There's a quiet tension at the heart of every job interview. Too loose, and it becomes a pleasant chat that predicts almost nothing about how someone will actually perform. Too rigid, and it turns into a stilted form-filling exercise that can't follow an interesting answer anywhere. The semi-structured interview has always promised a way to split the difference — and as AI starts conducting interviews at scale, that old idea is suddenly the most important one in hiring. Here's what the research actually says about it.

Structure is a dial, not a switch

The single most useful finding in the interview-research literature is that structure isn't binary. It's a continuum, and an interview's ability to predict job performance climbs steadily as you add structure to it.

When researchers have coded large numbers of interview studies by their degree of structure, the pattern is consistent: the least-structured interviews predict performance poorly, while the most-structured ones predict it far better — in several analyses, roughly twice as well. The foundational meta-analyses in personnel psychology reach the same conclusion from different angles: structured interviews have higher validity than unstructured ones, and situational interviews tend to out-predict more general formats.

This is exactly why "semi-structured" is so often misunderstood. People hear it as "structured, but watered down." The research says otherwise. An interview stays high on the structure continuum as long as its core elements hold — the same job-relevant questions, asked of everyone, scored against consistent criteria. What semi-structured adds on top of that is room to probe. And critically, the literature treats pre-planned follow-up questions as part of a structured interview, not a violation of it. Structure and follow-ups aren't opposites; the best interviews have both.

Situational or behavioural? The question type matters

Within a structured interview, two question types dominate the research, and they aren't interchangeable.

Situational questions ask what a candidate would do in a hypothetical scenario. They rest on the premise that intentions predict behaviour, and they've shown solid, repeated predictive validity across many studies. Behavioural questions ask what a candidate did do in the past, on the logic that past behaviour forecasts future behaviour.

The nuance worth designing around: situational questions work well for clearly-defined and more junior roles, but they lose predictive power for higher-level positions, where behaviour-description questions perform better. There's also a subtler finding from recent meta-analytic work — interviewers asking different question types to assess the same underlying trait can reach genuinely different conclusions about a candidate, which means a tailored, construct-specific design improves prediction. The lesson isn't "pick one." It's that question type should be matched deliberately to the role and the specific competency you're trying to measure, rather than pulled from a generic bank.

The probing problem

Here's where semi-structured interviews get genuinely hard, and where honesty matters.

Probing is the whole point of leaving room for follow-ups — and done well, it's valuable. When follow-up prompts are asked in a consistent manner, they can effectively reveal deeper aspects of a candidate's thinking that a fixed question never reaches.

But unscripted probing is also one of the best-documented ways bias creeps back in. Even within an otherwise structured interview, interviewers can unintentionally introduce bias through tone, manner, or follow-up probes. The defining weakness of the semi-structured format is that an interviewer's beliefs can shape which questions get asked and how answers are interpreted, making responses harder to compare systematically. The freedom that makes probing powerful is the same freedom that lets inconsistency — and bias — slip in.

So the design challenge is clear: keep the depth that adaptive probing provides, without surrendering the consistency that makes structure work in the first place.

What changes when AI does the interviewing

This is where AI genuinely shifts the picture — and where two serious 2025 studies do the talking.

In a large randomised experiment from Stanford and USC, applicants for a developer role were routed either through traditional resume screening or through an AI structured interview, with final human interviewers blind to which path each candidate took. Candidates who went through the AI interview were substantially more likely to pass that final human interview — the AI was surfacing genuinely stronger people, not merely different ones. The same study found something a resume simply can't catch: a meaningful share of applicants had asserted proficiency in a required technology that the interview revealed they lacked.

A separate large experiment from the University of Chicago Booth looked at the conversations themselves and found that the AI's adaptive questioning drew out more of the behaviours that actually correlate with strong hiring outcomes — richer, more interactive responses — while reducing unhelpful noise. Blind reviewers also rated the AI-led interviews as higher in question relevance and overall quality than the human-led ones.

What makes this possible is that an AI can do something humans struggle with: adapt and stay consistent at the same time. A human interviewer who improvises follow-ups drifts — different candidates get different probes, scored against shifting impressions. A well-built AI can ask a tailored follow-up while still applying the identical scoring rubric to every candidate. It can be flexible on the surface and rigid underneath.

But there's a specific trap the research flags. A model can learn to key its scoring off the pattern of follow-ups it chose rather than what the candidate actually said. The design principle that follows is non-negotiable: the AI may adapt its questions for depth, but scoring must be anchored to the substance of the candidate's responses, against a fixed rubric.

Why the structure pays off in fairness

The reason all this structure matters isn't just accuracy — it's equity, and the mechanism is mechanical rather than aspirational.

When every candidate receives the same questions and identical scoring rules apply, there's no interviewer forming a gut impression, no rapport effect, and no route for background similarity to sway the score. That's not a hope; it shows up in the outcomes. Structured interviews have been found to substantially reduce the adverse impact of unstructured ones. Higher structure consistently brings higher validity, better rater agreement, and less adverse impact all at once — the fairness and the accuracy come from the same source.

There's a practical dividend too. Consistent, job-related, documented questioning is far easier to defend as fair — which is no longer optional in a world where recruitment AI faces real regulatory scrutiny.

The takeaway

The semi-structured interview was always a good idea trapped in a hard trade-off: humans couldn't probe freely and stay consistent. AI dissolves that trade-off. It can hold the structure that makes interviews predict performance and reduce bias, while adding the adaptive depth that fixed scripts never allowed — as long as it scores what the candidate says, not the path it chose to get there. Done that way, a semi-structured AI interview isn't a novelty. It's the most rigorous version of the format we've ever been able to run at scale.