TL;DR: Structured vs unstructured interview: a structured interview asks every candidate the same job-related questions in the same order and scores answers against set criteria; an unstructured one does not. In the most recent large meta-analysis (Sackett et al., 2022), structured interviews had an operational validity of .42 for job performance versus .19 for unstructured, in studies of human-rated interviews.
Updated 28 October 2026.
Key takeaways
- Structured interviews had the highest operational validity of the common selection methods in Sackett et al. (2022): .42 (ρ), versus .19 for unstructured interviews, which is more than double (Sackett et al., 2022).
- The 1998 Schmidt and Hunter values (.51 structured, .38 unstructured) were revised downward because earlier meta-analyses over-corrected for range restriction. Per the abstract, most of the selection procedures that ranked high in prior summaries remain high in rank, but with mean validity estimates reduced by .10 to .20 points (Sackett et al., 2022).
- The ranking is robust to the authors' sensitivity cases; the levels are not fixed, and the spread is large. The 80% credibility lower bound for structured interviews is .18.
- The Black-White standardized difference is .23 for structured interviews versus .32 for unstructured. That is a data point, not evidence that structured interviews are unbiased.
- These estimates are for human-rated interviews. We found no meta-analytic validity estimate for AI-scored interviews.
Search for structured vs unstructured interviews and in our reading the results split in two: qualitative-research methods on one side, hiring on the other. This page is about hiring, and about one question: which format predicts later job performance, and by how much? The 1998 figures (.51 structured, .38 unstructured) were revised in 2022 to .42 and .19, and the gap between the two widened.
You get both sets of numbers side by side below, with the full ten-row table, how the interview figures were built, and the caveats that go with them. Structure can also be delivered and scored at scale through AI-delivered structured interviews; the research below doesn't cover those, and the AI section explains why that matters.
What is the difference between structured and unstructured job interviews?
Structured interview. A structured interview asks every candidate the same job-related questions, in the same order, and scores each answer against predefined criteria, usually on anchored rating scales. The aim is that differences in scores reflect differences in answers, not differences in the conversation.
Unstructured interview. An unstructured interview has no fixed question set and no scoring key. The interviewer chooses what to ask, follows the conversation where it goes, and forms an overall impression.
The structured definition used here, which in our reading matches the common vendor definitions: same questions, same order, anchored scoring for every candidate.
This page covers selection interviews for hiring. Qualitative research interviews, sociological methods and clinical interviews such as the SCID are out of scope.
What is a semi-structured interview?
A semi-structured interview keeps a fixed core of questions and allows bounded probes. In my framing it sits between the two ends of a continuum of structure. That continuum is mine, not a research taxonomy, and the studies below measure only the structured and unstructured ends.
Which is more valid, structured or unstructured?
Structured, by a wide margin on the revised numbers. The source is Sackett, Zhang, Berry and Lievens (2022), "Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range," Journal of Applied Psychology 107(11), 2040-2068 (DOI 10.1037/apl0000994).
- Structured interviews: operational validity ρ = .42 (it was .51 in 1998). SD of ρ = .19. The lower bound of the 80% credibility interval is .18. Black-White standardized difference d = .23.
- Unstructured interviews: ρ = .19 (it was .38). SD = .16. The lower bound is -.01. d = .32.
Structured is more than double unstructured on the revised numbers (.42 versus .19), and the gap widened from .13 to .23.
"Operational validity" is the correlation between an interview rating and later job performance, corrected for criterion unreliability. It is not an accuracy percentage. A validity of .42 does not mean the interview picks the right candidate 42% of the time. Whenever "twice as effective" appears in print, it needs the .42 and .19 behind it and the source.
The same paper revised general mental ability (GMA, or cognitive ability tests) from .51 to .31 (d = .79). Read that as a revised estimate after the authors removed a range-restriction correction they judged to be an overcorrection. It does not mean cognitive ability stopped mattering.
The size of the spread matters as much as the mean. A lower bound of .18 means a well-built structured interview can still perform poorly in some settings. These are averages across studies, not promises about your hiring process.
