ChatGPT Can Now See Everything You Do On Your Screen

OpenAI shipped what you could call the most controversial feature that you could give any AI. ChatGPT can now be Big Brother and look over your shoulder as you work throughout your day. So it remembers everything. And up until now, AI only remembered what you would tell it. Now it knows what you actually did. And that on the one hand is extremely powerful. And I want to talk about that and give this a spin. On the other hand, also really scary. We'll talk about that too. So let's start with the facts. Okay, this is a brand new feature available in the ChatGPT desktop app. And this is probably the most limited rollout that you could imagine. It's only available in the desktop app on Mac, then you need a Pro, Business, or Enterprise plan. Okay, so for individuals, at least $100 a month to access this. And then it's also unavailable in Europe, Switzerland, UK. So I'm on a Mac desktop computer here. I'm on a Pro plan, and I VPNed into the US to access this. And then if you go into the settings here inside of the ChatGPT desktop app right here, if I click the settings right here, you will see a new tab called Computer History in the Integrations category. Let ChatGPT follow your work. ChatGPT can summarize your activity across the apps and websites you use without ever recording your screen or audio. Ask about what you were working on, get help without explaining everything again. And this is the best part actually, or discover opportunities to automate repeated tasks. Ladies and gents, this is the missing ingredient in AI today. It has become so capable. And so many users have become extremely good at, you know, piecing different systems together and tinkering with it full time and automating a part of their life and their work. But that has been extremely technical. And me as a teacher of these things, there's just certain concepts that you can't even communicate without getting into terminals and vector databases and multiple memory systems working together. I'll elaborate on that a little bit. But first, let me show you one example from the blog post, which I think will communicate the power of this really well. And it's this one, reuse workflows. Okay, so without computer history enabled, you would say something like, I just finished my video upload workflow. Make a skill for next time. And it will tell you, okay, please share the steps and the decisions along the way, and I can turn them into something reusable. With this computer history feature, created a video upload skill from the steps in your recent activity. You can review it before using it again. And not just that, because it records everything on your computer, you can then also ask which tasks that I did could you automate the next time. Like, so powerful. But also so scary because all of a sudden ChatGPT app is watching everything you do. So if you were scared before of, you know, OpenAI taking your data and your ideas, well, this is the ultimate version of that. It literally sees everything. And as per usual, I'll be your scapegoat here, where, you know, I'll go ahead and I'll turn this on. And let's see what this allows us to do. Okay, you're in control. Sure. I can pause, clear my history, and manage which apps are included at any time. Consider pausing or excluding sensitive apps such as health or finance from your history. You also agree that you'll pause computer history during communications with others, such as video conferences or text messaging apps, unless you have their prior express consent. I mean, they just put this here for legal reasons, but basically this would be spying on every conversation that you have with everyone, including text messages, social media that you'd use on your desktop computer, and all Zoom calls. Let's start with customized apps. Exclude these apps. Okay, I'll exclude Messages. And I'll also exclude my OpenClaw Agents Web Hub. I mean, it's kind of tricky, this default of exclude these, because then you have to think of every single thing that you want to exclude. I'm going to go to include only these apps, the Finder, TextEdit. And then I guess Chrome is going to be essential. And then yeah, let's continue with this. So Google Chrome, TextEdit, and Finder. And there it is. That's the base setup. So no computer history yet. So how about this? Let's put this on one side here and open a new Chrome tab. Chrome window right here. And let's just do some work. Alright, so just opening up Google Docs. And what do we see here? Skyside session start, you have no captured active workflow in this 10-minute window, the segment only recorded session start even. Okay, so it's recording now. Let's just do something simple. Let's say I am writing up a document on community onboarding redesign. That's something that we're doing in the Advantage Club right now. It's been a huge project, we use a bunch of AI in the process. And it's fresh on my mind. So I'll do this community onboarding redesign. All right, that's a good little list. And then let's say we use AI in this process. How about that? I'm literally going to open up a new ChatGPT chat, copy this over. And I'm going to say, okay, turn this into a one-page project plan to share with the team. Oops, and I haven't shared the context. So here it is. And now let me reply. Replace the text with all of this. Almost a