An AI Scale for Writers

I.

Oh man, things are getting hot. The Hinternet Foundation will be hosting a symposium on AI, literacy, art, and the humanities in Paris this coming November, and we are beginning to fear that by then whatever any of us might say on the matter will only be able to be heard, by at least some others, as “fighting words”. We ourselves are cautious skeptics, trying to make sense of a world-historical technological shift with vast implications for most facets of human life, and to do so without rushing to any easy conclusions or facile prescriptions or proscriptions.

Our work is getting harder all the time. In just the past few weeks it has come to seem to us that AI use, culturally speaking, occupies much the same spot that masking did five years ago. That is to say that it has emerged as a razor-sharp wedge issue, and has accomplished a near-perfect division of society into two camps, both of which believe the other is absolutely, categorically wrong.

We would need to check the numbers, of course, but we also detect some significant continuity between the respective compositions of each side of the earlier pair of opposed camps, and each side of the current one. That is, those claiming today to be “AI-free” appear predictably to be the same who, circa 2021, were claiming that nothing valuable to human community is lost when you can no longer see the mouths and noses of your fellow human beings. And in turn those insisting on laxer attitudes toward AI-generated texts often seem to be the same who, five years ago, were prepared baldly to deny some fairly well-established facts of virological causation out of simple personal distaste for Anthony Fauci.

But whether or not there has been a “pro-mask to anti-AI” pipeline, or an “anti-mask to pro-AI” one, it seems to us that we should at most briefly bemoan the fact that our society, under pressure from profit-hungry proprietary algorithms, just keeps bifurcating in this way, and that then we should move on, without dwelling on the bitter memory of past theaters of our evidently endless, indeed materially unendable, culture-war.

If what we can learn about the AI controversy from the masking controversy is limited, this is in part because our newest rupture is in reality significantly less binary than the previous one, and than most involved in stoking it have so far been able to see. To put the matter bluntly, unless you have been living in a cave for the past 15 years —a real cave, that is, not a mancave, a gamer cave, or a gooner cave, which of course all have great wifi—, you are in fact “using AI”, to the extent that it is by now built into basically all of the technologies that you, as a writer, are almost certainly using to get your words out into the world.

And no matter how independent-minded you are, AI is using you in turn, in the sense that the negative style and the idiotic quirks it has dumped on the world now constitute a significant part of our default communicative frame. To some extent all this is an inevitable and predictable development in the history of technology. Madame de Sévigné’s liquid correspondence could only have poured from a plume put to paper. Frank O’Hara’s stark poems could only have come from his encounter with the partially automated language-machine known as a typewriter. And the language you have been generating, fellow writers of the 21st century, cannot fail similarly to bear the signatures of its own technological era.

It is in light of all this that we at The Hinternet have sometimes been unable to resist taking jabs at those supercilious performances of purity that we so often see on Substack, when writers declare of themselves that their work is “AI free”.¹ Ironically, it is often the writers most eager to announce the artisanal provenance of their work who are, at the same time, most faithful to a strict normative conception of good middlebrow writing that itself amounts to a sort of automation, to the extent that it outsources to society, to style guides, or to publishing-world gatekeepers, what in its fully unautomated expression would implicate only the writer’s inner voice, and the writer’s fingers.

II.

No one, we say again, is AI-free. But of course this does not mean that we may as well turn all the work of our imaginative and intellectual faculties over to the machines, or that we can now reduce the art of writing to an art of prompting. What it means, instead, is that we would be far better served by an honest scale of AI reliance than on any facetious self-reported claim of total AI virginity.

If this scale were to come into wide use, writers and readers could decide for themselves what degree of AI involvement they find acceptable. We could then dispense with the impossible ideal according to which only entirely AI-free writing is honest, meritorious, or worthwhile, and we could begin to think of the principled commitment to a minimal role for AI not as a universally binding moral-aesthetic standard governing all writing, but as something much more like the set of commitments to which the filmmakers behind the Dogme 95 manifesto adhered. Those Danes made some awfully good films, without soundtrack music, without artificial lighting, etc. Good for them! And good for Quentin Tarantino, too, for dropping in that historically incongruous but ever so sick Rick Ross track at just the right moment of Django Unchained (2012). If you don’t like this particular example, take the use of Burt Bacharach’s “Raindrops Keep Falling on My Head” in Butch Cassidy and the Sundance Kid (1969) instead, which irked the critics immensely, while undeniably demonstrating the primary conviction behind this modest proposal of ours: that there are just so many “udogmatisk” ways to make art!

