ChatGPT voice Mac screen: can it see your window? Yes, with Appshots in 2026

Yes, but only in a narrow way. ChatGPT Voice on Mac can look at your frontmost window when you ask it to, and it reads that window’s image and text. The OpenAI docs I checked describe a capture on request, not a live feed of your whole screen.

Voice left the Mac app, then came back in July 2026

Older guides describe a Mac voice mode that no longer exists. OpenAI’s release notes said Voice on the macOS desktop app was retiring, effective January 15, 2026. Voice stayed on chatgpt.com, iOS, Android and the Windows app, according to 9to5Mac’s report on those notes.

The Mac app got voice back on July 23, 2026. TechCrunch reported on July 24 that the new version runs on OpenAI’s GPT-Live voice models. If a guide tells you to find a voice button in an older menu, check its date first. The menu names and steps in a 2025 guide may not match what you see today.

This matters because many people already use these tools. Pew Research found that 49% of US adults use AI chatbots, and 44% use ChatGPT. The survey covered 5,119 US adults and ran from February 17 to 23, 2026. Work use is common too. Pew found that 38% of employed adults use chatbots for tasks at work.

Which plans can use ChatGPT Voice on Mac

OpenAI’s docs list Plus, Pro, Business, Edu and Enterprise for voice in the desktop app. Edu and Enterprise start with a two-week early-access period before the feature turns on by default. Availability also depends on rollout and workspace settings. So a plan on the list may still show no menu item on day one.

If you are on the free plan, the docs do not say you get Screen context. Check your plan in your account settings before you spend time on the steps below.

How Screen context works on a Mac

The setting is called Screen context. On a Mac, it runs through a feature OpenAI calls Appshots. When you say the trigger phrase, ChatGPT takes an appshot of your frontmost window and uses it as context for your question.

An appshot holds two things. It has the window’s image, and it has the window’s accessible text. According to OpenAI’s docs, that text can include content outside the visible scroll area. A long page can send more than you see on screen.

Here is a worked example. Say you have a 30-page lease open, and you are scrolled to clause 4. You start a voice chat and say “Take a look at this.” ChatGPT captures that window. If the lease text is accessible, the capture can include clauses 5 through 30 too. That helps when you ask about a clause near the end. It is also a reason to think before you open a private document.

TechCrunch ties Appshots to macOS. It does not say whether Windows gets the same feature, so this guide covers only the Mac version.

Turn it on step by step

1. Open the ChatGPT desktop app and press ⌘ , to open Settings.
2. Choose Voice, then turn on Screen context.
3. Start a voice chat. Allow microphone access when ChatGPT asks, and choose a voice.
4. Read the screen context review prompt before you continue. Do not click past it without reading.
5. If macOS asks, allow ChatGPT under System Settings > Privacy & Security > Screen & System Audio Recording. Then turn on ChatGPT under Privacy & Security > Accessibility.
6. Test with a blank note or a harmless window. Bring that window to the front, say “Take a look at this,” and ask what it sees.

The test step is the one people skip. A harmless window shows you what the capture looks like before a real document is open. It also shows whether the permissions are working at all. Once the test answer makes sense, move to a document you wrote yourself. Save the sensitive files for later.

Why macOS asks for two permissions

macOS treats screen content and accessibility text as separate kinds of access. Screen & System Audio Recording lets an app capture what is drawn in a window. Accessibility lets an app read the text and controls that assistive tools can see. The docs name both, because the appshot carries the image and the text.

If you allow only one, the result may be incomplete. Turn on both, then run the test window again. If the answer still ignores text you can see, check the Accessibility toggle first. Many first-run problems come from a permission that was never switched on.

What it can and cannot see

The table reflects what the pages I could open say. OpenAI may add more later. Check the Learn docs again before you plan a workflow around a feature that is not in them.

Capture or live share: why the difference matters

A live share keeps watching. A capture takes one look when you ask. That gap changes how you work.

With a capture, you control the moment. You pick the window, bring it to the front, and say the phrase. Nothing new reaches ChatGPT until you do that again. That makes the feature easier to predict. It also means you have to remember to ask.

