Mercor AI Interview: What It Records and Who Sees It

A Mercor screening interview is conducted by an AI interviewer, not a person, and Mercor's own talent documentation says there is no alternative: "Our AI interview is the standard and only method for screening candidates," with no chat, call, or other medium available at that stage (Mercor talent docs, AI Interview). It runs about 20 minutes, it records video and audio, it produces an AI-generated transcript, and you get up to three attempts.

That combination — mandatory, recorded, machine-scored, and non-retrievable — is unusual enough to be worth reading carefully before you click start. This guide covers what Mercor captures, who can see it, how the retake rule actually works, what its AI-use policy permits, and what its March 2026 security incident means for a recording you cannot download. Facts are current as of 10 September 2026 and drawn from Mercor's published documentation, its own incident disclosure, and news reporting.

What the Mercor AI interview is

Mercor is a talent marketplace that matches people — often for AI training, evaluation, and expert data work — with client projects. Its documentation describes the AI interviewer as a system that will "pose role-specific questions, generate a transcript of your responses, and evaluate your performance for the listing" (Mercor talent docs).

Two structural details matter more than the question content:

• The result is reused. Once you complete an interview, that result carries across other applications that require the same interview. You are not re-interviewed per listing. Mercor's release notes on the revised application flow confirm completed steps carry over automatically to new applications (Mercor talent docs, new application steps).
• There is a separate practice interview. Mercor's documentation states the practice interview is free, is not tied to any application, can be taken as many times as you like, and is not shared with companies.

The practice surface is the single most useful thing in the docs, and it is easy to miss. If a recorded, machine-scored 20-minute interview is going to be reused across every application you make, rehearsing on a surface that no employer sees is the rational first step.

What Mercor captures

Source: Mercor talent docs, AI and data usage.

The screen-share line deserves attention because it changes what "an AI interview" means technically. A one-way video interview captures a webcam feed. An interview that asks for a full-screen share captures your entire display, including notifications and whatever else is open on it. If you are used to reasoning about what a live human interviewer can observe during a technical interview screen share, the same reasoning applies here, except the observer is a recording that persists.

Mercor's technical requirements are narrow: a computer with a working camera and microphone, using Chrome, Edge, or Safari. Firefox and mobile browsers are not supported. Test the devices in the waiting room rather than at the first question.

The retake rule, precisely

Mercor documents three attempts: "Candidates are entitled to retake an interview up to three times across all applications related to that particular interview" (Mercor talent docs). The reset is self-serve — open the Assessments tab on your dashboard, find the interview showing as Submitted, use the three-dot menu, and select Retake interview.

Two limits follow from that wording. The three attempts are pooled across every application tied to that interview, not granted per application. And once they are used, further resets require contacting support; Mercor's release notes say an additional retake "isn't guaranteed and is only meant for technical or troubleshooting issues" (Mercor talent docs).

One claim circulating in third-party Mercor guides is that only your most recent attempt counts. Mercor's published documentation, as of 10 September 2026, does not state which attempt is used for evaluation. Treat that as unknown rather than settled, and plan on the assumption that a weak attempt may not be erased by a stronger one.

What you cannot get back

You cannot obtain your own recording or transcript. Mercor's documentation says recordings and transcripts are stored internally and used for evaluation, and that it is not able to share them with candidates.

That is a meaningful asymmetry. The system that decides whether you advance produces an artifact about you that you are not permitted to read. Mercor does offer account deletion through Account Settings, with the caveat that it may retain data where law requires and that deletion "does not remove data from companies that may have already accessed it."

If being screened by software without a human alternative is the part you object to, the request options are narrower than most candidates assume — disclosure, deletion, and disability accommodation carry more weight than a request to change the interview format. We covered what those requests are worth, and what withdrawing costs, in what you can actually request before an AI-led interview. For Mercor specifically, the documentation forecloses the format request outright.

What the AI-use policy actually permits

Mercor publishes a separate LLM Usage Policy. Read as written, it draws the line at authorship rather than at tool use:

• Permitted: checking grammar and diction, and refining phrasing or tone, provided the underlying ideas and structure remain yours.
• Prohibited: using a large language model to compose your responses, to rate model responses, to generate rationales or explanations, or to predict code behaviour or runtime results.

