In February this year, a safety researcher named Mrinank Sharma resigned from Anthropic with a long, public letter. Resignation letters from AI labs are a genre of their own now, and this one was widely shared and picked over. One reply stood out, from a developer named David Minnigerode posted on Threads: “Sorry that is. definitely tl;dr. But also kinda ai;dr Some of those sentences…yeesh.”
Two days later, a pseudonymous developer’s blog post titled “ai;dr” reached the front page of Hacker News. The whole argument is in one line: “Why should I bother to read something someone else couldn’t be bothered to write?” Futurism covered it within the week, and Fast Company declared that we’re now in the era of “AI unless proven otherwise.”
The term wasn’t all that new in February 2026. User holothuroid proposed it on Mastodon in September 2024, as a joke about extending the tl;dr standard (as well as suggesting pw;dr for paywall, didn’t read – which I think will be attractive to both my academic colleagues, and anyone who has tried to read a mainstream Australian newspaper online…). It took eighteen months for the phrase to really catch on though, and the lag itself tells us something: a shorthand for refusing to read machine-suspected text only becomes useful when you suspect it everywhere.
I’m fascinated by the impact of GenAI on writing and communications. I’ve written about authorship and synthetic media throughout my PhD, from digital plastic through to the effort economy of slop earlier this year. Those posts were largely about the producer’s side of the bargain: if the effort to consume what you’ve made exceeds the effort you put into making it, you’re producing slop.
AI;DR is the same deal, but from the reader’s side. It’s what happens when readers decide, in advance or in media res, that the contract between author and reader has been broken, and they check out.
Signs of the times
TL;DR is a judgement about cost: your text asks too much of my time (and I don’t have the attention span). AI;DR is a judgement about provenance: nobody paid the cost of writing this, so I owe it nothing. The judgement gets made based on surface features – linguistic tells that suggest to the reader that AI has been used in the production of the text. Words like “delve” and “navigating” (the Age of AI, usually) have becomes memes in their own right since the release of ChatGPT. Tidy lists of three, em dashes, and other rhetorical and grammatical devices are under a microscope. Wikipedia’s editors keep a detailed catalogue of the signs of AI writing, built from years of cleaning up machine-generated articles.
The Wikipedia page warns that none of these signs proves anything, because the models learned all of them from human writing, and that the list is “descriptive, not prescriptive.” AI signs are, however, genuine at the level of populations: a study of 15 million biomedical abstracts found the word “delves” spiking after ChatGPT’s release, and estimated at least 13.5% of 2024 abstracts had been processed with LLMs. But the same researchers are clear that you cannot take a population-level pattern and use it to convict an individual text.
“Delve” was a perfectly ordinary word before 2023, and according to the Guardian’s reporting on how outsourced human feedback shaped the models’ vocabulary, the use of Nigerian business-English trained classifiers shaped the output of early GenAI models as much as the training data itself. Unfortunately, whether humans are starting to sound like AI, or AI is starting to sound like a very specific demographic of humans, it doesn’t seem to matter. Once a text is branded as AI, there’s little the author can do to claw back their reader’s attention.
Sophisticated use? Think again
Think about an author who uses these tools the ways we’re told a skilled writer should use them. The ideas are the author’s. They feed the model their own notes and research as context. They dictate and transcribe rough sections, ask for suggestions, discard many of them, and rewrite until the piece sounds like their usual writing. Studies of real co-writing sessions show these patterns of granular, iterative, writer-controlled work, with hundreds or even thousands of small decisions about what to accept and reject.
From the perspective of AI;DR, none of it matters. The “sophisticated use” of GenAI happens entirely in the hands and the head of the author; it never reaches the page or the screen. This reminds me of a study from the 90s in which participants were asked to tap out a song they were hearing in their heads for listeners to identify. In Elizabeth Newton’s experiment, the tappers picked a well-known tune like Happy Birthday or The Star-Spangled Banner, and the listeners had to guess it based only on the tapping. Despite tappers estimates that 50% of the listeners would get the tune, just 3 of 120 attempts were correct. The melody existed only in the heads of the tappers, and the attempt to communicate it broke down entirely.
With GenAI, the author’s intent and their use of the technology are almost irrelevant. The moment a reader spots what they believe to be a sign of AI, they switch off, regardless of the care taken during drafting and editing. For example, in a series of experiments at Cornell and Stanford, researchers presented people with Airbnb host profiles and told them some were AI-written. Profiles that participants suspected of being machine-written were trusted less, and participants largely agreed on which profiles to suspect, because they shared the same “folk theory” of what AI text sounds like. Every profile in the study was written by a human: it was the suspicion which did all the damage.
