Your Interview Recording Can Be Scanned Months Later

Most interview integrity tooling used to be something you could see happening: a lockdown browser, a webcam permission prompt, a bot that joined the call with a name like "Notetaker." A newer category does none of that. It runs on the recording after the interview is over, works with files from any meeting tool, and can be pointed at a conversation that happened weeks or months ago. If an employer recorded your interview and still has the file, that file can be re-examined later without anyone telling you in advance.

This article explains what that class of tool reads, what it cannot see, how much weight its findings deserve, and what it changes — and does not change — for candidates.

What changed in September 2026

On September 24, 2026, HeyMilo announced HeyMilo Red, spinning its interview cheat-detection technology out of its recruiting platform into a standalone product. The significant detail is not the detection list but the deployment model. Red's own product page states it works "on its own, with recordings from any interview tool," and that a customer can "Upload a file, paste a link, or call the API." There is no requirement to migrate applicant tracking systems, install a proctoring browser, or put a bot in the meeting.

That is a structural change. Previous vendor tooling in this space generally had to be present at interview time. Tofu, for example, describes its deepfake detection as "Real-time detection in live interviews," running "silently in the background during interviews" — a capability that only exists if the employer deployed it before the call started. A scanner that accepts arbitrary recordings has no such constraint. Any interview an employer already recorded becomes eligible input, including interviews conducted before the employer bought the tool.

HeyMilo says it is developing "Proctor Bench," a benchmark intended to measure cheat-detection accuracy and false-positive rates. As of this writing that benchmark has not been published, so the accuracy of the product is a vendor claim rather than an independently verified result.

A post-hoc scan reads the recording, not your computer

This distinction matters more than any individual detection feature, and it is where a lot of candidate reasoning goes wrong.

Live proctoring systems inspect the machine. They ask the browser what it can see, watch for focus changes and tab switches, enumerate displays, and sometimes require a desktop application that restricts what else can run. Those are questions about your computer's state, and they are answerable only while the assessment is running. Our guide to what an in-meeting AI proctor can observe covers that model in detail.

A recording scanner asks a completely different question. It has one video track, one audio track, and a transcript. Everything it concludes has to be inferred from what the camera and microphone captured: where your eyes went, how your voice sounded, when you paused, whether a second voice is audible, whether your phrasing resembles generated text.

The practical consequence is that the entire category of "does this tool show up in a screen share or a capture" is irrelevant to this kind of analysis. It is not looking at your screen. It is looking at your face and listening to the room. A tool that stays out of screen capture neither helps nor hurts against gaze-and-voice inference, because the two operate on unrelated evidence. Anyone selling you certainty in either direction is overstating what is knowable.

Why old recordings stay in scope

Retroactive scanning only works if the recording still exists, and in the United States there are reasons it usually does.

Under EEOC recordkeeping rules implementing Title VII and the ADA, employers must keep "all personnel or employment records for one year," a period that covers applicant materials, and for an involuntarily terminated employee the records must be held "for one year from the date of termination." Employers frequently retain longer than the minimum for litigation-hold and internal audit reasons. Separately, recordings accumulate through ordinary meeting tooling rather than through any integrity program — a point covered in our guide to AI notetakers capturing job interviews.

So the realistic retention picture for a recorded interview is at least a year, held by a party whose tooling can change during that window. If you want a recording removed, the request has to be made to the employer and is governed by their policy and your jurisdiction; we walk through the mechanics in how to ask for an interview recording to be deleted.

What the signals are actually worth

Red's product page groups thirteen indicators into three families. They do not carry equal evidentiary weight, and candidates should understand the difference.

Identity signals

Face-swap deepfakes, identity mismatches, and undeclared participants. This family is the strongest case for the technology, because the underlying claim — the person on this recording is not the person who was hired — is falsifiable and consequential. It is also the one with the clearest legitimate employer interest.

The caveat is reliability. Deepfake detection degrades sharply outside curated benchmark conditions, and vendor-reported accuracy figures are typically measured on datasets the vendor chose. Compression, poor lighting, low bandwidth and webcam artifacts are exactly the conditions a real interview produces and a lab dataset usually lacks.

Answer-authenticity signals

AI-generated responses, paraphrased content, off-screen reading, covert coaching, and synthetic voices. This is the family candidates should treat most skeptically, because part of it rests on AI-text detection applied to a transcript.

That technique has a documented fairness problem. In Patterns, Liang, Yuksekgonul, Mao, Wu and Zou found that widely used GPT detectors "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified," and concluded by cautioning "against their use in evaluative or educational settings, particularly when they may inadvertently penalize or exclude non-native English speakers" (arXiv:2304.02819). An interview transcript is evaluative use by definition. A candidate speaking a second language, or a candidate who rehearsed answers into a structured format, produces exactly the signal pattern these detectors associate with generation.

"Off-screen reading" has a similar problem in the other direction: looking away from the camera is also what people do when they think, when they use a second monitor for an approved shared document, or when they have a notes page the interviewer explicitly allowed.

Room-occupancy signals

Multiple people present, candidate absence, external voices, and unauthorized devices. These are the most literal and the most environment-dependent. A housemate walking through frame, a delivery, or a thin wall are ordinary facts of interviewing from home, and none of them indicate anything about the interview itself.

What a finding means procedurally

Red presents results on a shared timeline where each finding carries "the check, the severity, the time window, and the reason, with a recommended action," and the vendor is explicit that "Your team makes the decision." That framing is more honest than a single opaque risk score, and it matters for candidates: a finding is an input to a human review, not an automatic disqualification.

It also means the burden of interpretation sits with a recruiter or hiring manager who may not have the context to evaluate it. The gap between "the system flagged a nine-second window at 14:22" and "this person cheated" is entirely filled by human judgment, and that judgment is where a candidate has standing to respond.

If you are contacted about a past interview, the useful posture is specific rather than defensive: ask what was flagged, when in the recording, and what the finding was. A timestamped claim can be addressed. A vague accusation cannot, and you are entitled to know which one you are facing.

What this does and does not change

For candidates preparing for interviews now, three things follow.

Assume recorded interviews are durable artifacts rather than transient conversations. Anything captured can be re-examined under tooling that did not exist when you spoke.

Treat environment as part of preparation. Room-occupancy and off-camera signals are the easiest findings to avoid and the ones most likely to be generated by circumstances that have nothing to do with integrity.

Be realistic about what any desktop tool can and cannot affect. Control is a desktop-native assistant for Windows and macOS that stays out of supported screen captures and does not take focus from the window you are sharing — properties defined against screen capture and window focus. They are real properties, and they are also simply not the surface a recording scanner examines. We say this plainly because the alternative is the kind of blanket invisibility claim that this category of product makes impossible to honor.

The more durable preparation is the one that survives any detection model: knowing the material well enough that your own explanation of it is the fastest thing available to you. Our desktop AI interview assistant guide covers how to evaluate tools on their actual, stated behavior rather than on marketing absolutes — which is the only basis on which any of this can be assessed honestly.