That is stricter than it first appears, because rating model outputs and writing rationales are the actual work in many Mercor projects. The prohibition is not only about the interview; it covers the substance of the task.

The policy also asserts that AI-produced submissions "will be identified and flagged accordingly," and reasons that human and model writing are generally distinguishable. Treat that as a vendor claim about intent, not as a verified detection capability: Mercor publishes no methodology, no accuracy figures, and no appeal process for a flag. Detection claims in this category are frequently overstated, in both directions — a point worth keeping in mind whenever a platform asserts it can tell, as we discussed in what AI-assisted coding interviews actually grade.

Separately, Mercor's Account Security and Conduct policy requires that "Each expert is expected to operate one account, in their own name, and to perform the work themselves," and lists account sharing, operating multiple accounts, and impersonation as reviewed conduct. The stated consequences escalate from reduced or reversed payment for the affected work, to removal from a project, to suspension of platform access. Sharing interview questions or screenshots is also prohibited.

There is context for that severity. In April 2026, Forbes reported that Mercor's founders were contending with employee fraud and North Korean infiltration (Forbes, 15 April 2026). A marketplace paying people worldwide for remote AI work is a natural target for identity fraud, which is why its identity rules read closer to a payroll policy than to a typical interview waiver.

The March 2026 security incident

If your recording cannot be retrieved by you, it is fair to ask how well it is held. Mercor disclosed that in late March 2026 it was affected by a supply-chain attack on LiteLLM, an open-source tool, in which "a malicious actor published compromised versions of LiteLLM designed to exfiltrate credentials from any system that installed them." Its update, published 25 June 2026, said only a very limited subset of its nearly five million experts had sensitive information exposed, that no employees were affected, and that notifications went out on 25–26 June with an offer of identity-protection services. Remediation included dependency auditing, credential rotation, independent penetration testing, and managed detection and response (Mercor, update on security incident).

Mercor's disclosure does not state that interview video or transcripts were among the exposed categories. Class-action suits filed on 23 April 2026 allege more: that exposed data included job interview recordings, facial biometric data, and contractor screenshots, and that recorded interviews were used to train AI. Mercor said it "strongly dispute[s] the speculative claims in these lawsuits" (PYMNTS, April 2026).

Those are allegations, not findings, and they sit against Mercor's own documented position that it does not use interview data to train external AI models and does not sell interview or résumé data — while acknowledging it does train internal recommendation models and that human contractors sometimes assist with assessment work. The honest summary for a candidate is that the training question is contested in active litigation and unresolved as of 10 September 2026.

What the documentation does not tell you

• No published rubric. Mercor says its AI evaluation aims for consistent, objective standards, but publishes no competency list, scoring bands, or pass threshold. You cannot prepare against the criteria because the criteria are not public.
• No stated scoring split. How much of the outcome comes from the model versus a human reviewer is not documented.
• Timeline. Mercor's documentation puts the wait for a next-stage decision at roughly two to four weeks.
• No transcript for appeal. With no access to the recording or transcript, there is no practical basis for contesting an evaluation.

What to do before you start

The preparation that actually pays here is unglamorous. Use the practice interview repeatedly, since it is free, unlimited, and invisible to employers. Fix the environment before the official attempt: supported browser, tested microphone and camera, a quiet room, and a display you are willing to share in full if the interview asks for it. Assume all three attempts may count, and treat the first as real.

On tooling, be precise about what Mercor's policy allows. Control is a desktop AI interview assistant for interviews, assessments, and screen-share workflows on Windows and macOS — but Mercor's LLM Usage Policy prohibits using a language model to compose your interview responses, and its conduct policy requires that the work be yours. The defensible use of any assistant in this context is rehearsal beforehand, not authorship during. If you want a structured way to run that rehearsal, our guide to preparing for an AI interview sets out a practice loop and a scoring rubric you can apply to your own recorded answers.