Does Zoom AI Companion Detect Screen Recording? (2026)

Direct Answer

Zoom AI Companion does not detect screen recording, and it does not proctor candidates. It is a server-side productivity layer — meeting summary, smart recording, in-meeting query, whiteboard generation — that processes only the meeting's own audio, video, transcript, and shared-screen content, never what is rendered on a participant's local screen outside the shared region. Zoom notifies all participants only when its own native cloud or local recording is started; third-party recorders such as OBS, QuickTime, ScreenFlow, or the operating system's native capture send no signal back to Zoom and trigger no notification. The persistent misconception that "AI Companion watches the candidate" inverts what the feature does: it summarizes the meeting for the host, and it has no mechanism to inspect a participant's wider machine.

How Zoom AI Companion Actually Works

Zoom AI Companion is a generative-AI assistant bolted onto the meeting product, and understanding it starts with knowing it only ever sees the meeting's own data stream. When a host enables meeting summary or smart recording, AI Companion ingests three inputs: the meeting audio (for transcription and summarization), the chat messages (for context), and the content of a shared screen (read via optical character recognition when someone is actively sharing). From those inputs it produces a summary of topics and next steps, smart-recording chapters, and answers to natural-language questions about what was discussed.

Every one of those inputs is something a participant transmitted into the meeting on purpose. The audio comes from the microphone the participant selected. The shared-screen content comes from a screen share the participant explicitly started. The chat comes from messages they typed. AI Companion has no separate channel into a participant's operating system, no agent installed on the candidate's machine, and no ability to enumerate windows, processes, or displays. It is structurally a consumer of meeting data, not a monitor of endpoints. That single architectural fact answers most of the questions candidates ask about it.

Does AI Companion run gaze tracking or eye-movement analysis on the candidate? No — actually it does no per-participant integrity analysis at all. Gaze and attention models belong to dedicated proctoring vendors. AI Companion's video handling is limited to the recording itself and the OCR of shared content; it does not score whether a participant is looking away from the camera.

What Each Zoom Feature Notifies Participants About

Zoom does notify participants about several events, but the notifications cover Zoom's own actions, not anything happening on a participant's external software. The native recording feature is the clearest example: when a host or participant starts Zoom's built-in cloud or local recording, every participant sees a recording indicator and is asked to consent, and on many configurations they hear an automated announcement. Screen sharing is visible to everyone the instant it starts, with the sharer's name attached. Participant join and leave events appear in the roster. Reactions and raised hands are visible.

Starting in January 2026, Zoom added a distinct AI Companion Policy Disclaimer: when a host enables meeting summary or smart recording, all participants must acknowledge and agree to remain in the meeting. This is a data-processing consent prompt, not a proctoring notice — it tells participants that AI features are running, the same way the recording banner tells them a recording is running. None of these notifications fire for a third-party recorder, because Zoom never learns the third-party recorder exists. The distinction between what triggers a banner and what does not is the entire story, and it parallels the analysis in the dedicated piece on whether Microsoft Teams detects screen recording.

What Zoom Can and Cannot See — by Feature and Plan

The detection surface does not meaningfully change across Zoom's pricing tiers, and it does not change when AI Companion is toggled on. The matrix below maps each signal against the plan and the AI Companion state, and the pattern is consistent: Zoom sees only what is transmitted into the meeting.

The only rows that read "Yes" are Zoom's own meeting events. Everything in the lower block — the things candidates actually worry about — is structurally invisible to the platform regardless of plan or AI Companion state. Upgrading from Free to Enterprise buys longer cloud retention, more administrative controls, and the AI Companion add-on; it buys no new ability to observe a participant's machine. The same architectural boundary governs Google Meet, and the side-by-side comparison lives in Google Meet vs Zoom AI detection.

