When someone sends you a draft of their work, whether that's a technical writeup, a recipe for the best goddamn brownies you've ever tasted, a heartfelt and heartbreaking poem, or a fanfic of your favorite slash, they're trusting you to treat that work with love and respect. Unfortunately, modern software has made it harder and harder for us to uphold the trust, love, and respect with which people share their words with us. AI assistants and "copilots" can abuse that trust to train on and plagiarize work, can suggest crude and inhuman edits, and can grant copyright permissions to giant corporations that fail to uphold that trust.
This page was not written by nor reviewed by a lawyer. This page was not written by nor reviewed by your lawyer. This page is not legal advice. Any statements about terms of service, intellectual property rights, and so forth are written by concerned laypeople to the best of their imperfect knowledge; they have been written with love and care, but not with expertise.
Similarly, this page was written on a volunteer basis, and while every reasonable effort is taken to be accurate and useful, no warranty is offered to that effect.
You may see similar disclaimers offered by AI chatbots: the difference that the authors of this page care, the truth matters to them, but ultimately they're human with all the corresponding limitations on expertise and knowledge that come along with that. AI chatbots do not and cannot care, and truth is simply not a goal that they are designed to achieve.
You've probably seen the term artificial intelligence, or AI for short, a lot lately. It's become the latest tech buzzword, completely eclipsing the hype of every other buzzword in recent memory. Every product, service, and device seems to advertise AI features, so it's a bit confusing trying to figure out what AI even is. Unfortunately, this page won't be able to help a lot there — the term is intentionally confusing, and as a marketing decision. For decades, "AI" has referred to different fields of research within computer science (often related to machine learning), or alternatively, to the kinds of techniques used to make convincing characters in video games.
As of late, however, the term AI gets much more commonly used to refer either to LLMs like ChatGPT, which generate plausible sounding text, or to GANs like Midjourney, which are most often used to generate plausible looking images. Increasingly, LLMs are included as features in modern software, presented as "chatbots," "assistants," or "copilots." They may appear as more sophisticated autocomplete engines, adding in more text that sounds vaguely like the text you've already written, or may be presented as tools for performing research, and even generating whole documents automatically.
Due to their inherently random and unpredictable nature, as well as their inability to synthasize new information, LLMs are often called "stochastic parrots". Due to the immense hype around LLMs, it's often hard to get reliable and correct information about them and how they work.
The basic principle is fairly straightforward, though: large language models work by learning the correlations between different words by analyzing immense amounts of text. Words that commonly appear together in training data are thus more likely to appear together in the outputs of an LLM. Effectively, LLMs plagiarize from their training data to find text that could plausibly follow a given prompt. Notably, LLMs are not in general able to provide factual information, as correlation does not imply truthfulness, accuracy, or veracity. LLMs are sometimes said to "hallucinate" as a result; this page avoids the term, however, as it describes how LLMs inherently work, and not a fault or flaw in their operation.
Both LLMs and GANs rely on immense amount of computational power, and require truly staggering amounts of data to create (or "train") in the first place. Whereas humans can learn to write and talk from a relatively small number of examples, LLMs are often trained on huge parts of the entire accessible Internet; far more than a human could ever reasonably read. As a result, data that can be used to train LLMs — that is to say, examples of human-written text — have become incredibly valuable. Sites like Reddit and Stack Overflow have made much of their recent revenue by selling user posts as training data, while many pieces of software you might use on a day-to-day basis can siphon off your data for training new LLMs.
In some recent versions of Microsoft Office, Word includes copilots by default. Your terms of service with Microsoft not only allow for this, but may allow Microsoft to use documents that you open to train new versions of AI chatbots, autocompletion engines, and copilots. If it's your words that you're sharing with Microsoft, and you go in knowing that's what could happen with your work, by all means — please don't let anything on this page dissuade you from doing what you want and consent to with your words. Microsoft Word is, for all the faults highlighted here, a mature product with decades of expertise behind it, and may fit your needs perfectly.
The problem, though, is that not everyone who sends you their work may necessarily consent to having their work shared with Microsoft in the same way. Your agreement with Microsoft is indifferent as to whether the documents you open with Word are your own or someone else's. Even opening a draft to provide comments may promise rights to that draft over to Microsoft, depending on where you live, what version of Word you're using, and what you've agreed to. For some people who might share their work with you, that's at most an annoying detail; for others, that might be a hurtful breach of the trust they've placed in you.
Technology shouldn't work that way. Just opening a document in your favorite text processor shouldn't be the kind of thing that can get your friends and colleagues hurt. Sadly, that is the case for the monopoly-driven world of tech in general and word processing software in particular. Navigating that is harder than it should be, and places an unfair burden both on the people writing work that they'd like to share with you, and on you in reading that work. It really sucks, and there's no getting around that right now. But being thoughtful and caring with how you approach the complexity of sharing and communicating with others in a monopoly-driven world of tech is worth the extra effort, the empathy, and the love that you put into it.
While Google Docs doesn't currently demand the same level of ownership over work that you open in Docs, they do claim AI training rights over any documents that you make public. It's easy to accidentally make a document in Google Docs public, making it more difficult to respect the trust placed in you by your friends, loved ones, collaborators, and respected enemies.
This is where this page gets particularly bleak: there are alternatives to Microsoft Word and Google Docs that can help you respect the trust and love that has been placed in your care, but there's not as many as would be nice, and they all come with tradeoffs that can make your life more difficult.
Primary amongst those alternatives is LibreOffice, a free and open-source suite of office applications that works on Windows, macOS, and Linux. Unlike Word and Google Docs, LibreOffice does not claim ownership over the documents that you open.
While LibreOffice does not have a mobile version, tools like Collabora Office can help you avoid AI training when using your phone or tablet.
Outside of LibreOffice and Collabora Office, some competitors to Microsoft Office and Google Docs may offer better terms of service. As of this writing, while Zoho Writer does sell AI-based products, their terms of service do not currently appear to claim ownership or training rights to documents opened with their services.
We got to this state of affairs because Word is an effective monopoly in the word processing space. Everything must be compatible with Word, must open nicely in Word, format correctly in Word, use fonts that ship by default with Word, and so forth. Word has an incredibly out-sized influence over how we communicate and share our work with others.
As a result, being intentional about how you reinforce or oppose that monopoly is important, and can help make it easier for everyone to choose something other than Word.
The easiest step that you can take is to consider using formats other than Word's own *.docx format. International standards bodies have settled on a different format, known as OpenDocument, as it's easier to share, archive, and preserve OpenDocument files. Word, Google Docs, and LibreOffice all understand and can interoperate with OpenDocument text files, using *.odt filenames. When saving and sharing your work, making sure to choose OpenDocument even if you also provide a *.docx can help reduce and oppose Microsoft's monopoly over the written word.
It's easier to avoid AI training and corporate monopolies when we work together. There's power in letting others know that your boundaries are important, and that there's alternatives that can help people respect your boundaries.
Consider adding a link to https://nopilot.xyz to your e-mail signature, your social media bio, or other places where you might share your work.