AI Writing and Forgeries: Why Nobody Wants to Read Your Slop

tl;dr - AI slop signed as your own writing breaks the social contract of reader and writer. It’s also, effectively, a forgery.

Chamath Palihapitiya is being accused of “Claude-posting.” New Brunswick MLA Bill Oliver appeared to read an LLM’s editing instruction aloud in the legislature. And when readers accused Kate Gilgan’s New York Times “Modern Love” essay of being AI-written, she denied pasting generated prose but acknowledged using several chatbots to conceptualize and edit it. Meanwhile, AI-generated essays are winning prizes. What should we make of this? Why don’t people want to read AI “slop”? I’ll hypothesize why and connect this to Denis Dutton’s work on forgeries in art. ¹

First of all, what is “slop”? When I say “slop,” I mean text primarily written by AI but with a human byline. Most readers familiar with AI will know its characteristic tics, like the frequent use of em dashes (—) or negative parallelism (“it’s not just X, it’s Y”).² Usually, the author generated the first draft through an AI program like Claude or ChatGPT. Maybe they seeded it with their own ideas, but AI does the actual writing. Slop happens when the primary “pen to paper” of writing sentences is delegated to AI.

Edit: It wasn’t clear in the first version of this essay, but my argument doesn’t rely on the AI writing being bad. That is a current problem, yes, but those arguments are both well-trodden and inherently subjective. Plus, future models may write as well as humans. So for this essay, I’m not assuming that AI writing is fundamentally poorer quality than human writing.

Slop can appear in both creative writing and “functional” / nonfiction writing, but I’ll only focus on the latter. Examples: posts, memos, essays, and blogs. AI writing has taken a foothold in these domains. At some big tech companies³, leaders have been caught pushing AI-generated memos to audiences of thousands. Their motivation is clear: if AI makes knowledge work more efficient, why not use it? Software is already being automated, so why shouldn’t large language models be used to produce language? Besides, a user might think, with functional writing the output quality doesn’t have to compare to Huckleberry Finn.

If that’s true, why do I, and many others, find AI writing so awfully offensive? People online claim that we’re developing “antibodies” for AI writing⁴. First, there’s a basic asymmetry between the writer of the text and the reader. In the past, the writer had a burden of production that was at least comparable to the burden of consumption. As a result, the reader and writer enter into an implicit social contract. This contract states that the writer of, say, a report on widget production took the time to understand the widget supply chain and its constraints. I assume that the author took the subject seriously, if only because of the time and energy it took to produce the piece. If a human author (pre-AI) slapped together words with no understanding of the subject, I’d be similarly angry at this contract’s breach. Whether or not I agree with a piece, I respect that the subject wormed its way through the author’s head. AI writing inverts the effort asymmetry, since the author can blast out slop while expecting readers to read the old-fashioned way.

Incidentally, the effort doesn’t have to come at the moment of writing. The widget expert that puts together an insightful essay in 10 minutes may have spent years “behind the scenes” understanding widget production. Similarly, a precocious youngster who breezes through high-quality writing is often doing an accelerated version of the same burden. AI is not; perhaps it is expending effort, but in such an alien way that our contract on it is broken.

Put more simply: the author couldn’t be bothered to write it, so why should I bother to read it?

Perhaps if the human “authors” didn’t sign their name to the writing, it wouldn’t be so bad. However, signing your name to slop is a form of forgery, and I think many “functional” writers miss the perfidiousness of this. Dutton lays out the problems with forgery in his essay “Artistic Crimes”; in particular, we care about the “performance” of the piece itself, i.e., what the writer went through in writing. For example, the Dutch painter van Meegeren forged Vermeers with exquisite care to prove a point about art criticism: dismissed by the critics as a painter, he planned to unveil his forgery after they had lauded it as a Vermeer. He was actually too successful; his first fake proved so profitable that he told no one and sold a dozen more. He only unmasked himself after World War II, when one of his “Vermeers” turned up in Göring’s collection and he was prosecuted for selling a national treasure to the Nazis. Even then, some critics refused to believe that an accused criminal had painted an acknowledged masterpiece, and his claim was only definitively confirmed by dating techniques developed years later. Ultimately, Dutton argues that critics are right to value the forgery less, as it belies the true work that went into the painting. Similarly, to write an AI essay with no attribution is to lie about the “performance” of the writing. This is also what writers of memos or other “functional” writing get wrong. Even business writing is not the mere transfer of facts between minds; it has an implication of performance.

Despite calling it “slop”, framing it as forgery doesn’t require the writing to be poor quality (although it often is). This addresses the question: if you can never tell that the piece is AI-generated, does that negate the entire argument? No, for the same reasons that Dutton lays out in his essay: the “wrong” of the forgery exists whether or not it’s caught. Van Meegeren’s copies, despite their high quality, are truly less compelling as Vermeers than real Vermeers whether or not Van Meegeren was ever uncovered in the lie. In other words, “nobody wants to read your AI slop” is meant as an intrinsic property of the work as forgery, not a measurement of audience response.

One could object to the slop-as-forgery argument on gradualist grounds: how much AI-assistance is too much? If you use AI as a first-pass critic⁵, is that slop? What about editing individual words or sentences, rather than generating a whole piece? These are difficult questions, and they force us to distinguish between tools and independent agents. Famous painters used assistants in their studios to help them finish tedious portions of work; they were often “tools” of the principal.⁶ Fully distinguishing between a tool and an independent agent is a challenging problem, and I won’t attempt to do it here. Rather, I claim that the level of technology as of Summer 2026 and the version of assistance I often observe, where AIs write everything from paragraphs to entire pieces, is too much. That is, I can’t tell you precisely where the line is, but I know that slop essays have crossed it.

AI could have genuinely good ideas (why not, if it can prove Erdős problems⁷). Would I read an AI-generated essay if it were signed under the AI’s name? Not today, but only because of their typically low quality. The combination of tics and superficially-right-but-ultimately-wrong insights has left me not trusting the models’ writing on anything substantial. However, I’m open to the possibility of their improving. Like any human author, the model would have to win (back) my trust in order to persuade me to read the essay. Presenting AI with its own byline could be a road to making its insights more palatable. As a reader, I choose to enter into the “contract” on honest terms. Who knows? Perhaps if AIs are proven sentient, they could even pass Tolstoy’s test⁸ for emotional experience in artists. But until then, I beg you: stop writing Claudeslop.