Why No AI?

Why don't you use AI?

That's a great question! There are a variety of reasons that I, like many others, believe that the use of generative or large language model (LLM) AI should be minimized or even eliminated in some cases. Others have written in more depth about the environmental impacts of AI (MIT / UCLA), the negative psychological effects on AI users (The Guardian / NPR / Stanford), and how AI exacerbates systemic injustice (HBR / Stanford / UNESCO).

For myself, the reasons I find AI use problematic in science writing are threefold: misinformation, lack of reasoning, and lack of creativity. First, AI is notorious for hallucinations and misinformation (MIT). It frequently invents seemingly plausible but incorrect facts, citations, or arguments. Humans - especially humans who are already using AI because they want to spend less time and energy on the writing process - are not especially good at closely fact-checking something that AI has written. It's very easy for AI to insert a fact into your manuscript that sounds plausible but is actually inaccurate. Given the amount of detailed, complex information in the average scientific paper, I'm convinced that widespread use of AI to edit or draft manuscripts will, and likely already has, introduced misinformation into the scientific literature. This problem may then snowball, as both real humans and AI will read published misinformation and repeat it as fact. We already know that AI has introduced hallucinated citations into the scientific literature (Nature), and citations are relatively easy to check compared to, e.g., how the configuration of a protein might impact a cellular process. The end goal of scientific research is to more accurately understand the world around us. Introducing misinformation into the scientific literature goes directly counter to this goal.

Moreover, despite the name, "AI" is not actually intelligent. Intelligence requires a process of reasoning, thinking, learning, comparing and contrasting. LLM AI works essentially by stringing together likely sequences of words (MIT). AI does not have the capacity to develop integrity, to truly understand and learn from mistakes, or even to actually "feel" something is good or bad; it simply responds to any prompts with a likely sequence of words. Part of what makes the human writing process powerful is that it allows us to make explicit arguments for our conclusions, and it forces us to reconsider and adapt when we write a sentence and then realize it isn't as strong as we initially thought. AI is, at its root, utterly incapable of the kind of reasoning required to develop sound science. Using AI in scientific writing will both result in poorly reasoned works of 'science' (see, e.g., research on AI-driven "workslop" HBR), and in scientists with poorer skills in reasoning (BBC). Choosing not to wrestle with the arguments and conclusions within your own work - leaving the writing (and editing) up to AI - will leave you vulnerable to fallacies, dead ends, and simply bad science.

Finally, AI is, by definition, a predictor machine based on previously gathered information. While it's possible for you to direct AI to mix together concepts that may not have previously been mixed together (see, for example, AI videos of topics such as "tigers grocery shopping"), AI is not capable of constructing a topic or argument that is truly novel. The advancement of science requires that we regularly examine our current knowledge and use it to create ideas and concepts that are new. It may be a trope for an area of research to be "paradigm shifting", but underlying that trope is the ground truth that our pursuit of science requires creativity and thinking outside of the box. AI, by definition, can only iterate - it can only create subtly shifted copies of previous knowledge. Relying on AI for scientific thinking, writing, and editing will produce works of science that lack real creativity. AI used without human intelligence will miss insights that could propel both fields of scientific knowledge and scientific careers.

Altogether - while AI may have some select uses that are worth the environmental costs when conducted with close human supervision, it is my opinion that the process of scientific writing, editing, and data visualization works best with no AI.

I already used AI in my manuscript. Can I still work with you?

Yes - BUT I insist that you let me know that you used AI. Depending on how you used AI, I may ask that you take additional steps to check for misinformation and hallucinated citations.

I put your writing in an AI checker, and it said it was 63% AI. What's up with that?

The problem with "AI checkers" is that they do not directly check for the use of AI. They check for patterns in writing - groups of 3, certain kinds of punctuation, rhythms in the text, etc - that are relatively common in writing produced by generative or LLM AI. These patterns are common in AI text specifically because they are common in the human writing samples that are used to train the AI. Thus, it is entirely possible that a sample of writing produced solely by human intelligence will test as likely generated by AI (University of San Diego). I conduct my editing work in word processors that document changes and can show that my work is iterative; additionally, if you require absolute proof that my work is human-only, I offer the option to send you a screen recording of my editing work on your manuscript.