Take a memo you are drafting in a word processor. You ask whether the second paragraph makes sense. Then you rewrite that paragraph and ask again. If you skip the second capture, the answer rests on the first version of the text. Say the phrase after every big edit.

When a capture is the wrong tool

Some jobs do not suit a window capture. A capture gives the model a picture of one window. It is weaker for work that spans several windows at once.

Say you need to compare a spreadsheet with an email. Capture the spreadsheet, then type the exact figures you care about into the chat. The capture gives the model the layout. Your typed numbers give it precision. That combination beats a capture alone for anything involving exact amounts or dates.

Questions that work well with a window capture

A capture gives the model the window. Your question gives it a job. Name the part you care about, and the answer gets sharper.

• “Take a look at this. Which sentence in this paragraph is hardest to read?”
• “Take a look at this. What does the error in this log mean, and which line caused it?”
• “Take a look at this. Summarize section 3 and list every date in it.”
• “Take a look at this. What does the second table say about Q3?”

Vague questions about long pages get vague answers. Ask about one section, one error or one table at a time. If the first answer misses, say what was wrong and name the section again in the same chat.

Who gets the most out of a window capture

Some people will find this feature useful right away. Others will find it adds a step they do not need.

• Readers of long documents. A contract, a report or a research paper can be read in pieces. Ask about the section in front of you.
• People reviewing a draft. You can ask about tone, structure or a confusing sentence without copying the text into a chat.
• Anyone debugging from a log window. Capture the window that shows the error, then ask what the line means.

If you mostly type short questions, the window capture may add little. A plain typed chat can be faster for those.

Privacy: check the window before you speak

The docs say to avoid sharing windows with sensitive information. That warning is short, but it covers a lot. Accessible text can carry more than the screen shows, so a window that looks safe may not be.

A simple rule helps. If you would not paste the text into a public chat, do not let the capture read it.

• Password managers and vaults. Close them before you start a voice chat.
• Banking, tax and medical pages. Keep them in a different window or a different app.
• Client files, contracts and HR records. Ask the owner first if the file is not yours.
• Long documents with text below the fold. The words you cannot see can still be captured.

Your organization can disable the feature. If your work Mac has no Screen context option, ask your admin before you assume something is broken. Read OpenAI’s privacy policy before you use voice for client work.

Common mistakes and fixes

• Expecting a live view. ChatGPT looks when you ask. Say the phrase again after you change what is on screen.
• Wrong window in front. The capture follows the frontmost window. Click the target window, or press ⌘ Tab to it, before you speak.
• Permissions denied. Turn on Screen & System Audio Recording and Accessibility for ChatGPT in System Settings, then try again.
• Screen context missing. Check Settings > Voice, the Appshots permissions, and any organization restriction. The OpenAI docs point to those three places.
• Asking about a long page with a broad question. Name the section, table or error you care about. A precise question gets a more useful answer.

ChatGPT Voice or a different tool for the job

ChatGPT Voice is a conversation about the window in front of you. It suits questions about an article, a bug report or a document you already have open. If you are deciding between ChatGPT and a Mac voice assistant, read ChatGPT on Mac vs Lazur: which one to use for which job in 2026.

Where Lazur fits

Lazur is a voice assistant for the Mac that lives in the notch. Hold ⌥ ⌘ to ask about anything on screen, and Lazur answers out loud. Ask where something is, and Lazur draws a circle and a label around it on your screen. Ask how to do something, and Lazur guides you one click at a time, waiting for each click.

Lazur does not click buttons for you. It is not a plugin or add-in for any app, and it has no partnership with any app maker. Hold fn to dictate clean text at the cursor in any app. Lazur says it listens and looks at the screen only while you hold the shortcut.

Lazur runs on macOS, on Apple silicon and Intel Macs. Windows is planned, not available. The 3-day Pro trial needs no credit card, and it starts when you first open the app. You can download Lazur for Mac here and test it on a harmless window first.