Recruiters given micro1’s AI interview reports selected finalists who were more likely to pass a later, blinded human interview. An August 2026 preprint revision reports a 17.5-percentage-point improvement among 140 finalists in one experiment. The system asks adaptive follow-ups and assesses transcripts and coding tasks; the experiment tested whether its reports improved selection.[1]
Ask for evidence of the relevant skill
Structured interviewing has a purpose that predates generative AI. The US Office of Personnel Management describes questions tied to job competencies, with answers evaluated against defined standards. A highly structured interview fixes both the main questions and the follow-ups. This helps interviewers compare evidence about the same capabilities.[2]
Two question types are particularly useful: asking what somebody did in a previous situation, or how they would approach a hypothetical problem. A claim on a CV can then lead to an explanation of a decision, a sequence of actions or the reasoning behind a result.[2]
An illustrative interview about handling customer problems
| Starting claim | Useful follow-up | Evidence to listen for |
|---|---|---|
| Resolved difficult complaints | Describe one case from first contact to resolution. | Actions taken and the reason for them.Source 2 |
| Improved response times | How did you measure the change? | A defined period and a comparable baseline.Source 2 |
| Worked across teams | What did another team need from you? | Specific information exchanged and responsibilities.Source 2 |
What the experiment could isolate
All 1,108 candidates in the first experiment had completed AI interviews before assignment. Recruiters received either CV-based information alone or that information plus the report. Final human interviewers did not know the assignment.[1]
Counting non-attendance as failure, the estimated improvement was 17.5 percentage points from a 28.6% control mean, with an 8.5-point standard error. The comparison is between the finalists each recruitment process selected. It does not mean that taking an AI interview improved an individual candidate’s skills: everyone in this experiment had already completed one.[1]
The study concerns junior technical recruitment, not subsequent work performance; two authors have micro1 affiliations. In the separate pipeline experiment, about 75% of invitees did not complete the 30–40-minute AI interview. Its unadjusted result among 70 finalists was less conclusive, counting non-attendance as failure (permutation p=0.149).[1]
Give candidates room to demonstrate what they know
The distinction matters for a candidate with a short employment history. A CV might supply little more than a skill name and a course. A well-chosen follow-up can ask them to explain an actual problem, making their reasoning available for review. The response must still be judged against the job’s requirements.
OPM also identifies interviewer time and the number of applicants as factors in administering structured interviews. Moving such a conversation earlier in recruitment changes where that effort occurs: more people can be asked for evidence before a shortlist is settled.[2]
For an applicant, that opportunity is valuable when the questions allow a relevant strength to emerge. For the employer, the useful report connects a judgement to the answer that supports it. Keeping that connection visible lets the next interviewer pursue a promising skill or investigate something still unclear.
Sources & further reading
- Quantifying the Benefits of AI-Assisted Recruitment, version 2 Aka and colleagues — arXiv preprint · 7 Aug 2026 · Accessed 12 Oct 2026
- Structured Interviews US Office of Personnel Management · Accessed 12 Oct 2026