page. All right, that's pretty good. And now let's say we save this document locally, right? Because we have to find— there we go. Download PDF. I have it here. And what I'll do is I'll just create a new folder on the desktop, say Project Plan 8/14/26. Open the Downloads folder and move this in there. Also rename this. So mind you, I'm not recording I'm just doing and ChatGPT is listening. I guess it's not listening. I guess it's watching. And now maybe we could open Gmail. And then I'll send this to the email I wish I had, igor@gmail.com. Here is the plan, the plan for you, brother, with a clean email, attach the project plan and send this. And now let's give this a shot. Let's just send a new chat, ask what have I done today? And there it is, an overview of what I did, community onboarding redesign, PDF learning plans, turned it into a one-pager, started working on a website. These were some earlier chats. Huh. Okay, so let's retry what we saw in the blog post that I thought was so impressive. I just finished my video upload workflow, make a skill for next time. Okay, I just finished the onboarding design work, create a skill for next time. Whoa. And it adds all of this context. Look, this is literal magic because it just has the recording of what I did and it adds it. No, I'm concerned. This is too powerful. This is too much of a leap. How will they be able to like realize how strong this is? I mean, yes, there's all these privacy concerns. Yes, I already see this comment section of people saying like, hell no, I'm not doing that. Fair enough. I understand. But also just understand the power of this. Like AI's capabilities are all about the context you give it. Okay. And I'm not saying that just with prompts. Look at all these harnesses around the most capable agents. Like basically, They prompt themselves multiple times, right? That's what thinking is. In order to get better context to make the right decision. One, they use skills, which are just packaged prompts. So more context on the process and on what you're doing. They use connectors to bring context from other places to themselves. And then they use memories, which is basically just getting context from the past, from a vector database or whatever memory system you might be using. It's all about getting more context. And this is like the ultimate version on context from your life, because you're literally doing these things. If it captures your screen, if it sees you typing, if it sees what apps you go into, it doesn't just have context on what you're doing, it also has the sequence of events. And that, dear ladies and gentlemen, that I really want to highlight as the big kind of outstanding feature of this. You don't even have that in all these agentic apps like Hermès Open Claw. All of these frontier memory layers, like they don't actually have access to your working machine and see it sequentially across time how your work has been evolving. Look at that. It created a skill called Rough Project Plan and turns rough notes into to-dos. Okay, let's try saying extend the skill with the steps I took after to see if it can conjure that up. But this is crazy. It's a crazy level of access to your data in your life. And it's also a crazy ability because this honestly is the big missing piece in making agentic AI easy for people. Like a combination between this and GrokBot that came out earlier this week. By the way, I created a great video on that. I really thought that more people should watch that. It's SpaceX AI's product that looks just like a messaging application, but you have multiple AI employees in there and they can just talk to each other and do things. It's so simple. You can't even pick a model or set custom instructions. It's just simple. And a combination of that interface that people just get because it's like a messaging application and something that records how you work day to day. Well, that's the magical pairing because you do your work and then there is some middle layer that figures out which parts of your work you could be handing off to agents. And then you just talk to agents like you would to a coworker in a Slack channel or a Messages chat or in WhatsApp, whatever you use. Those are the 2 user-friendly pieces. And honestly, like these agents that were really hard to set up and to maintain like OpenClaw and Hermes, They got one piece right, which is like the interaction layer and the memories. But the big question for everybody always was, what do I even do with this? And the answer is like, yeah, if you gave it access to your main workstation and it watched you for a week, it could figure those things out on its own. That was never a point of discussion because it's like those systems are completely open and crazy. And that makes me want to talk about security for a second with this because, look at that, fair enough, it didn't even see that I went into Gmail. So, you know, maybe room for improvement. Maybe I missed something. But let me tell you about security with these things. So with OpenCLaw, the way I've been approaching it is, you know, separate Mac Mini, and it just doesn't have a single one of my real accounts. It has its own email account, it has its own Google Workspace account, it has its own Notion account, it has its own API keys. It's completely sandboxed, it doesn't touch any of my day-to-day workflows. That's why