Robots can’t write. This was the conclusion I reached at the end of last summer, after taking a close look at the ermerging “large language model” that we now all know as “ChatGPT” but I was playing around with as “GPT-3” back then. I still believe that while artificial intelligence can produce impressively articulate texts it is not doing it by writing. In this post, I will not try to defend this idea; instead I want to identify three important consequences that I will take up in later posts one at a time. I just need to get these three ideas off my chest.

First, language models can’t be authors. They have no authority to state facts or opinions; indeed, they know no facts and hold no opinions. No matter how much of a text you have let an AI generate for you, you cannot attribute authorship to it. It can’t take responsibility for your mistakes and, therefore, can’t take credit for your work. AIs do not make any decisions about what they are saying, nor do they have any sense of their rhetorical situation. They don’t imagine a reader or observe any ethical relationship to them. “Machine learning” notwithstanding, they don’t actually learn anything from the criticism their output receives, they just adjust their parameters. They can’t execute what Foucault called “the author function”.

Second, since they can’t be authors, language models can’t be cited either. There is, properly speaking, no text to cite; there’s is only a record (if you keep it) of your interaction with a machine. Recently a university embarrassed itself by citing ChatGPT’s contribution to an email to its students as a “personal communication”. At the time, they probably just thought they were being transparent, but it is important to keep in mind that AIs aren’t persons and don’t communicate. Such a citation is nonsense. If you use the words that an AI suggests to you to describe something or analyse something or summarize something, you must stand by those words as you would your own. Saying you got it from ChatGPT is like saying the idea came to you in the shower. Go ahead and tell your reader such things if you want. It’s not a citation.

Finally, since they can’t be authors, and therefore can’t be cited, language models can’t be acknowledged. In our acknowledgements, we mention people who have made meaningful contributions to our work for which they are not explicitly cited, and which did not rise to the level of making them co-authors of the text we have written. The point here is that we can acknowledge contributions from entities that could have been cited in or co-authored or work. Yes, you can acknowledge friends and family members who have no scholarly authority or expertise. But the truth is that they could write a book and would then be able to claim to authorship. That’s just how it works. You can also acknowledge institutions (like your department or funding agency); institutions are named as the authors of documents all the time too. In my view, acknowledging the contribution of artifificial intelligence is therefore the first step down a road we don’t want to go toward “robot rights”. We would be committing a fateful category mistake.

I have heard people say that we should think of AI, not in terms of plagiarism (I agree), but in terms of “co-creation”. (There’s already a Wikipedia article about “collaboration” between humans and AI.) I think this, too, misunderstands the contribution that language models make. We do not cite or otherwise acknowledge the contributions of Word or Google in our writing. We should treat ChatGPT the same way. It is simply a machine we use to make our writing better.

Simplifying somewhat, your theory section either activates your reader’s expectations of your object or stimulates their curiosity about it. Your analysis (which I will talk about next week) will either disappoint those expectations or satisfy that curiosity. With this in mind, I will talk about how to write the methods section of a research project tomorrow afternoon. The important thing to remember is that there should be an interesting tension between the theory and the analysis. Your method operates in that tension.

In general, your methodology is just an account of what you did and why you did it that way. This has the interesting consequence that you can write it based directly on your own experience. As I’ve said before, you won’t tell the whole story, but you will be talking about things you did yourself for reasons you yourself determined. If you conducted interviews, you chose some people to talk to, decided what to ask them about, and then met with them, recorded the conversation, perhaps even transcribed and coded it. If you did field work in an organization you decided when you would be where and what you would be doing there. You then tried to be there and do that at the planned time. All of these actions can be described and the main task of your methods section is provide those descriptions.

One of my favorite statements of method can be found in Erving Goffman’s preface to Asylums:

In 1955-56 I did a year’s field work at St. Elizabeths Hospital, Washington D.C., a federal institution of somewhat over 7000 inmates. …

I started out in the role of assistant to the athletic director, when pressed avowing to be a student of recreation and community life, and I passed the day with patients, avoiding sociable contact with the staff and the carrying of a key. I did not sleep in the wards, and the top hospital management knew what my aims were.

Here, in plain language, we are told the conditions under which he made his observations. We can decide whether we think they are valid before we read his conclusions. This is especially important if those conclusions intend to teach us something new.

