Is Using AI Autocomplete Cheating?

If you are a student, a journalist, an author, or just someone who wants or needs to write, you have a lot of dubious ways to integrate AI into your process.

You could, of course, use AI to generate a work whole cloth. You could use it to “rewrite” what someone else wrote.  Finally, you could use AI to edit your own work, including completely rewriting your content.

Then there are a number of (largely) ethical ways to use AI. AI proofreading and spell checking are largely accepted. AI can also be useful in organizing thoughts, brainstorming, and finding sources (that you need to verify and cite).

However, a new contender has emerged. It’s called AI autocomplete, and it’s billed not as a writing tool but as a typing aid. It claims to be able to help even the fastest typists write more efficiently.

But is it cheating or a form of AI plagiarism? That depends both on how you use it and the environment you use it in.

Understanding AI Autocomplete

Disclosure: I have paid for a discounted one-year subscription to CoTypist. I was also a beta tester for the product. However, I have also tested others mentioned in this article, including CoTabby and KeyType.

Most people are familiar with autocomplete. Whether on a mobile device or a computer, when one starts typing a word, their device will often offer a list of possible words or short phrases to finish it with. This is particularly useful on mobile devices, where the small keyboard makes it difficult to type efficiently.

Many of these autocomplete tools are already powered by AI. However, most also existed before AI, and for regular autocomplete, a full LLM is not necessary.

AI autocomplete takes autocomplete a big step further. Instead of suggesting the next word or short phrase, AI autocomplete suggests ways to complete a sentence or even start the next one. 

The current standard bearer for AI autocomplete desktop is CoTypist, a Mac application that adds AI autocomplete to your device. However, commercial competitors such as Typeahead have emerged. There are also open-source alternatives such as CoTabby and KeyType, both of which are in the early stages of development.

However, the idea is broadly the same across all of these tools. With these systems, you set up a local AI model on your device, you train the model on your writing, and then it begins to feed you suggestions as you type. It learns your writing style and attempts to predict what you would organically type next. You can then accept the changes by pressing a button, usually tab, or reject them by continuing to type.

Ideally, an AI autocomplete tool should not make editorial decisions. Instead, it should help you type what you were already going to say. That said, in my testing/use of CoTypist, I found that I did accept some “wrong” suggestions that were either “close enough” or “better than what I was going to say.”

In most cases, the recommendations are limited to a small number of words. However, with these apps, you can request longer recommendations, up to a dozen words or so. That said, the longer the recommendation, the less likely it is to be accurate, and the more slowly the recommendation will be generated. Because of this, most apps recommend keeping the suggestion length short, usually around 4-5 words. 

But does that make a difference? Over the past month, I’ve been experimenting with a variety of AI autocomplete tools to find out.

My Experience Using AI Autocomplete

Though I’ve tried CoTabby, KeyType, and CoTypist, I’ve only had significant success with CoTypist. As of this writing, it’s by far the most developed of these tools for macOS. That said, I expect that to change over the coming months.

Helpfully, CoTypist keeps track of the suggestions you accept and lets you know how many words you’ve saved. Most days I used it, between email, two articles here, and other projects, I would save around 500 words per day. However, this didn’t translate to a large amount of time saved per day.

I am a fast typist and can average around 100 words per minute. If we just factor in the pure typing time, that’s about five minutes saved per day. But that’s also not the whole story. It takes time to read suggestions and decide whether to accept them. I also semi-regularly accepted suggestions that I later deleted, costing me time. 

However, this also possibly undersells the time saved. In some situations, the AI autocomplete suggestions knew what I wanted to say before I did. This was especially true when completing sentences. That was when the suggestions were most useful and were common points that I would stop to think about what to say.

As a result, it’s difficult to say how much time was actually saved. Simply put, the variables that determine how long a task takes are simply too difficult to eliminate. Some things take more research, more time to think, and are more difficult to format. That had a much bigger impact on the time spent than the AI autocomplete.

That said, I did find that I preferred having the tool available than not. Especially once it was adequately trained on my writing style, it felt like a useful tool, even if the time saved wasn’t as significant as I had hoped. That said, the days I didn’t use it did not feel significantly different. 

But that raises the question: Is it cheating? I think that the answer is complicated.

Is AI Autocomplete Cheating?

As with most things AI, the answer to this question depends heavily on where/how you use the tool and whether you disclose its use. Obviously, using this in a typing class would be a significant problem. It would also be an issue with students who are learning to write as they are still developing their own voice, and AI autocomplete could hinder that process.

But what about regular essays, research papers, articles, or even books? I think the answer gets much murkier.

In my tests, I never felt that the tool was trying to take away my authorship. Instead, it felt like it was trying to guess what I would say next and getting it right a decent amount of the time. But, while that’s true for short phrases, what happens when those phrases get longer? Instead of 4 or 5 words, what happens when AI autocomplete suggests a whole sentence or entire paragraph? The technology is only going to improve, and those predictions and completions are going to grow in length and accuracy.

This is another example of an AI gradient. However, this time, it’s just one tool rather than multiple approaches to the whole of AI. At what point does AI autocomplete (or autocomplete in general) go from being a typing aid to a form of AI plagiarism? That’s an impossible question to answer. At what point does AI go from being an assistant to a co-author? Is it four words? Seven? Twelve? A hundred?

There is no easy answer to this, and it becomes almost a question of philosophy rather than a technical one.

That said, I did run articles where I used AI autocomplete and ones written at nearly the same time through five different AI detection tools. All five tools reported that both sets of articles were human-written. There was no difference in the results or the certainty of the results between the two sets. 

Ultimately, I think that this is more of a future issue than a current one. Today’s AI autocomplete tools are largely extensions of what existing tools have been able to do for many years. They don’t raise the same issues of authorship that purely generative AI does. 

However, that doesn’t mean that there aren’t spaces where AI autocomplete should be barred. But, as with many things AI-related, it’s something that has to be decided on a case-by-case basis.

Bottom Line

In the end, I’m going to continue to use/experiment with AI autocomplete tools and follow this space. I’ve updated my AI disclosure policy to reflect this change. However, I’m going to be more interested in how these tools grow and improve in the future.

Right now, I think that the positives outweigh the negatives when it comes to AI autocomplete. Though my benefits have been modest, there are those who struggle with typing (for whatever reason) and could benefit from such a tool.

That said, this is an interesting picture of what the future of AI might look like. These services don’t rely on an internet chatbot or a big datacenter. Instead, you run the model on your own device, and it uses the data there to generate a response.

To be clear, this isn’t new. There’s been a long-running push for on-device AI. Apple and Microsoft both have made strides in this area. Tools like LM Studio have long made it easier to run AI models locally.

All in all, the future of AI may be less about big data centers and more about on-device capabilities. That presents a new set of opportunities and challenges, especially in education. 

It’s something that decision-makers need to start thinking about now, because tomorrow might be too late.