AI Chat Privacy

AI chatbots are all the rage. They’re powerful tools that do amazing things, but come with serious caveats and concerns. Today I’ll focus on privacy.

Background

Artificial Intelligence, as a technology, has been around since at least the 1950’s. But the current AI boom started with the advent of Large Language Models (LLMs) around 2017, and in particular the public release of OpenAI’s ChatGPT in late 2022. Within five days of its debut, ChatGPT 3.5 had one million users – and 100 million users within two months. Its simple, text chat interface and simulated conversation made it instantly accessible to anyone. Soon after, other big players got in the game and released competing chatbots like Google’s Bard (which became Gemini), Anthropic’s Claude and Meta’s Llama.

Image generators actually hit the scene prior to ChatGPT. DALL-E, Midjourney and Stable Diffusion debuted earlier in 2022, allowing people to describe an image using text and have the tool create the image. They were quirky at first, and did really weird things with people’s arms, legs and fingers. In fact, some enterprising people created 6th finger rings, so that the wearer could claim that pictures of them were faked.

There are whole books on these subjects already – not just the technologies but the effects AI is having on jobs, products, the environment and society – despite the fact that we’re only at the very beginning of this new “AI era”. (Many books are being completely written by AI, too.) I’m particularly looking forward to Carissa Véliz’s upcoming book on AI. (I’ll be interviewing her about it when it comes out, but we already recorded a great interview on AI ethics last fall that you should check out.)

Privacy and AI

There are many serious problems with the way we’re implementing AI today. And to make matters more complicated and dangerous, we’re also rushing to give these AI systems agency – that is, empowering them to interact with other apps and services on our behalf, autonomously. I’m not going to get into all of that here. However, I will say that I don’t think most of these problems are inherent to AI itself. We could have (and still can) do things differently. We shouldn’t throw the baby out with the bathwater.

In this article I want to address one very specific problem with modern AI: privacy. Due to the processing power required to run these large language models (the part we call inference), most chatbots today send your queries and prompts to the cloud for processing. That means that your entire ‘conversation’ history is visible to whatever AI provider you’re using (OpenAI, Google, Meta, Anthropic, etc). And as we all know now, that information can be used to learn a lot about you, which may in turn be sold to, leaked by, or shared with third parties. There are very few guardrails on AI data collection – it’s a real wild west gold rush situation.

These AI tools, despite their issues, are powerful and can be genuinely helpful. And regardless of how you feel about the technology, it’s everywhere now. We have to deal with it and we should be striving to limit the amount of personal information we’re exposing as we do.

Local AI

There are two primary ways you can make your AI experiences more private: using a purely local AI tool or using a privacy-focused cloud AI service. Let’s start with local AI. You don’t have to run AI tools in the cloud. Modern devices have enough power to run basic and even some complex AI tools locally – as in, on a device that you own and resides physically with you. That may be a smartphone or a computer (or some day, maybe a dedicated appliance in your home). These tools may talk to the internet to get software updates, but all of the processing of your queries happens on the device. As of this writing, that’s the ideal situation for AI privacy.

When it comes to smartphones, the basic approach by both Apple and Google is to do as much AI processing on the device as possible, but use the cloud when necessary. Things like speech-to-text (dictation) and simple photo enhancements can be done locally, for example. Unfortunately, right now we don’t have a lot of control over or visibility into this division of labor. In my view, Apple is doing a better job of this right now than Google. But in the end, Google is already using your local data to target you with ads, anyway, so it’s sort of moot where they do your AI processing. Apple isn’t perfect on this, but it’s a lot better than Google.

On your computer, you have a lot more options and a lot more control. Computers have more capabilities than smartphones (though the difference today isn’t as much as it used to be). You can download powerful AI tools that explicitly keep all data local, never sending it to the cloud and not sharing it with anyone.

As computer chips get inexorably faster and become more optimized for AI processing, we can expect more and more processing to happen on our devices – and that’s a good thing for privacy.

Private Cloud AI

Most of the popular AI cloud tools today are cloud-first and have very little in the way of privacy protections. You should never send information to a cloud-based AI tool that you wouldn’t want to become public some day. However… having said that, there have been some concerted efforts to create online, cloud-based AI systems that respect their users’ privacy. Unfortunately, the only viable way to process your queries today is with unencrypted plain text. That is, there currently is no practical way to have true end-to-end encryption (E2EE) with AI systems. While work is being done to do AI processing on encrypted queries (an amazing tech referred to as homomorphic encryption), it’s just way, way too slow right now.

That means that some part of the system has to be able to read your queries – and that means that technically it could be extracted, logged, saved, and/or shared. And since this happens on computers you don’t own and can’t inspect, we have to take the AI provider’s word on what they’re doing with your data.

However, many of these privacy AI providers go to great lengths to anonymize your queries – so that even though the text is unencrypted, they can plausibly say that they have no way to know who sent that query. And by extension, no third party could figure that out, either – be they criminal hackers or law enforcement or intelligence agencies. (Of course, you can give yourself away in the chat conversation itself.)

Recommended AI Tools

So, if you want to use these powerful tools, how can you do it and preserve privacy? Again, the AI landscape is changing constantly and often drastically. I’ve collected a handful of privacy-focused AI tools below that I feel comfortable recommending, given the caveats above. The most private option is always a fully local AI tool, where your queries are only processed on your device (never in the cloud) and your chat history never leaves your control. However, the cloud-based options below are about as private as you can get today. You can also use the more popular AI tools, too – just don’t give them any sensitive personal information.

Local AI tools

These are tools you download and run on your computer (Mac, Windows and Linux). They allow you to run one of the literally thousands of free, public large language models. Any given model can be optimized for speed, accuracy, size, and many other parameters. This leads to a very confusing number of choices. For starters, I would install one of the variants of OpenAI’s ChatGPT open source models – they should all have “gpt-oss-20B” in their name. If you can, choose a download from OpenAI or Ggml.org. NOTE: These models are about 12GB, so it will take significant hard drive space and take a while to download.

• LM Studio
• Jan
• Ensu

Privacy-focused cloud AI tools

If you don’t have a modern computer capable of running one of the local AI tools above or if just want the simplicity of using a web-based tool, then give one of these a try.

• Proton Lumo. You can use it immediately, with no account setup. (Note that I interviewed the head of this project in late 2025.)
• Duck.ai. This is from the DuckDuckGo private search folks. It uses popular AI chatbots (you choose which) and hides your identity.
• Confer. This was created by Moxie Marlinspike, who also created the Signal app. It requires a rather odd passkey-based setup that is flaky.

Apple has a really interesting tech called Private Cloud Compute that anonymizes AI processing, but currently it doesn’t have full chatbot capabilities. It’s used to do Apple Intelligence processing when your device isn’t powerful enough. We’ll see if this eventually becomes a competitor for the likes of ChatGPT.