Stop Sending Your Private Data to OpenAI's Servers

Author: Matt Williams · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Stop Sending Your Private Data to OpenAI’s Servers · Channel: Matt Williams · 08:55 · Watch on YouTube

Topics: AI models, Local and open-source models, Privacy and data protection

Summary

Matt Williams explains how to connect local Ollama models to the ChatGPT desktop app on macOS so prompts can be processed on the user’s own machine. He walks through enabling the integration, adding local models, and switching between them and OpenAI models, while warning that switching models within a conversation can send its contents to OpenAI. The video describes data retention and app-access concerns, then outlines local models’ privacy benefits and practical limitations. His recommendation is to use cloud models for low-stakes tasks and local models for sensitive information.

Key points

The ChatGPT desktop integration with Ollama is available on Mac; Ollama itself also works on Windows and Linux.
The setup requires the installed ChatGPT app, enabling the Ollama integration, and restarting ChatGPT.
Local Ollama models appear in the model picker alongside OpenAI models and can be switched during use.
Switching to an OpenAI model within a conversation that contains private information can send that conversation’s data to OpenAI.
The video says ChatGPT conversations and uploaded files are retained in the cloud, and opting out of training does not stop retention.
The video warns that Work with Apps can include content from open Mac apps in messages sent to OpenAI.
According to the video, local model prompts and responses stay on the device, and local models continue working without an internet connection.
Local models can load more slowly at first and may have a quality gap, but the speaker says they can be suitable for many tasks.
The speaker recommends using local models for sensitive material such as legal, medical, financial, or proprietary information.

Chapters

Start	Chapter	Summary
00:00	Why use local models	The speaker introduces the privacy risks of sending sensitive information to cloud services and previews using Ollama with ChatGPT.
01:33	Setting up Ollama with ChatGPT	He demonstrates enabling the integration on Mac, adding a local model, and switching between local and OpenAI models.
04:10	What local models keep private	The video reviews ChatGPT data retention and app-access concerns, then explains how local inference keeps prompts on the device.
06:49	Limitations and when to switch	The speaker discusses setup scope, performance trade-offs, and his recommendation to use local models for sensitive tasks.

Quotes

They go to another computer. — 00:00

The only difference is which brain answers your questions. — 03:37

If it stops working when the internet dies, it was never truly local. — 06:49

Switch to Ollama for anything that could hurt you if it leaks. — 07:50

Concepts and entities

Ollama, ChatGPT, OpenAI, Ollama Cloud, modelos locais, work with apps, computer history

Full transcript (original language)

00:00 When you use ChatGPT, your chats don’t just stay in the app window. They go to another computer. That may be fine for a grocery list, but not for your tax return, or your bank login credentials, or your company’s source code. But, you don’t have to give your private information to OpenAI just to use the desktop app. Every prompt you type, every file you upload, and every app ChatGPT reads through work with apps, gets stored in the clouds and get used

00:31 to train their models. Now, I was on the founding team at Ollama, and today I’m going to show you how to swap in local models, so your most private data never leaves your machine. The setup is almost the same as the Claude desktop integration that I showed recently, but with one weird quirk in how you pick your models. Unfortunately, for now, this integration with ChatGPT is only available on the Mac. But, don’t go away yet if you’re on Windows or Linux.

01:01 I’m going to be showing how you set it up on my Mac, but on other platforms, Ollama still works, and many other apps give you the same functionality as the ChatGPT desktop app. If you’re new to local models and Ollama, getting started is easy. Visit ollama.com and click the download button. There are other tools you could use to run AI locally, but Ollama is the open-source app that is by far the most popular, and definitely the easiest to use without giving up any

01:33 power. Once you’ve downloaded it, open the downloaded file and drag Ollama into applications. Now, launch Ollama. At this point, I go into a Mac-specific part, so you can skip [clears throat] ahead to this timestamp if you’re on Windows or Linux, because I still have info for you there. Now, I’m going to assume that you already have the ChatGPT app installed on your your We’re specifically talking about the installed app rather than the website for ChatGPT. Now, open the Ollama desktop app. In the sidebar, click apps,

02:05 then find ChatGPT, and flip the integration on. You can also toggle it from the menu bar, which is usually how I do it. The app will tell you that ChatGPT needs to restart, so go ahead and do that. If you haven’t pulled any models from Ollama yet, try using Ornith 1.5 9B. I’m copying the full name of the model from the Ollama website. Just go to models and scroll down a little. I chose this one because it’s really powerful while being small enough that