So then, without further divagation, let us present our scale. We have modeled it on the Mohs Scale of Mineral Hardness, widely used in mineralogy and related fields, which until we began working on our own variant we had mistakenly been calling “Moe’s Hardness Scale”. In 1812 the German naturalist Friedrich Mohs devised this qualitative ordinal system, with diamonds at 10 and talc at 1, by observing which minerals can scratch which others, and which those others can scratch in turn. There will be no scratching on our scale, and we will move in order from lowest to highest, beginning, unlike Mohs, at zero.

So then:

0 — You actually do live in a cave, or have found some other way to go about your life entirely unconnected to the internet or to any media that indirectly transmits habits of language formed on the internet. If that is the case with you, paradoxically that also cannot be the case, since by definition you cannot be reading this.

1 — You have never so much as opened ChatGPT, Claude, etc., but you live in society like the rest of us, you send and receive e-mails, get your news from sources like the New York Times, occasionally post to Facebook or Instagram or TikTok, and, almost certainly, interact on a fairly regular basis with people who do use commercially available LLMs.

2 — You occasionally use ChatGPT for narrowly focused tasks, like finding out how the luggage-storage system works at the Düsseldorf Bahnhof, but never with the intention to share with others, under your own name, anything the LLM tells you. Over time, however, you begin to detect new idiosyncrasies in your own habits of expression. You may find yourself writing of some phenomenon or process or historical fact that there is “nothing mysterious” about it, even though no one has suggested there is anything mysterious about it, and there could be no reason at all to suppose that there is, just as you may have read some days before that there is “nothing mysterious” about the train-station lockers.

3 — You sometimes use LLMs for “bullshit” writing tasks, the ones you would rather not be required to do at all, and that you judge “beneath” you to the extent that they involve nothing of your creative faculties or your personal style — so, pretty much all work e-mails, for example, or syllabi, or grant-proposal budgets. Occasionally you recall an earlier time when you thought it important to express yourself carefully and truthfully even in such tasks as these, and you wonder whether the boundary you have sought to set up, between these bullshit tasks on the one hand, and your meaningful tasks on the other, is really so impermeable as you might have hoped.

4 — You occasionally carry on lengthy conversations with LLMs about topics of interest to you, as for example what Claude Debussy might have said if you were to go back in time and to play John Coltrane’s A Love Supreme (1964) for him. You learn all sorts of things about modal composition and so on in the process — or at least you think you learn such things, even if you know theoretically that what the LLM told you is all just so much probabilistic gap-filling. A few days later you’re writing an essay on Substack about, among other things, the evolution of jazz into a high art form in the post-war era. You try to suppress any memory of the earlier dialogue, knowing that it was unverified information and any serious writer would not content himself with that. But it’s hard entirely to suppress all the lively thoughts and speculations that dialogue triggered in you, and inevitably some traces of the ideas explored, if not of the artificial language used to explore them, enter into what you end up writing under your own name.

5 — When you used to write fiction involving, say, the third-highest-ranked women’s basketball team in Latvia, and you needed to assign a name to one of the players, you would go to the Wikipedia article on, say, the Latvian national women’s volleyball team; you would take the first name of one of the players, the last name of another of the players, and you’d consider that part of the job done. Now you go to your LLM and you write: “Give me ten plausible names of Latvian women’s basketball players,” and you choose your favorite from there.

6 — You do the sort of things described in 3-5 above, and, in addition, you sometimes consult LLMs to add ornament or grain to a sentence published under your own name. For example, you query ChatGPT as to what kind of paper Madame de Sévigné used, and then instead of writing that she “put plume to paper” you write that she “put plume to rag paper”. You know you could have got this information by other means, but you also find yourself able to judge with fairly high confidence that the training-data of LLMs have included some pretty reliable information about the history of paper-making. You have the strong sense that your judgment of verisimilitude is something like the judgment of an expert chicken-sexer, to use a stock example from the philosophy of science. You feel something like moral certainty that further reading from trusted sources would only confirm what you’ve just been told by the LLM, and anyway the stakes are pretty low here and you’re unlikely to be called out on it so you may as well just go ahead.

7 — You love crafting your own sentences and paragraphs, but you find the final process of revision and polishing tedious, and anyhow you have so many commitments that it’s just not realistic any longer to do that part yourself. (It used to be realistic, but somehow it no longer is.) So you run your text through the machine with the prompt: “Please identify any serious grammatical errors or stylistic infelicities.” You consider the problems it points out one by one, and incorporate the changes it recommends as you see fit.