I feel pretty good about experimenting with it. This is the polar opposite. So I guess the key to, you know, keeping this even like remotely secure is really scoping this In terms of what apps it can access. So, you know, already giving it access to Google Chrome is wild, because Google Chrome is like, I don't know, half my life in it. And if you haven't asked yourself this question yet, it's probably really time to draw a line in the sand and say, okay, like, what information am I okay with sharing with these tech companies? And what am I not willing to share? But as you can tell, I'm pretty excited about this. I think this feature is absolutely massive. I want to try one more thing before I let you go here. And that is, Which one of the tasks I did today could I automate? Turning onboarding docs into project plan, highly automatable. QA for PDF learning plan delivery, highly automatable. It's even looking into the to-do list that it saw within the document. Oh my God, are you seeing this? Like it's literally considering the bullets that I kind of put in here in the beginning because it saw that is some work that I have on my desk right now. Checked what documents you worked on, highly automatable. Put on focus music to lock in. Yeah, that's also highly automatable. Yeah. QA for recommendation engine, partial. Okay, and building refining courses is low in automation. Fair enough. So are you getting this? This is not just the things you do. This is also the things within the documents that you're working on. So if you're just like reading emails, it's going to consider everything in every email. If you just scroll over a document, it's going to get all pages, pull that into the context. And if you ask something like, hey, what can I automate? It's really gonna tell you. So this is probably the best, you know, automation audit tool on the market right now. Like insanely useful. I mean, this is gonna spawn a huge discussion on the internet around, you know, safety boundaries. And just as a reminder, the last time somebody tried to track everything you do on your computer with Microsoft Recall, they got like the biggest backlash ever. They were like, are you crazy, Microsoft? I'm not gonna give you all of my information on what I'm doing. And they're like, all right, all right. I guess when OpenAI is doing it, it's cool. I don't know, we shall see. This is a bold feature that they're only releasing because the competition is heating up and they want to differentiate themselves. I mean, this has to be a reaction to the GrokBot thing from Elon. Because honestly, Grokbot is like 2 to 3 features away from being just a better consumer experience than ChatGPT and Codex combined. It's just simple. It's just intuitive. Talk to your agents. Again, check out that video. I'm a real believer in that interface. Okay. But this is a whole different beast. And that's today's video. That's what I wanted to show you. As per usual in these Friday videos, I also highlight some stories that were on my radar. I just want to do that quickly because OpenAI showed off a new version of GPT-5.6. SOL, which they call ultra fast, which is up to 14 times faster. So you know, could be reading all of your context, then at 14 times the speed, you could be getting things done. Okay, it's gonna be impressive. They're releasing that through the API for mainly coding, customer support, and things like that. Oh, then I should also note that this application, ChatGPT Desktop, is now available across Linux, Windows, and Mac. So they added Linux support, which is nice because people can run this on their virtual machines really easily. But this feature right here that I showed you, again, the computer history is Mac only. I'm just gonna turn it off for now. I need to sleep over this and make up my mind if I want to be using this and how I want to use this. But yeah, I think those are the big releases. This is a milestone. I can't understate that. And finally, I want to say you have to realize that the timing of Microsoft's release with Recall was way before we could do a lot of these agentic things. I mean, this is the year of agents in human history. We will look back and say, hey, 2026, that's when the new Opus models came out. And that's really what made projects like OpenCLAW and then Hermes and Claude Code and Codex and CoWork and now ChatGPT Work really function. It's the increase in intelligence in the models that made this work. We had projects like this before, but these new models, they make it work. But up until now, most people weren't able to gather the context to have it be useful for them. With things like Computer History to gather the context and things like GrokBot to then manage the agents while they're working on things, That changes. My prediction is going to be that before the end of the year, you're going to see all of these big AI companies compete for these 2 features. One, computer history and agentic interface that is just multiple chats like you have in Messages, because that's what we know. That's what we love. That's what we're used to already on a day-to-day basis. And then computer history is going to be the thing that's going to solve the what are the use cases that are relevant to me issue that most people face. All right, with that being said, my name is Igor Pogany, and I hope you have a wonderful weekend. Stay safe.