02:36 most Macs can run it, even some of the cheapest ones. Now, here’s where it gets a little weird. In a recent video, I showed the process for Claude desktop. We have to select the models we want from this drop-down for each existing model. For ChatGPT, we add them to a block in what looks like a text box. I find it kind of surprising that they did something that has such an awkward UI, especially after using a very different method for essentially the same thing

03:07 just above it in the settings window. The annoying part is five models are already selected, and you have to remove one before you can add yours. I’d much rather be able to add a sixth or seventh and then get told to remove the extra ones. Instead, you have to hunt around and delete a model first. Now, it’s not a deal breaker, but it just feels a bit awkward and clunky. Once you’ve got your local models added, restart the ChatGPT desktop app one more time.

03:37 When it opens again, your local Ollama models appear right alongside the OpenAI models in the picker. The UI looks the same. The shortcuts work the same. You can switch between a local model and an OpenAI model mid conversation if you want. The only difference is which brain answers your questions. And that single swap locks your private data on your hardware. Of course, if you shared a private document with a local model, then switch mid conversation to an open AI model,

04:10 all the data will go there, too. So, be careful. In that case, it’s probably best to switch to a new chat. Before you get too comfortable, you need to understand exactly what you’re avoiding every time you hit send. And welcome back to the Windows and Linux folks. You can pay attention again. The ChatGPT macOS app follows the exact same data retention rules as the web version. That means every conversation, every uploaded file, and every screenshot you add to a chat lives in the cloud tied to your open AI account.

04:43 By default, your content feeds into model training unless you dig into settings and opt out. And even if you do opt out, open AI still keeps your chats for about 30 days or so after you delete them. The opt-out stops training use, but it doesn’t stop retention. Then there’s work with apps. When this is enabled, ChatGPT can read up to about 200 lines from open apps on Mac. That includes your IDE, your terminal, your notes, and your text editors. So,

05:15 if you have a database password in a terminal window or proprietary code open in Xcode, that content can get bundled into your message and sent to open AI servers. It’s convenient, but it means even more of your local activity is silently uploaded. Open AI also rolled out a feature called computer history in August of 2026. If you opt in, it records your clicks and typing and your app switches. The memory files sit unencrypted on your Mac and any other program running as the

05:47 same user can read them. Even before anything hits OpenAI servers, your activity is sitting exposed on your machine. So, now you know what you’re up against, and here’s how local models actually change the game. >> [snorts] >> When you run an Ollama model on your own machine, your prompts and responses never leave your device. The inference happens on your GPU, and there is no outbound network traffic for the actual chat. Ollama’s privacy policy is explicit

06:17 about this. When you use local models, they don’t collect, store, or transmit your prompts. They don’t see your data at all. Now, I mentioned this with local models, but they don’t collect anything with their cloud models, either. I use Ollama Cloud for day-to-day work, but I swap to local models depending on the task. You can verify this yourself. Pull your network cable out, or turn off Wi-Fi, and local models continue working. The responses feel identical because the processing is happening on your chip,

06:49 not theirs. That is the test. If it stops working when the internet dies, it was never truly local. Now, let’s be honest about what this does and does not fix. This setup is per user. It protects your own machine, but it won’t magically secure a whole team. If your company needs shared access, everyone needs to set this up individually. Now, there is a new team and enterprise version of Ollama that may help, but I haven’t looked into that yet.

07:19 For most tasks, the quality gap between OpenAI’s models and those on Ollama is smaller than you’d expect. Local models can load slower on first use because the model has to fit into your RAM, and your RAM and processing is slower than what OpenAI has. And its models stay in memory across lots of users. After that first load, the speed feels normal. And quite often, local models can perform faster than what you get with OpenAI models. And if a local model

07:50 isn’t cutting it for specific tasks, switching back to a cloud model just takes a few seconds. For public research or low-stakes brainstorming, OpenAI models are totally fine. Switch to Ollama for anything that could hurt you if it leaks. Think legal documents, medical information, proprietary code, or financial records. That’s the split I recommend, and it’s not just tinfoil hat conspiracy theory here, either. OpenAI has been caught a few times sharing user data from prompts

08:22 with other users. So, there is a real risk here. But thankfully, you don’t have to give your private data to OpenAI just to use the ChatGPT desktop app. Keep the integrations and shortcuts you like, just swap the brain to Ollama. If you want more local AI tricks, hit subscribe and the bell. I cover Ollama and automation every week. We’ll see soon if I can increase that cadence further. Thanks so much for watching. Goodbye.

08:55 My coffee.