The full validity table: 1998 vs 2022
Rows are sorted by the 2022 estimate, highest first. The 1998 column is Schmidt and Hunter's, as reproduced in the Sackett et al. comparison (Schmidt & Hunter, 1998, DOI 10.1037/0033-2909.124.2.262).
| Predictor | Schmidt & Hunter 1998 | Sackett et al. 2022 ρ | Change | SD of ρ | 80% credibility lower bound | Black-White d |
|---|---|---|---|---|---|---|
| Employment interview, structured | .51 | .42 | -.09 | .19 | .18 | .23 |
| Job knowledge tests | .48 | .40 | -.08 | .13 | .23 | .54 |
| Biodata, empirically keyed | .35 | .38 | +.03 | .09 | .26 | .33 |
| Work sample tests | .54 | .33 | -.21 | .09 | .21 | .67 |
| Cognitive ability (GMA) tests | .51 | .31 | -.20 | .14 | .13 | .79 |
| Integrity tests | .41 | .31 | -.10 | .20 | .05 | .10 |
| Assessment centers | .37 | .29 | -.08 | .09 | .17 | .52 |
| Employment interview, unstructured | .38 | .19 | -.19 | .16 | -.01 | .32 |
| Job experience (years) | .18 | .07 | -.11 | .11 | -.07 | .49 |
Source: Table 3 of Sackett et al. (2022).
Five notes go with the table:
- Validities are for predicting job performance, not training or tenure.
- The Black-White d column is the only subgroup contrast in the table. Some d values come from non-applicant or mixed samples, and the authors say the picture is incomplete.
- The 2022 column mostly uses a criterion reliability of .60.
- These are meta-analytic means with large spread, not guarantees for a given employer.
- The studies are of human-rated interviews, not AI-scored ones.
What did Sackett et al. (2022) change, and is the 1998 table still right?
The levels changed. Structured interviews stayed among the strongest predictors and moved to the top.
Schmidt and Hunter (1998) summarized 85 years of findings across 19 selection procedures (Psychological Bulletin 124(2), 262-274). They reported work samples at .54, GMA at .51, structured interviews at .51 and unstructured interviews at .38. The top combinations were GMA plus a work sample (.63), GMA plus an integrity test (.65) and GMA plus a structured interview (.63).
McDaniel et al. (1994) had found, across all 245 coefficients, fully corrected validities of .44 for structured and .33 for unstructured interviews (figures as cited in Sackett et al.).
Sackett et al. rebuilt the interview number in three moves. They dropped range-restriction corrections they judged unsupported, corrected for criterion unreliability at .60, and combined the two meta-analyses by sample size. The result was .42 for structured and .19 for unstructured.
The interview figures were built from two earlier meta-analyses, McDaniel et al. (1994, DOI 10.1037/0021-9010.79.4.599) and Huffcutt et al. (2014, DOI 10.1111/ijsa.12078). Table 2 of the paper lists their study counts, sample sizes and observed correlations. Schmidt and Hunter had drawn on only a subset of the McDaniel studies, so the two sets of estimates do not rest on identical data.
Sackett et al. also tested the result by varying the assumptions, with a more generous criterion reliability and the authors' maximum plausible range-restriction correction. Structured interviews stayed on top in every variant.
The paper is contested. The University of Minnesota's research portal lists a reply by Sackett and colleagues to a commentary by Oh et al. on correcting for range restriction (Minnesota research portal). What can be said now: every correction choice moves the levels, and none of the sensitivity cases in the paper displaces structured interviews from the top.
Why are unstructured interviews less valid, and what makes an interview structured?
The research compares formats, so the useful question is what separates them in practice. Structure means the same job-related questions, anchored rating scales, scoring before an overall judgement, and limited off-script probing. An unstructured interview leaves each of those to the interviewer on the day.
Use the checklist below to see how structured your own process is. It is our synthesis of common definitions, not a published instrument.
Degree-of-structure checklist (answer yes or no):
- Every candidate gets the same questions.
- The questions come in the same order.
- The question content is based on a job analysis.
- Answers are scored on anchored rating scales.
- Each question is scored before any overall judgement is formed.
- Multiple raters score, or one rater applies the same standard throughout.
- Probing off-script is limited to a defined set.
The more items you can tick, the closer your process sits to the structured end. If item 4 is a no, structured interview questions and rubrics by role covers the question text and the scoring anchors.
Are structured interviews less biased?
The data gives a narrower answer than the claim usually does. In Sackett et al. (2022), the Black-White standardized difference is d = .23 for structured interviews and d = .32 for unstructured ones (Sackett et al., 2022).
Treat that with care. It is the only subgroup contrast in the table, some values come from non-applicant or mixed samples, and the authors say the picture is incomplete. This page does not claim structured interviews are unbiased. Evidence on interviewer bias beyond these numbers was not reviewed here, so none is offered.
