Screen Privacy for AI Tools: 6 Approaches Compared
"Screen privacy" gets used for at least three different problems, and a method that solves one can be completely useless against another. This is a straight comparison of six common approaches against the threats people actually have.
The three problems, separated
Almost every disappointed review of a privacy product comes from a mismatch here, so it is worth being precise about which one you have:
- The reader. Someone close enough to read your actual text. A colleague at the next desk, a stranger on a train.
- The identifier. Someone who cannot read a word but recognises the app from its shape and colours. Anyone walking past. This is the one most tools ignore.
- The record. Network logs, browser history, a monitoring agent installed by IT, a screen-recording tool in a meeting. Nothing on your screen affects this at all.
Keep those straight and the comparison below becomes much easier to act on.
The comparison
| Approach | Stops the reader | Stops the identifier | Stops the record | Main cost |
|---|---|---|---|---|
| Privacy filter | Yes, off-axis only | No | No | Dims the display |
| Virtual desktops | Yes, if you switch in time | Yes, if you switch in time | No | Constant vigilance |
| Incognito mode | No | No | Local history only | Widely misunderstood |
| Second monitor | Partly | Partly | No | Desk space, not portable |
| Browser zoom out | Slightly | No | No | Unreadable for you too |
| Interface disguise | No | Yes | No | Needs an extension |
1. Privacy filters
A polarising sheet over the panel that collapses the viewing angle to roughly 60 degrees. Genuinely effective against the reader approaching from the side, and the only option here that works without any software.
It does nothing about the identifier, because shape and colour survive the angle change. It also dims your display and can interfere with touchscreens. Around £30–60 for a laptop panel.
2. Virtual desktops and window switching
Free, built into every operating system, and effective against both the reader and the identifier — conditional on you switching before you are seen.
That condition is the whole problem. It converts a privacy question into an attention question, and you end up half-watching the room. It also fails completely in the case where someone is already behind you when you notice them, which is exactly when you need it.
3. Incognito / private browsing
Worth addressing because it is the most commonly suggested answer and the least relevant one. Incognito controls what is written to local browser history and cookies. It has no effect whatsoever on what is visible on your screen right now.
It also does far less to the record than most people assume: your employer's network, your DNS resolver and the site itself all see the same traffic either way. Useful for keeping a shared family laptop's history clean. Useless for an office.
4. A second monitor, angled away
Put the AI on a secondary display turned away from the room. Physically simple and works reasonably against both the reader and the identifier, depending entirely on your desk layout.
Costs desk space and does not travel. If you work from cafés or hot-desk, this is not your answer.
5. Zooming the browser out
Included because people genuinely do this. Dropping to 50% zoom makes the text hard to read from a distance — and hard to read for you, which is the point of failure. The layout stays just as recognisable, so it does nothing about the identifier.
Not recommended. It trades your own usability for a small gain against one threat.
6. Disguising the interface
The one approach aimed squarely at the identifier. Rather than hiding the window, it restyles it: GPTDisguise renders ChatGPT inside a replica of the Google Docs UI, so the thing on screen has the silhouette of a document — toolbar, title bar, sidebar, margins.
Its strength is that it is passive. There is no shortcut to hit in time and no window to switch, so the vigilance cost of virtual desktops disappears. Its limit is equally clear: it does nothing about the reader. Someone leaning over your shoulder reads your prompt as easily as before, because the text is still text.
On the record problem it is honest by construction. The disguise is CSS and JavaScript applied in your own browser — there is no GPTDisguise server, so no new copy of your conversation exists anywhere. But equally, your traffic to OpenAI is unchanged and unhidden. If a monitoring agent is installed on your machine, an interface skin is irrelevant to it.
Try the disguise
Free plan covers the Google Docs disguise with ChatGPT. No account needed.
Add to Chrome — it's freeFirefox version available in beta.
What to actually pick
Match the tool to the problem you have, and combine only where the problems genuinely stack:
- Café or train, strangers nearby: privacy filter. The reader is your threat and nobody there cares which app you use.
- Office, colleagues walking past: interface disguise. The identifier is your threat, and it is the one a filter cannot touch.
- Both: filter plus disguise. They cover different axes and do not conflict.
- Worried about the record: none of these. That is a policy and IT question, and any product claiming otherwise is overselling.