

It is also a better instrument for most of what an early-stage interview is trying to find out. Written reasoning can be read twice and checked, so every score carries the reasoning behind it. Companies read the transcript with identifying details removed, never a recording of how it looked to answer.

No personally identifying information is sent to any language model. Not the talent’s name, not their email, not their employment history, not anything that could be used to work out who they are. The model sees the role, the track, the talent’s experience level, declared skills and industries, the conversation and the scoring criteria.

This is worth stating precisely because it is the kind of claim vendors make loosely. It is enforced at the architecture level: the interview service composes prompts from a bounded, non-identifying set of fields, and the boundary is one of the platform’s standing invariants rather than an option a tenant can switch off.

It matters for a reason beyond privacy. A model that cannot see who is answering cannot condition its scoring on who is answering. Removing identity from the model’s input is the same protection that anonymous screening gives a human reviewer, applied to the component that scales fastest.