If you are going to disappoint your reader’s expectations, for example, (and, don’t worry, I’ll say more about this “disappointment” next week) you are going to need to gain your reader’s trust in your data. Otherwise the reader is more likely to reject your analysis than let you use it to challenge their theory. Since the reader has all kinds of good reasons to presume that the theory is true (as do you, I should say), it would be natural to resolve its tension with your analysis by suspecting there is something wrong with your data. You anticipate this suspicion (which is really just some healthy skepticism) by explaining, in your methods section, how you avoided the most familiar sources of error. You also point out the limitations of your method so that the reader doesn’t make too much of your conclusions (making them too challenging). Overall, you tell the reader everything they need to know to be as confident as you are about the quality and relevance of your data to your research question.

Here’s another sample from Goffman’s preface:

The limits, of both my method and my application of it, are obvious: I did not allow myself to be committed even nominally, and had I done so my my range of movements and roles, and hence my data, would have been restricted even more than they were.

Notice that he presents this as both a limit and a strength. He takes the measure of the scope of his data.

If, on the other hand, you intend to satisfy your reader’s curiosity, you have to establish your authority to relate the facts of the story. This could involve classical methodological issues, like the ones I’ve already mentioned: Who did you talk to? For how long? With what questions in mind? Did you make a recording, a transcription? Did you keep careful notes of events that you witnessed? But it can also involve something that especially researchers in qualitative fields are taking increasingly seriously, what they call your “positionality”*. Who are you to tell this story? What gives you the authority to state these facts? How did you get yourself into a position to speak credibly on the issues you have studied? We can find at least two examples of this kind of statement in Goffman:

I want to warn that my view is probably too much that of a middle-class male; perhaps I suffered vicariously about conditions that lower-class patients handled with little pain.

Permission to study St. Elizabeths was negotiated through the then First Assistant Physician, the late Dr. Jay Hoffman. He agreed that the hospital would expect pre-publication criticism rights but exert no final censorship or clearance privileges.

Note, again, that he simply and plainly describes how he got into possession of his data, and why we can trust his presentation of his results. In my view, it is an exemplary statement of method in the sense that, after we have read it, we’re inclined to trust the basis of analysis of the institutions he is about to present to us. He presents himself as both thoughtful and experienced with his subject matter. He constructs himself as a plausible authority on the subject.

_______________
*This is a relatively new trend in academia, though the “reflexivity” it implies is older even than Goffman’s nascent example. At the moment I’m reading Katja Thieme’s paper “Spacious Grammar” in Discourse and Writing 32, 2022, which has led me to Gillian Rose’s “Situating Knowledges” in Progress in Human Geography 21(3), 1997, which led me to Linda McDowell’s “Doing Gender” in Transactions of the Institute of British Geographers 17(4), 1992. I’m going to write a follow-up post of “positionality statements” sometime soon.

Probably the most controversial thing I say in my “How to Review the Literature” talk is that you shouldn’t write a literature review. (I normally credit — or blame, if you will — Ezra Zuckerman, but for some reason I forgot this year. Sorry, Ezra.) I also understand if it comes as something of a surprise to my audience. After all, my first talk was called “How to Write a Research Project” and next week’s talk will be “How to Write the Theory Section.” It would be reasonable to assume that last week’s talk could just as well have been called “How to Write a Literature Review,” not merely “how to review the literature,” but, in fact, the difference is important. I was not talking about how to write at all; I was talking about how to review the literature, i.e., how to search for it and how to read it when you find it.

Ezra is right to say that you are looking, in part, for the limits of the existing literature. It’s often suggested that you should be looking for “gaps”, but, in so far as these exist, you should actually try to fill them in with presumptions. (This idea, that theories are presumptions, is something I learned from Steve Fuller many years ago and which I’ll unpack in tomorrow’s talk.) That is, if one glaring feature of the literature in your tradition is the lack of any work on the subject that you are interested in, then you should read the literature looking for what your reader expects in their ignorance, not simply for a basis to claim that the reader probably doesn’t expect anything at all and will therefore (presumably!) be happy to learn whatever you find out. Ezra says you’re looking for the sense in which a theory is “compelling but flawed”. We might also say you’re trying to assure your readers that you think they’re rational before you tell them they’re mistaken. They have good reasons to to expect what they expect; though you think they’re wrong, you don’t think they’re stupid. Indeed, you know the story of how they came to think that way. If you hadn’t discovered the facts you did, you’d probably think like that too.

I came up with this analogy about a week ago and tweeted it. I’ve since been thinking about it, and I think it holds up under scrutiny. So I thought I’d write a quick post.

