For a video call, start by bringing the face you are watching closer to your camera, then test how your gaze appears in that particular setup. Looking at a person on your screen and looking toward the lens are different physical actions. An AI Eye Contact effect can change how your gaze looks in the outgoing picture, but it cannot establish that you are listening, concentrating or looking at a particular person.
The useful goal is a comfortable, clear conversation. You do not need to hold an unbroken stare or calculate a perfect angle. A short comparison of your camera position, meeting layout and available effects can reveal a more useful adjustment than repeatedly judging a single screenshot.
Why watching the screen can look like looking away
Your camera records your eyes from its own position. The face you are watching may sit well below that camera, or on a different monitor altogether. Following that face therefore changes your gaze direction relative to the lens. Someone can be following a conversation while appearing to look down or sideways in the video.
Separate the visible direction of the eyes from an interpretation of the person. A camera image does not tell you whether someone is interested, honest or attentive. Our guide to AI emotion labels and their limits explains why a facial label is not a direct readout of an internal state. A gaze correction is another reason to be cautious: the displayed eyes may already have been altered.
Consider a laptop with the camera above its screen and the other participant's tile near the bottom. Moving the tile nearer the camera is a reasonable setup experiment. If you instead place your notes on a second screen, the apparent sideways glance may remain. These are different arrangements to compare, rather than evidence that a person's expression needs fixing.
What a controlled study can and cannot tell you
In Gao and colleagues' 2025 Journal of Vision study (full text), 17 participants judged photographs of actors looking at different vertical locations. The judgments concerned perceived eye contact and whether the gaze looked upward or downward. Both analyses favored looking approximately two degrees below the camera lens. The DOI is 10.1167/jov.25.11.8.
That result concerns a controlled set of photographed faces and observers, not every live call, display or person. It does not establish an ideal gaze target for your laptop or prove that an AI effect improves a conversation. Treat it as a reason to check the actual picture rather than assume that one instruction works universally. The setup comparison below is an editorial application, not a tested intervention from this experiment.
Understand which layer changes the picture
There may be a physical camera, a device-level effect and a meeting application's own camera choice. Changing a meeting window is different from changing the pixels in the camera feed. Before comparing results, identify the source and the effect you are actually using.
Microsoft's Windows Studio Effects documentation describes Standard Eye Contact as a subtle correction for looking down from the camera toward the screen; it does not adjust left or right eye movement. Teleprompter uses a stronger correction for scanning around the screen. Availability depends on compatible hardware and supported effects. The documentation also explains that a camera-level effect can apply to apps using that camera even when an app does not expose its own control. Check both the selected camera and supported device settings.
In its 2023 Maxine Eye Contact technical explanation, NVIDIA describes estimating gaze, creating a redirected eye region and blending it into the video frame. It also describes transitions toward the original gaze when redirection moves outside its working conditions, and handling obscured eyes. This is a manufacturer's explanation of that pipeline, not independent evidence that every current version preserves every blink or eye detail. It suggests useful moments to inspect: looking away, blinking and returning toward the camera.
The current NVIDIA Broadcast FAQ describes a virtual Windows camera that you select inside the calling app after enabling effects in Broadcast. This is a different route from simply choosing the original webcam. Hardware and software requirements still apply. These examples explain why a control in one window need not describe the exact feed used by another; they are not a recommendation to buy hardware or install an effect.
A short comparison before the real call
- Start with an understandable baseline. Note the physical camera, its position, the selected video source and any known effects. If supported, switch gaze correction off for the first comparison. If you cannot establish whether processing remains active, record that uncertainty.
- Bring the relevant tile closer to the camera. Move the face you are watching toward the part of the screen nearest the lens. Keep the camera, seat and lighting steady while comparing. Do not obstruct the lens or make the screen uncomfortable to read.
- Include ordinary movement. Say a short sentence, glance at your notes, blink and return to the screen. Use your own test clip if the app supports it, or an agreed test with another person. Avoid recording or sharing someone else's call as a casual setup check.
- Change one available setting. If a supported gaze effect is useful to test, enable it and repeat the same actions. Keep the same source selection and layout so you can distinguish the effect from a camera switch.
- Check the actual route into the meeting. Confirm which camera the calling app uses. A preview from a separate utility may represent a different source. Where practical, ask the other person about the received picture; a local preview alone does not verify their view.
- Choose a usable setup and stop. Compare normal playback as well as a closer look at the eyes. Keep the arrangement that is comfortable and understandable. An available effect can remain off when its result is distracting or difficult to assess.
For example, compare a brief introduction with the participant tile low on the screen, then near the camera. Only afterward compare a gaze effect. If you move the tile, change webcams and enable several effects together, a different result will not tell you which change mattered. You can still choose the clearer version, but the comparison will be less informative.
Use what you see to choose the next step
| Observation | Useful check | Decision boundary |
|---|---|---|
| You look down while watching the other person's tile | Compare the tile nearer the camera with the same seat and source | Do not infer inattention from the earlier image |
| A correction is visible in a utility but not in the meeting preview | Check the calling app's selected camera and available effect controls | Do not assume every preview uses the same feed |
| A sideways glance remains after a supported correction | Check that feature's documented direction and scope | A vertical adjustment need not correct horizontal gaze |
| The eyes change abruptly during a blink, turn or return | Repeat that movement with the effect off, keeping other settings steady | If the effect remains distracting, use a simpler setup |
| A single screenshot looks awkward but playback is understandable | Review the surrounding sequence at normal speed | One frozen frame does not represent a whole conversation |
These checks do not diagnose the software or measure social skill. They help identify a practical next step without treating your face as the problem. Background choices are a separate variable; our video-call background guide covers that part of the setup.
Keep a video setup separate from a photo assessment
A frame taken while you blink or speak can be a poor still portrait even when the video is fine. Our guide to choosing a photo from a video explains that separate selection task. A gaze-corrected frame also records an edited appearance, not an unaltered reference of where your eyes were directed.
If you separately want photo-based appearance feedback, you can explore FaceScore with a clear, representative still photograph. Keep the capture and editing conditions in mind when interpreting a result. FaceScore's photo feedback does not verify video eye contact, attention or a person's value.
Use the call itself as the practical test: can you follow the discussion, read what you need and appear clearly? Camera placement and an optional effect are tools for that task. A comfortable setup with an understood video source is a useful outcome, even when it includes ordinary glances away.