An AI interviewer is a software agent that asks hiring questions and scores the answers, usually as a one-way video or a voice call. A job simulation is a role-specific task where the candidate produces work you can score against a rubric. They are not the same screen. One captures how someone talks about the job. The other captures whether they can do a slice of it.
Amazon made Amazon Connect Talent generally available on 17 September 2026: AI agents run structured voice interviews and competency assessments, and recruiters keep the final call. University of Georgia researchers, writing in Information Systems Research, found a different risk. Across eight studies, job seekers who knew an AI agent would score a one-way video interview showed more exaggeration than people a human evaluated. Video analysis confirmed the self-reports. The AI rater did not penalize stretched answers; human raters generally did. Telling candidates how they would be evaluated brought exaggeration back toward human-interview levels. Transparency helped authenticity. It did not make the AI score a measure of job performance. Roth, Bobko, and McFarland (Personnel Psychology, 2005) estimated work-sample validity near 0.26 observed and about 0.33 corrected — tied to the work, not the story about the work.
This guide is for lean talent teams deciding what to buy next. It separates conversational AI screens from role-task evidence, and shows where Dotof’s AI Role-Simulation Assessments sit: prove the work first, then spend live conversation on what the task could not answer.
What is an AI interviewer?
An AI interviewer conducts a question-and-answer screen without a recruiter in the room, then rates the transcript, audio, or video. In 2026 the common forms are asynchronous one-way video and voice agents that adapt the next question. Amazon Connect Talent, generally available from 17 September 2026 in AWS US East (N. Virginia) and US West (Oregon), is a current volume example: candidates join a structured voice interview at any hour, the system also administers competency assessments, and a person reviews scores, transcripts, and notes before a hiring decision. That design solves scheduling. The core signal is still language about the work, plus whatever items the assessment adds.
Treat the label carefully. A vendor can call a scored conversation an “assessment.” If the candidate never produces the job’s output — a reply, a plan, a prioritization, a piece of analysis — you bought an interviewer, not a simulation.
What is a job simulation?
A job simulation, also called a work sample or role simulation, places the candidate inside a controlled version of a real task. A support lead might triage three tickets and write the customer reply. A salesperson might rank a pipeline and draft the next email. Scoring uses a rubric tied to the job, not an inference from tone of voice. Every applicant faces the same situation, and the output is something a hiring manager can read. That is why simulations fit skills-based hiring better than a résumé screen or a charm-heavy first call. For the sequencing against a live human interview, see job simulation vs interview.
Why do candidates exaggerate in AI interviews?
Akshat Lakhiwal, Che-Wei Liu, Hillol Bala, and Hung-Yue Suen studied this in “From Opacity to Transparency: User Behavior and Downstream Effects in Algorithmic Evaluation” (Information Systems Research, 2026; DOI 10.1287/isre.2023.0579). Relative to human evaluation, algorithmic scoring of asynchronous video interviews increased stress and deceptive impression management and reduced honest self-presentation. The AI rater did not reliably separate honest answers from embellished ones. Human evaluators looking at the same material generally penalized the less authentic behavior.
Explaining the criteria — facial expressions, sentiment, keywords, teamwork, skills, work style, personality — offset the exaggeration and restored behavior toward human-evaluation levels. The authors also found a second trap. When later evaluators were shown the AI score, they anchored on it and discounted their own judgment, even though the score did not distinguish honesty from deception. Transparency can make candidates more authentic. It does not, by itself, repair a weak predictor. If you run an AI interview, say what is scored, what is not, and that a person makes the decision. Do not treat the number as proof of job performance.
What predicts performance: talk or a work sample?
Work samples rank ahead of unstructured interviews because the predictor is a sample of the job, not a story about the job. The Roth, Bobko, and McFarland 2005 estimates — about 0.26 observed, about 0.33 corrected — are the figures to use. The older 0.54 number still appears in sales decks and overstates the evidence. Interviews still matter once capability is known: motivation, judgment, and how someone will work with this team are hard to read from a task alone. An AI interviewer automates the conversation. It does not convert that conversation into a work sample.
If the shortlist is a ranked transcript, you have scaled a weak early signal. If the shortlist is ranked on a role task, the next conversation — human, or an AI screen for high-volume roles — can ask what the task left open. Pair that with integrity controls. A simulation score is only as trustworthy as the conditions it was taken under; see how to prevent AI cheating in hiring assessments.
When should a lean team use each?
- AI interviewer — scheduling is the bottleneck, questions are structured, and a human still decides. Use it to collect consistent answers at odd hours, not to declare who can do the job.
- Job simulation — the role has observable day-one work, a wrong hire is expensive, and you need comparable evidence before panel time. Generate the task from the live job description so it is not a generic quiz. The steps are in from job description to role simulation.
- Both, in this order, for most professional roles — simulation first to prove capability, then a short human interview for judgment and motivation. An AI voice screen can sit beside that sequence for surge hiring. It should not replace the task.
Do not ask the simulation and the interview the same questions. Rewrite the conversation around what the work sample could not show. Keep a written override: no model score, interviewer or simulation, should be the only vote. The practical gates are in the responsible AI hiring checklist.
How does Dotof fit this decision?
Dotof AI Role-Simulation Assessments start from the job description, generate a realistic task, run it in a supervised browser, and return a structured report hiring managers can compare. The point is evidence of work, not a score on how fluently someone described their past. Use an AI interviewer if you need after-hours question coverage. Use the simulation when the decision you cannot afford to get wrong is “can this person do the job.”