What to do with the number: don't borrow it. Adverse impact is assessed on your own selection data, and the standard check is explained in four-fifths rule and impact ratios.
When is an unstructured or semi-structured interview still the right call?
Sometimes. What follows is practitioner judgement, not a research finding.
Unstructured or semi-structured conversations suit rapport and role-clarity discussions, where the point is for both sides to learn about the work. They suit senior or novel roles, where a fixed script can't capture what the job demands. And they work well as a second stage after a structured screen has narrowed the field. The problem comes when they are the only stage.
The overlap with skills evidence is worth using. If you've defined the skills a role requires, the next step is turning each requirement into a scored interview question, which is the approach in skills-based hiring with structured interviews.
Can you run a structured interview with an AI interviewer?
In design, yes. Structure is the active ingredient in the research, and an AI interviewer that delivers the same template questions and scores each answer against a stated key is one way to get structure at scale. The research supports structure. It does not support any particular AI scorer.
Sackett et al.'s estimates are for human-rated structured interviews. AI-scored interviews need their own validity and adverse-impact evidence. In SIOP's TIP, a piece on the Sackett et al. paper concludes that structured interviews "may in fact be the strongest predictor of job performance", which is a claim about human-rated formats (SIOP TIP). We found no meta-analytic validity estimate for AI-scored interviews.
That gap is the argument for explainable AI in hiring: until validity evidence exists, each score has to carry the answer, the key and the reasoning, so a human can check it.
Candidates are already meeting these interviews. Criteria Corp's annual candidate experience report (n=2,500+, 10 March 2026) found 68% of job seekers are willing to forgo a résumé for assessments and structured interviews (Criteria Corp; vendor survey). Greenhouse (n=2,950, 1 May 2026) found 63% have faced an AI interview and 70% were never clearly told AI would evaluate them (Greenhouse; vendor survey). Both sets of figures sit in the candidate survey data on AI interviews. Several vendors now market structured or skills-based AI interviews. Ask each one whether the core questions are fixed for every candidate, whether scoring is anchored to a written key, and whether the reasoning behind each score is visible.
ZeaHire is one such system. Interview templates hold questions and instructions, and can be AI-generated or written by hand. Each answer is scored with the AI's reasoning visible. Five dimension ratings come with written evidence. Every ZeaHire score comes with the answer and the reasoning for a human to check. No validation study of ZeaHire's scores exists, and none is claimed here.
If you're evaluating tools on this basis, what to look for in AI interview software sets out the checks.
Frequently asked questions
Which is more valid, structured or unstructured interviews?
Structured interviews, at .42 versus .19 in the most recent large meta-analysis (Sackett et al., 2022). That is more than double on the revised numbers. The estimates are for human-rated interviews.
How much more predictive is a structured interview?
On the 2022 estimates, .42 against .19 is more than double. That ratio compares two validity coefficients. It is not a percentage of hires predicted correctly, and "twice as effective" should always appear with the .42, the .19 and the source.
What did Sackett et al. 2022 change?
It corrected an overcorrection for range restriction in earlier meta-analyses. For most of the procedures that ranked high before, mean validity estimates fell by .10 to .20 points, and structured interviews emerged as the top-ranked selection procedure, ahead of work samples, which led in 1998.
Is the 1998 Schmidt and Hunter table still right?
Not for the levels. The 1998 values (.51 structured, .38 unstructured) were revised to .42 and .19. The ranking at the top, with structured interviews among the strongest predictors, held.
Why are unstructured interviews less valid?
The research compares formats and measures the result, not the mechanism. What separates the formats is whether questions, order, scoring anchors and probing are fixed in advance or left to the interviewer on the day.
Are structured interviews less biased?
The Black-White standardized difference is d = .23 for structured interviews and d = .32 for unstructured. The authors say the subgroup picture is incomplete, so this page does not claim structured interviews are unbiased. Adverse impact should be checked on your own selection data.
Can you do a structured interview with an AI interviewer?
In design, yes: an AI interviewer can deliver the same template questions and score each answer against a stated key. The validity evidence, however, is for human-rated interviews. We found no meta-analytic validity estimate for AI-scored interviews.
What is a semi-structured interview?
A semi-structured interview keeps a fixed core of questions and allows bounded probes. It sits between the structured and unstructured ends of a continuum of structure, and the research above measures only the two ends.
See structured scoring in practice
See structured scoring in a demo. Each answer is scored against a stated key, with the reasoning shown for a human to check.
Comments
No comments yet — start the conversation.