The Participant-Notification Event, in Pseudocode

The cleanest way to see why third-party tools go unnoticed is to look at how the recording notification actually fires. The event is driven entirely by Zoom's own recording state, not by any inspection of the participant's machine. A simplified version of the logic that decides whether to show the banner and announcement:

on_recording_state_change(meeting, actor):
    # This only ever runs for Zoom's OWN recording engine.
    if actor.recording_kind in ["zoom_cloud", "zoom_local"]:
        meeting.set_recording_indicator(visible=True)
        for participant in meeting.participants:
            participant.show_consent_prompt(REASON_RECORDING)
            if meeting.audio_announcement_enabled:
                participant.play_announcement("This meeting is being recorded")
        return

    # Third-party recorders never reach this function.
    # OBS, QuickTime, ScreenFlow, OS-native capture, and screenshot
    # tools run in the participant's OS, emit no signal to Zoom's
    # servers, and have no callback into meeting state.
    # => no banner, no consent prompt, no AI Companion summary entry.

on_ai_companion_enabled(meeting):
    # Separate, parallel consent path (2026 policy disclaimer).
    for participant in meeting.participants:
        participant.require_ack(AI_COMPANION_POLICY_DISCLAIMER)
    # Processes meeting audio/transcript/shared-screen only.
    # Never inspects a participant's local windows or processes.

The function that shows the banner is reachable only from Zoom's internal recording engine. There is no branch in which an externally running recorder, a screenshot utility, or an overlay window calls into it, because those tools live in the participant's operating system and never communicate with Zoom. This is the same reason the platform cannot tell you whether a participant pasted from an external source — a topic covered in can interviewers see paste events.

The Four AI Companion Features, and Why None of Them Proctor

AI Companion is not a single thing — it is a bundle of four user-facing features, and walking through each one shows that none has a proctoring function hiding inside it. The meeting summary feature consumes the transcript and produces a structured recap of topics, decisions, and action items; its only input is what was said. Smart recording takes a cloud recording and segments it into chapters, attaches highlights, and generates next-step suggestions; it operates on a recording the host already chose to make, not on any live inspection of participants. The in-meeting query feature lets a participant ask "what did I miss" and answers from the transcript so far. Whiteboard generation turns a text prompt into a diagram.

Every one of these is a transformation of meeting content that already exists, performed on Zoom's servers after the content has been transmitted into the meeting. There is no fifth feature that watches a participant's machine, scores their behavior, or flags suspicious activity, because Zoom does not sell AI Companion as an integrity product and has not built one into it. A recruiter who enables AI Companion gets better notes; they do not get a proctoring dashboard. This is the cleanest way to dispel the "AI Companion is watching me" worry: enumerate what the feature actually does, and the surveillance interpretation has nowhere to attach. The way candidate-side tools relate to this kind of meeting AI is explored further in is ChatGPT in a Zoom interview detectable.

Server-Side Processing vs Endpoint Inspection

The deepest reason Zoom cannot proctor is the location where its processing happens: on the server, on data already sent, never on the participant's endpoint. A proctoring system, by contrast, has to run code on or very close to the candidate's machine — a browser lockdown extension, a desktop integrity agent, a webcam analysis pipeline tied to a specific test window. That is how it gains the ability to enumerate displays, scan processes, or lock the environment. Zoom installs no such agent for the purpose of monitoring participants. The Zoom client on a participant's machine is a communication client; it captures the camera and microphone the user selected and sends them, and it renders the meeting it receives.

Because the heavy lifting happens server-side on already-transmitted streams, AI Companion's knowledge of any participant is bounded by what that participant chose to send. It cannot reach back through the client to ask the operating system what windows are open, what processes are running, or how many monitors are attached. Those queries are exactly what an integrity agent performs and exactly what a meeting client has no reason to perform. The architectural gap between "summarize what was transmitted" and "inspect the endpoint that transmitted it" is the gap between a productivity tool and a proctor, and Zoom sits firmly on the productivity side. The same server-side-versus-endpoint distinction is what separates conferencing tools from the assessment platforms analyzed in does Webex detect AI in interviews.

Does the meeting host get a report of who recorded or screenshotted the call? No — actually no such report exists, because Zoom only knows about its own recording engine. There is no per-participant capture log for third-party recorders or screenshots, so the host has nothing to review on that front.

What Zoom Structurally Cannot Detect

There is a clean architectural boundary around what Zoom can observe, and everything on the far side of it is invisible by construction rather than by oversight. Zoom receives a camera feed, a microphone feed, chat text, and — only when a participant chooses to share — a screen or window feed. It does not run an agent on the endpoint. That means a third-party screen recorder capturing the Zoom window leaves no trace in the meeting, because the capture happens in the operating system's graphics pipeline, downstream of anything Zoom controls. A screenshot is the same: it is an OS action Zoom never sees.

