If you can ask a human, you can ask an AI

Large language models (LLMs) are being used every day in a variety of writing tasks that range from emails all the way to journalism and academic research. Yet many react to their increasingly common usage as regrettable or, worse yet, abhorrent. Writers like Washington Post’s Megan McArdle have been shamed for admitting on X that they use AI as a sounding board and editorial assistant.

For some, these criticisms stem from moral reservations about using AI at all, such as concerns about copyright violations and worries about their environmental impact (though see Andy Masley’s excellent Substack for much-needed context). While these supply-chain ethical questions are important, many others believe these technologies should never be used in the writing process itself. Are these people right?

They are not. In this article, I suggest and defend an extremely simple principle for the permissible uses of AI in the writing process:

Simple Principle: Anything that is acceptable to ask a human to do is, other things equal, acceptable for an LLM to do. (Equivalently, anything it would be wrong to ask a human to do is wrong to ask an LLM to do.)

If this principle is sound, it would provide important practical guidance for how we should use these technologies to help us write.

Let’s start by considering the case of Alex Preston, a freelance journalist with the New York Times. Recently, Preston used AI to write a book review of Jean-Baptiste Andrea’s Watching Over Her. This was discovered by a reader who noticed that his review had lifted certain sections from a book review published in the Guardian. Intuitively, one should think that what Preston did was unacceptable. The Simple Principle explains why. Would it have been acceptable for Preston to ask someone to write the review on his behalf and state that he alone was the sole author? No, it would not. In both the human and LLM case, Preston is doing something that no author should do, namely, take credit for someone else’s work. Other things being equal, it doesn’t matter who Preston asks to do this, it matters that he did it.

Of course, Preston’s case is, philosophically speaking, low-hanging fruit for the Simple Principle. Let’s now examine what verdict it issues about Megan McArdle, the journalist we encountered in the introductory paragraph. It is instructive to see what precisely she was criticized for, and I copy in full the guilty Tweet:

I use AI to do research (i.e., find things to read, explain parts of academic papers I find ambiguous or confusing), transcribe interviews, generate pushback on my column thesis, suggest trims when I’m over my word count, sharpen podcast interview questions, and perform a final fact check on columns and editorials. But mostly it’s compressing the ancillary tasks to the main job: reading, thinking, and writing.

In fact, this tweet ironically does the job that a good LLM-summary does well, i.e., generate a list of the many tasks any writing project will involve.

In any case, I believe that the Simple Principle gets it right on every one of these tasks. Would it be unacceptable for McArdle to ask a human to help transcribe interviews, generate pushback on her thesis, suggest trims, sharpen questions, or perform a final fact check? I believe the intuitive answer here is also ‘no’.

The Simple Principle would, therefore, permit her to use an LLM to complete the very same tasks which are ‘ancillary’ to the main job. This point is crucial. The principle is compatible with the obviously-true idea that an author retains ultimate responsibility; just as one cannot blame a human assistant for a published error, the author remains the agent responsible for the final output.

The Simple Principle is intuitive because it is derived from a deeper idea about the purpose of writing. As the philosopher George Sher nicely puts it, ‘we put our words on paper in order to make contact with other minds’. Though these words are taken from an essay on the point of writing philosophy, it clearly applies to all forms of writing.

What matters is that the words on paper (or on a screen) are ours fundamentally. If those words aren’t fundamentally ours – but we pretend that they are – then we have done something wrong. This is true whether the words are written by a human or a large language model. If the words aren’t ‘ours’, then we should acknowledge that. We should not lead others to think they are ‘in contact with’ our minds when they are not.

Unfortunately, a great deal of stigma surrounds the use of LLMs in the writing process. If the Simple Principle is correct, then stigmatizing people who use LLMs in ways licensed by the principle is objectionable. Moreover, those who criticize the use of LLMs in writing often do so from an immensely privileged perspective – they may already be good writers, can afford to pay a human research assistant, or have a network of intelligent minds to probe whenever they want. LLMs can help level the playing field for many more would-be writers without such resources.

In short, much of the resistance to using LLMs in writing stems from this: we are actively negotiating where to draw the line rather than believing there are no lines to be drawn. The Simple Principle draws the line exactly where we thought it was.

What do you think? Let me know in the comments!