The ‘AI interview' doesn’t exist.

Well it kind of does, but the key point is, ‘The AI interview’ isn’t one thing. What we actually have is a slew of hiring products hitting the market, all with different features, that make different uses of different types of AI to achieve different outcomes. Each one brings different, and important, trade offs.

📽️ Modality - text, voice or video. These products can be 100% text based, with candidates typing their answers, offer an audio recording option or be video based. Different modalities offer different trade offs. Text strips out the most construct-irrelevant variance from delivery. No accent, appearance, lighting, or fluency-under-camera-pressure contaminating the signal. Async and untimed lowers anxiety for most people. But, it's arguably more susceptible to AI assistance, feels less modern and could penalise those with poorer writing skills.

↔️ Interaction direction (one-way async vs two-way conversational) - Trade offs are clear here. One way is more structured and defensible from a fairness standpoint, but arguably is less of an interview and more of a structured response test. Lack of probes in one-way interaction may disadvantage some candidates who didn’t fully understand the initial question.

❓ Question logic - static or dynamic. Static questions stay defensible because identical content keeps your adverse-impact or validation analysis clean, whereas dynamic trades that away for greater depth on key areas.

💾 What the AI is doing - Interviewer, rater or both. AI as an interviewer strips out a load of human-to-human variance in the quality and consistency of the interview but not all candidates enjoy the experience. AI as a rater can potentially remove various human biases and cognitive overwhelm issues, if done well. The two in combination can be powerful, but only if you’ve put the work in: stack an AI interviewer on top of an AI rater without validating both and good luck disentangling which one actually moved the score. At the very least,  you’ll have a pile of interaction effects to investigate, at this point I recommend you call Ben Hawkes, this is one of his nerd-out areas.

🤐 What content is analysed. There is no defensible trade off here. The only content analysed should be the words said, or spoken. We don’t have the evidence for tone, facial expressions or other para data.

Got questions? Want some clear, evidence backed advice on choosing an AI interview tool? You know who to call. Hint, it’s not Ghostbusters.