The list extends to every "off-stream" surface. An overlay window positioned over the Zoom window is not transmitted to Zoom unless the participant shares the exact screen region it occupies — and tools designed to render outside standard capture surfaces stay out of even an entire-screen share. A second monitor that the participant does not share is simply not part of any feed Zoom holds. A virtual camera or a virtual audio device (the kind created by loopback drivers) is accepted as an ordinary input; Zoom does not run forensics to distinguish a physical sensor from a software source. Content rendered on another virtual desktop, on a phone, or on a tablet is on a device Zoom has no relationship with at all.

Does sharing your entire screen let Zoom see your other monitor? No — actually an entire-screen share captures only the single display you selected in the share dialog, not every connected display at once. A second physical monitor is a separate capture source, and Zoom transmits only the one you chose.

Live vs Async: Where the Confusion Comes From

The reason candidates conflate Zoom AI Companion with proctoring is that they have used both kinds of platform and the categories blur together. Zoom is a live, synchronous communication tool: a human interviewer is present, watching the camera feed and the shared screen in real time. Proctoring platforms — the kind discussed in the webcam proctoring software breakdown — are integrity systems that install browser lockdown, run gaze and overlay heuristics, and produce a suspicion report. AI Companion is neither of those things; it is a note-taker for the host.

In a live Zoom interview, the detection surface is a human, not an algorithm. The interviewer can notice a candidate glancing repeatedly off-camera, a long unexplained pause, or an answer that arrives suspiciously polished. That is ordinary human judgment, the same judgment present in any face-to-face conversation, and it is unrelated to whether AI Companion is summarizing the call. AI Companion does not surface a "suspicion score," does not flag the candidate, and does not retain anything beyond the meeting summary and recording the host explicitly enabled. The framing of how candidate-side assistants relate to live interviews is developed further in how AI interview assistants work.

Single-Window vs Entire-Screen Share: The One Setting That Matters

The single share-dialog choice — one window or the whole display — changes more about candidate privacy than any AI Companion setting. When a participant shares a single application window, Zoom transmits only that window's pixels; notifications, other applications, the taskbar, and other monitors never enter the stream. When a participant shares an entire screen, everything on that one display flows to Zoom for as long as the share is active, including notification banners and any window that happens to be on top.

This matters in two directions depending on who is sharing. If the interviewer shares their screen to present a problem, the candidate's machine is never transmitted at all — the candidate is only ever on camera and audio. If the candidate shares their screen to write code, the single-window choice keeps everything except the editor out of view, while the entire-screen choice exposes the whole display. Neither choice changes what AI Companion does; AI Companion's OCR reads whatever screen content is shared, but it reads it as meeting context for the summary, not as an integrity signal. The same single-window-versus-full-screen tradeoff appears in remote interviews across tools, as covered in the practical guide to remote technical interviews on Zoom and Google Meet.

Interviewer Shares vs Candidate Shares: Two Different Threat Models

Who initiates the screen share completely changes which machine is exposed, and candidates often reason about the wrong one. When the interviewer shares their screen — to present a coding prompt, walk through a system-design canvas, or display a shared editor — the candidate's machine never enters any stream. The candidate is present only as a camera feed and audio, and AI Companion, if running, summarizes the conversation and OCRs the interviewer's shared content, not the candidate's. In this configuration the candidate's local screen, second monitor, and any overlay are entirely outside Zoom's view by default, because nothing on the candidate's side was ever shared.

The mirror case is when the candidate shares their screen to write code in a live coding round. Now the candidate's chosen surface is transmitted, and the single-window-versus-entire-screen decision becomes the determining factor for what the interviewer sees. A single-window share of the editor exposes only the editor; an entire-screen share exposes that whole display. Even in the entire-screen case, only the one selected display flows to Zoom, and surfaces designed to render outside standard capture stay out of the stream. Recognizing which of these two threat models applies is the difference between worrying about the wrong things and reasoning accurately, and it generalizes across every conferencing tool, including the comparison drawn in Google Meet vs Zoom AI detection.

What Detection Means in Practice for a 2026 Candidate

The practical takeaway is that in a Zoom interview the meaningful observer is the human interviewer, not the software, and the software's reach is narrow and well-defined. AI Companion will produce notes; it will not flag the candidate. The recording banner will appear only if someone uses Zoom's own recorder; it will not appear for anything else. The interviewer will form impressions from the camera feed and the conversation, exactly as they would in person. None of this is new surveillance — it is the ordinary visibility of a video call plus an opt-in note-taker that summarizes what was already said.