8 — You’ve written a book that needs some powerful paragraphs, synthesizing the entire argument and all the leitmotifs woven through it, for a conclusion. You’re exhausted and you just can’t squeeze any more words out of yourself, so you upload the file, you ask the machine to study it and to generate some paragraphs for you. You’ve written the book in English, but you’re also fluent in Norwegian, so you ask the machine to give you the paragraphs in this latter language. Once you have them, you translate them freely into English, and the further you go, the more freely you find yourself translating, as the text itself awakens your ideas and reminds you that you do have some things of your own to say after all. The final version, which you publish under your own name, has an LLM as its causal origin, but the causal sequence is essentially untraceable and the language is essentially your own. You used an LLM as “starter yeast”, so to speak.

9 — You love your work as a writer of non-fiction books and of public lectures, but you hate the part where you also have to furnish the usual accompanying paratexts: abstracts, catalog copy, all the different types of front matter, and so on. So you do the “meaningful” stuff yourself, but turn to an LLM for the promotional padding. Fearful that someone might detect the true origins of these parts of your work, and falsely suppose that the “meaningful” part must be equally artificial, you run through several LLM-generated iterations of the various paratexts, requesting each time that the machine “humanize” it, or that it work harder to duplicate the idiosyncratic style of the main text that it has already studied, in order not to be found out. You used an LLM as “starch”, as it were, or as “caulking glue”, to hold together or to solidify, as demanded by publishing-world conventions you yourself never chose, what remains substantially your own work. At least you tell yourself that’s what you’ve done, but you continue to worry, like a naughty child, that you’re gonna get spanked.

10 — You are a lazy borderline-illiterate person who wants to make some money online. You know there are plenty of suckers out there basically as alienated from language as you are, and willing to pay to skim some trifling thing about, say, how to dazzle a literary agent with a great opening paragraph, or about how charmed a life spent in France is, and all the other usual stuff that clogs our feeds. So you go to ChatGPT and command it to write an essay for you on one of these topics. You post it to Substack and it gets 11.2k hearts, over twenty times more than the most-loved piece ever published at The Hinternet, and hundreds of comments from people declaring your work “brilliant”. A vanishingly small number of precious egghead scolds will complain that your piece was “obviously written by AI”. They will call your work “slop”, and you will cry all the way to your payment processor’s portal.

III.

We suppose that more or less everyone who is reading this falls somewhere between 1 and 9 in their current habits. We do not want to have anything to do with those who score a 10. They belong to a different and incommensurable world from ours. By contrast we sympathize with everyone in the long middle of the scale, and know, from personal temptation, what they’re going through. We ourselves have gone up to 4, and have often considered the many attractions of 5-9. In the future we would like to impose on ourselves so to speak a “Dogme 26” that would prohibit us from going higher than 3. We would continue to read other authors who go up to 9, and would be grateful to them for sharing their score with us.

Beyond such efforts at practical resolution of our current quandary, we at The Hinternet truly do not think that there is any fundamental moral problem at issue, or that any claim as to the wrongness of LLM-enhanced writing can be defended on the basis of universal principles alone. Practices evolve, and eventually they change beyond recognition. We can actively work towards an idea of what we would like our practices to be, of the direction in which we would like to see them evolve, and try to get there from here, even under the considerable weight of these new machines that are transforming language so rapidly and so indifferently to any consideration of our preferences. It would be very nice indeed if we could do so without constant moralizing and condemnation.

We are living through an absolutely bonkers historical moment. Everything is topsy-turvy, and the fate of writing entirely uncertain. None of us really knows what’s going to happen. Under these conditions it really does not make sense to be so hard on each other.

The Hinternet Foundation is a California-registered 501(c)(3) educational nonprofit. Donations are tax deductible to the fullest extent allowable by law.

1. A related phenomenon is the strange mystification of “style” that we have seen among Substack writers who are in a position, or believe themselves to be in a position, to teach others something about that slippery notion. George Saunders declares for example that “A Sentence Is a Way of Being in the World”. Please. Let’s not exaggerate. A sentence is the vehicle of a proposition. It does its work and then we move on. At least we used to move on, but now we have the option of liking (an option taken in this case 700 times, at last count) or sharing (112 times) this sentiment, which enables Substack users to affiliate themselves to whatever mystery Saunders imagines he is fingering, but to do so in a more or less automated way. “Fuck your lecture on craft, my people are dying,” wrote the Palestinian poet Noor Hindi. We applaud this intervention, though we also note that essentially any fact at all about our world of infinite joy and infinite suffering could have come, salva veritate, after the comma.