Suggestions are proffered in roughly descending order of “difficulty.”
An asterisk signifies that it is, to my mind, the best option for multiple reasons and by a significant margin, but may be annoyingly technical and slightly more “risky.” A dagger (†) signifies that it's what I use.
Warning: this document spills over into general-purpose opining; the first bolded option in every list is a good enough start if your interest exclusively pertains to the project of the title. Second warning: I am, by training and disposition, a humanities pleb; all technical explanations are reductive and provisional.
Misc.
Word processing. I'm of the opinion that most word processors are, minimally, overtuned for the purposes we writers mostly use them for. (Maximally, they’re bad.) An application like Microsoft Word (or, to a lesser extent, Google Docs) is built for formatting and typesetting, not writing; I regard it as closer, philosophically speaking, to something like Adobe InDesign than to, say, a physical notebook. That's why these programs can feel so overstuffed with features and irritating to navigate. And one of my fundamental beefs with contemporary software is that it’s cluttered, and clutter is bad for thinking. If interested, I would recommend considering tools beyond the Google Docs/Microsoft Word modality, though there might be a bit of a learning curve. (At the very least, I think it’s important to know that there are decisions undergirding every aspect of every piece of software, including word processors.) Anyway, I use OpenOffice for finalizing formatting and Emacs for everything else.
Plain text. One of the advantages of Obsidian is that it is basically an organizational interface for dealing with plain text files, though it does feature basic formatting options. (Formatting is handled by way of a “markup language” called Markdown, but you won't have to worry about that much in practice, unless you want to.) The advantage of plain text files is that, as long as the encoding standard is accessible, everything can read them and the will be readable as long as computers exist. There are plain text files on the internet identical to those read on computers in the 1970s; a plain text file made today would be readable on an Apple II. As long as you have them backed up on reliable local storage media, there is nothing Microsoft or Apple can do to take them away from you.
Email clients. An email client is an application you use in order to interact with emails. This is different from an email provider, which (a) takes care of all the weird stuff necessary to stay ahead of spam and (b) hosts your emails on their server. For NCSU email, Google is the provider, and the Gmail.com website is the client. The impact of this distinction on how to think about avoiding AI requires a bit of explanation.
Modern AI is based around specific models. These models are extremely complicated predictive algorithms, the most famous of which were created by engineers who receive nine-digit signing bonuses, though they were refined into basic functionality by thousands upon thousands of undercompensated and precarious gig workers (especially in the Global South, but also in places such as, oh, I don't know, Hendersonville, North Carolina), then packaged and released.
When I open a ChatGPT account for the first time and begin engaging with the chatbot, I am interacting with the GPT-5.2 model in its basic form. Every user starts from the same point. However, as soon as you begin to interact with ChatGPT, it begins to “learn” from your specific prompts, altering its performance in sophisticated and unpredictable ways.
What this means is that, for the purposes of most applications of AI, there are two training phases: the primary, pre-consumer phase and the individualized, consumer phase. Most major AI platforms at least claim that data from the personalized phase are not feeding back into general-purpose training. Given that LLMs are prone to spitting out verbatim transcriptions of training text, this is a truly basic, truly necessary privacy feature.
As it stands, Gemini, Google's AI model, is integrated into the Gmail client. When it interacts with your emails, it does so within the context of that second, personal phase of training. The barrier between the two is allegedly intact.
Because these companies are insane liars, broad skepticism of FAANG privacy claims is fair play (and these companies are already horseshit with privacy stuff). Personally, I see no reason to doubt this specific claim just yet. Still, Google recently announced plans to expand this personal training to every service in the entire Google ecosystem. Given that consumer AI services remain catastrophically unprofitable, and given that LLMs have already trained on the entire extant internet, I think things are going to get even more chaotic and ugly very soon, and who knows what that will entail.
The takeaway for email: if (or when) Google uses (or admits to using) your personal email for AI model training, the contents of your emails will still be visible to them no matter what client you're using (unless both you and the senders/recepients of your sensitive emails are jumping through difficult encryption hoops I do not personally understand). If you're concerned with the eventual fate of your emails as a body of training data, an email client is not going to solve that problem; downloading and deleting your emails off of Google's servers would be a better bet, though even that is no guarantee. But it's nice as a “get this stupid bullshit out of my face”-type intervention, which I think is extremely (and underratedly) important.
Created: 2026-01-27 Tue 20:44
Last edited: 2026-01-29 Thu 11:18
Originally converted from org-mode to HTML with ox-hugo