This stands in sharp contrast to a browser-based proctored assessment, where an integrity agent really does instrument the endpoint, lock the environment, and produce a suspicion report. Conflating the two is the central error: a Zoom call is a conversation with a productivity assistant attached, while a proctored coding test is an instrumented environment with an integrity layer attached. A candidate who understands that distinction stops attributing assessment-platform capabilities to a meeting tool. The broader question of where AI assistance sits relative to interview ethics is treated in is using AI during a coding interview cheating.

There is also a consent and legal layer worth separating from the technical one. The fact that Zoom cannot technically detect a third-party recorder does not make recording a meeting without consent advisable: many jurisdictions operate under two-party or all-party consent rules for recording audio, and a company's interview policy may prohibit recording outright. The technical reality — no banner, no notification, no AI Companion entry for an external recorder — is a statement about Zoom's architecture, not a statement about what is permitted. Candidates reasoning about a Zoom interview should treat the platform's detection limits and the interview's stated rules as two independent constraints, the same way they would for any conferencing-based round, including those at companies whose processes are profiled in pieces like the Shopify technical interview process and the Linear technical interview process, where the format and tooling are set by the employer rather than by Zoom.

What Candidates Get Wrong

Most confusion about Zoom AI Companion comes from a handful of recurring misconceptions, each of which collapses once the platform's data boundary is clear.

• Assuming AI Companion is a proctor. It is a summarizer. It produces meeting notes and recording chapters for the host. It does not score participants, flag behavior, or run any integrity heuristic. The 2026 policy disclaimer is a consent prompt, not a monitoring notice.
• Believing the recording banner detects all recording. The banner reflects Zoom's own recording engine only. A third-party recorder produces no banner because Zoom never learns it is running. The absence of a banner does not mean nobody is recording.
• Thinking an entire-screen share exposes every monitor. A share captures one selected display or one window. Other monitors are separate sources that must be selected individually, so they are not transmitted unless chosen.
• Treating a virtual camera as a red flag Zoom raises. Zoom accepts virtual cameras and virtual audio devices as ordinary inputs and does not announce them to other participants or run forensics on the source.
• Confusing the interviewer's human judgment with platform detection. A live interviewer noticing off-camera glances is human observation, not an AI Companion feature. The two are completely separate, and conflating them leads candidates to over-attribute capability to the software.
• Assuming a higher Zoom plan adds surveillance. Enterprise tiers add retention, admin controls, and the AI Companion add-on — never the ability to inspect a participant's machine. The detection surface is identical from Free to Enterprise.

Frequently Asked Questions

Does Zoom AI Companion detect screen recording done with OBS or QuickTime? No. AI Companion processes only the meeting's own audio, video, transcript, and shared-screen content on Zoom's servers. A third-party recorder runs in the participant's operating system, downstream of anything Zoom controls, and emits no signal back to Zoom. There is no banner, no consent prompt, and no entry in the meeting summary.

Does Zoom notify participants when AI Companion is turned on? Yes, but only as a consent disclaimer. Since January 2026, when a host enables meeting summary or smart recording, all participants must acknowledge the AI Companion Policy Disclaimer to remain in the meeting. This tells participants that AI processing of the meeting is active; it is not a statement that any participant is being individually monitored.

Can the interviewer see my second monitor on Zoom? Only if you explicitly share that monitor. Each connected display is a separate capture source in the share dialog. A second monitor you do not select is never transmitted to Zoom, so neither the interviewer nor AI Companion sees its contents. Sharing the entire screen captures the one display you chose, not all of them.

Does AI Companion store a recording of me without my knowledge? No. Smart recording and meeting summary only run when the host enables them, and the 2026 disclaimer requires every participant to acknowledge that before the meeting continues. If you did not see and accept the disclaimer, AI Companion's summary and smart-recording features are not running on that meeting.

Is using a virtual camera detectable by Zoom AI Companion? No. Zoom and AI Companion accept a virtual camera as an ordinary video input and process its stream like any other. There is no forensic check that distinguishes a hardware sensor from a software loopback, and no notification fires to other participants about the source type.