Home › Blog › Open-plan office privacy

How to Use ChatGPT Privately in an Open-Plan Office

28 July 2026 · 6 min read

Open-plan desks were designed for visibility. That works fine until the thing on your screen is a chat window that invites questions you would rather not field mid-task. Here are seven approaches that actually hold up in a busy office, roughly from least to most effective.

First, worth separating two problems that get muddled together. Confidentiality is about people reading your actual words — a client name, a salary figure, a draft resignation letter. Conspicuousness is about people recognising the app from across the room and forming an opinion about it. Most advice online solves the first and ignores the second, which is why it feels insufficient. A privacy filter stops someone two desks away from reading your prompt, but the person walking behind your chair still clocks the interface instantly. ChatGPT's layout is one of the most recognisable on the web now.

1. Fix your desk geometry first

Before any software, spend ten minutes on where you sit. The single biggest factor in screen privacy is what is behind you. A monitor with its back to a wall is dramatically harder to read over than one facing an aisle, and no amount of clever software compensates for a screen pointed at the main walkway.

If you cannot move desks, angle the monitor 15–20 degrees away from the highest-traffic approach. It feels like nothing, but LCD panels lose contrast off-axis fast, and the difference between a legible screen and an unreadable smear is often a small rotation.

2. Privacy filters: good, with a real cost

A polarising privacy filter narrows the viewing cone to roughly 60 degrees, so anything off to the side goes black. They work as advertised for the confidentiality problem.

Two honest downsides. They noticeably dim the display, which matters if you already work in a dark office or care about colour accuracy. And they do nothing about the person standing directly behind you — the exact case you probably care most about. Treat a filter as a partial measure, not a solution.

3. Window management and virtual desktops

Keep the AI in its own virtual desktop and bind a switch shortcut you can hit without looking: Ctrl + ←/→ on Windows, Ctrl + ←/→ between Spaces on macOS. The advantage over Alt+Tab is that nothing flashes on screen during the transition, and there is no window edge peeking out from behind whatever you switch to.

This is genuinely effective and costs nothing. Its weakness is that it is reactive — it depends on you noticing someone approaching, which in practice means you are monitoring the room instead of working. That vigilance is its own tax.

4. Reader and focus modes

Some people paste AI output into a plain text editor and work there, going back to the chat window only to send prompts. It reduces exposure a lot, and it has a side benefit: it forces you to actually read and edit what came back rather than accepting it wholesale.

The cost is friction. Every round trip is a copy, a switch and a paste, and in practice most people abandon it within a week.

5. Change what the interface looks like

This is the approach that addresses conspicuousness directly rather than working around it. Instead of hiding the window, you change what the window appears to be.

GPTDisguise is a browser extension that renders ChatGPT inside a replica of the Google Docs interface — the toolbar, the title bar, the sidebar, the document margins. You type prompts and read answers as normal, but the shape of the thing on screen is a document. Someone passing behind you sees the layout they expect to see at a desk, and moves on.

The mechanism is worth understanding, because it determines what the tool can and cannot do for you. It is CSS and JavaScript restyling the page in your own browser. There is no server in the middle, no proxy, and no copy of your conversation anywhere new — the extension has no backend to send it to. Your prompt goes to OpenAI exactly as it would without the extension.

That also sets the limit honestly: this changes what a passer-by perceives. It does not hide anything from your employer's network logs, an installed monitoring agent, or your browser history. Those are different problems requiring different answers.

6. Keyboard shortcuts beat mouse-reaching

Whatever method you settle on, bind it to a key. The tell is never the screen — it is the person lunging for their mouse when a colleague appears. A shortcut you can hit while still typing looks like nothing at all, because it is nothing at all.

In GPTDisguise this is the toggle shortcut, available on the Premium plan. The broader point stands regardless of tool: reaching for the mouse is what draws the eye.

7. Consider whether you need to hide it

Worth stating plainly, because the honest answer is sometimes no. Many organisations now have an explicit AI policy, and a good number actively encourage these tools. If yours does, the cognitive overhead of concealment is a cost you are paying for nothing.

Where discretion genuinely helps is the middle ground: no formal policy, but a culture where using AI invites a conversation you do not want to have six times a day. Reducing the visual noise is not deception, it is just declining to make your tool choice everyone else's business. If your workplace has an explicit rule against these tools, no extension changes what you have agreed to — that is a conversation to have with your manager, not something to route around.

Putting it together

Most people land on a combination: sensible monitor placement, a disguised interface for the passer-by problem, and a keyboard shortcut so nothing on your desk moves when someone walks up. Filters and virtual desktops are worth adding if your specific situation calls for them.