Back in November I heard Gary Marcus describe large language models as “autocomplete on steroids” and I immediately free-associated to Terence McKenna’s “stoned ape” theory which, we might say, holds that human language is autocomplete on mushrooms. That’s a bit glib, but it suggests a more serious analogy. “PC is the LSD of the 1990s,” said the acid guru Timothy Leary once (he meant “personal computers” not “political correctness”) and today he would be even more correct to suggest that AI is our LSD. Now, I don’t actually mean that there is some interesting similarity between artificial intelligence and psychedelic experience. These are very different things. But what I’ve noticed is that my reaction to AI, and the range of reactions of my peers in academic writing instruction, is similar to the reaction to the introduction of LSD into mass culture in the 1960s.

To set the scene, let me state my reasonably educated but utterly non-expert view of LSD. (I’ve checked my facts in this post using Wikipedia.) Lysergic acid diethylamide was first synthesized by Albert Hofmann in 1938, and, in 1943, he discovered its psychedelic properties by accident. It was then used in psychiatric research and treatment in the 1950s and 1960s, while the CIA also began to weaponize it for “intelligence” purposes. (Maybe you can already see the pattern!) Its use in research made it widely available to academics and their students, and it was soon adopted by the counter-culture of the 1960s. Timothy Leary was one of its most ardent proponents, famously urging young people to “turn on, tune in, and drop out,” and was described by Richard Nixon as “the most dangerous man in America”. Despite being “non-addictive with low potential for abuse” and known to “induce transcendental experiences with lasting psychological benefit,” LSD was “scheduled” in 1968, making it illegal to use both medically and recreationally. But by this time it was, as it were, “too late”; Sgt. Pepper had already been, if you will, inspired and conceived, and, indeed, released to the public.

I’m sure you can imagine many clever ways to replace the people and events in this story to produce a pretty close approximation of the narrative around GPT. But, for me, the analogy suggests a number of important things.

1. There’s no way back. Our students will increasingly use ChatGPT (and other language models) to inspire and, no doubt, produce their written assignments at university. I’m sure they’ll also find ways to use it “recreationally”.
2. Prohibition will not work. Punishing students for using it where it seems relevant to them and is technically possible will merely undermine our authority. (I’ll leave it to you to work out the analogy to the “war on drugs” in its details.)
3. Language models are perfectly safe. They are non-addictive and, so long as students continue to read and write on their own, will not damage their minds.
4. The quality of artificially generated text will improve. As will its “potency”. That is, both the style and the content of the output of language models will become increasingly effective in all sorts of applications.
5. Language models can both motivate and inspire students to produce writing they might not otherwise have come up with.

This may seem like an endorsement, which brings me to my final reason for liking the GPT/LSD analogy. Since the Summer of Bots, I have experimented extensively with GPT-3 and ChatGPT, and I have thought a great deal about it. And, if I had been a professor of philosophy or psychology in 1963, I think I might have experimented with LSD and mushrooms and DMT and other drugs, at least until they were forbidden. But…

6. I will not encourage students to use GPT to assist them in their writing projects.

That is, like LSD, while I’m comfortable with it myself, and while I grant that it has probably helped many artists and thinkers have experiences that have helped them produce interesting work to the benefit of themselves and our culture, I am not comfortable with the idea of integrating artificial intelligence into the process of developing the natural abilities of students to make up their minds, speak their minds, and write it down. Yes, I know that many students will use it anyway, and I’m not going to warn them off it, but I will not personally advise them to see how it might help solve their writing problems. I’m simply not sure I know how to use it to help them become better writers.

If they do use it to make the Sgt. Pepper’s Lonely Hearts Club Band of the college essay, that’s great! It was possible to enjoy that record without dropping acid too. But I am not going to engage explicitly with their experiments with artificial intelligence or suggest particularly effective ways of getting the most out of it. Like I say, I’m not sure I know enough about it.

Of course, if LSD had not been forced underground in 1968, there’s no telling what uses mainstream psychiatry and psychology, philosophy and poetry, would have found for it, and what place it would therefore now have in academic life. In some alternate universe, acid trips might today be familiar parts of the college (and even high school!) curriculum, as common as field trips! I truly hope that we make the best of artificial intelligence too. I hope we don’t let a moral panic awaken our prohibitionist impulses.

Let’s turn on, tune in, and stay calm.