The pattern can be applied across many contexts and kinds of media. News labelled as AI-generated is rated less trustworthy even when readers can’t identify anything wrong with the articles. Identical artworks are valued less under an “AI-created” label. When readers see an AI label, they assume the whole thing was automated, with no human judgement anywhere in the process. And disclosure, the thing every academic integrity policy that allows AI use asks for, makes it worse rather than better: across thirteen experiments, people who disclosed their AI use were trusted less than people who said nothing. Writers seem to sense this. One study found users don’t feel like the authors of AI-generated text, but decline to say the AI was involved. So, both the careful hybrid writer and the “lazy” prompt-paster receive the same sentence from readers with an eye for what they determine AI slop.
Bad evidence with real consequences
It would be one thing if readers were actually any good at detecting AI, but much of the time they aren’t. In early controlled studies people distinguished AI text from human text at close to chance, and the cues they relied on were predictably wrong. Detection software fails at scale too: none are accurate or reliable enough to act as the sole source of “truth”, and the false positives land on non-native English writers and autistic students with structured writing styles far more often than on anyone else. With around two-thirds of teachers using detection tools, even small error rates add up to a lot of wrongly accused students. Some of those students now keep version histories of everything they write, evidence of their own humanity filed in advance, just in case. I wrote everything I could ever want to write on these tools back in early 2024, when I said that AI detectors in education are a dead end.
Meanwhile, the signs themselves are getting less reliable, and from both directions at once. Writers who know the most obvious AI tells are stripping em dashes and flagged vocabulary out of their prose, to the genuine grief of em dash loyalists. At the same time, human writing that involves AI anywhere in the process is drifting measurably closer to the machine register. There’s an arms race happening between the “sophisticated users” and the equally sophisticated readers, and it’s driving a wedge between people online in increasing volume.
I notice the same gut reactions in myself, for what it’s worth. Because I’ve been neck deep in these technologies for a long time, I fancy myself something of an expert in spotting AI use beyond the obvious tells. In truth, my “by eye” detection methods are probably just as useless as everyone else’s. But there are certain tropes from certain models that send me crosseyed, immediately zoning out and scrolling past text even if I respect the author or the topic. For me, the most eye-glazingly tedious of these come from what I call “Claudespeak”.
Claudespeak isn’t just a vocabulary or a set of rhetorical devices. By whatever obscure training methods they use (and possibly on purpose, to convey the sense of personhood they’re obsessed with), Anthropic have created an entire library of idioms unique to their model. Claude will tell you everything is “the move”, whatever that means. It hedges the start of almost every paragraph, with painfully lame sentences like, “Here’s the part that I think matters most for anyone who writes, and it’s the uncomfortable part” and “Here’s the idea I want to put on the table”. I have a suggestion, Claude: rather than telling me you have an idea that you’re planning on putting on the table, just put it on the damned table.
Linguist Kimberly Pace Becker calls this Claudish, and compares it to other linguistic tells from across models. I like Claudespeak because of its resonance with duckspeak, from George Orwell’s Nineteen Eight-Four:
“It was not the man’s brain that was speaking, it was his larynx. The stuff that was coming out of him consisted of words but it was not speech in true sense: it was a noise uttered in unconsciousness like the quacking of a duck.”
George Orwell – 1984
Whichever way you look at it, the web genuinely is filling with slop, and refusing to read suspected slop is a reasonable way to protect your attention. That’s what makes this situation so uncomfortable: every individual reader is behaving sensibly, and the collective result is a reading culture that convicts real writers on punctuation and word choice.
Photo by Aaron J Hill on Pexels.com
Broken contracts
Where does this leave writers, teachers, and the rest of us?
For conscientious writers who use the technology, the question “was AI involved in this text?” is so complex it’s unanswerable, and readers can’t verify the answer anyway. Maybe the important questions were answered a century ago in the politics of ghostwriting: whose thinking is this, and who is accountable for it? Audiences have long tolerated speechwriters and ghostwritten memoirs when the ideas and the responsibility genuinely belonged to the named person, and they’ve reacted badly when that condition failed and the contract between author and reader is broken.
For educators, the uncomfortable news is that this suspicion economy is already operating in classrooms, and in both directions. Students scan their teachers’ work for signs of AI; teachers run student work through detectors that demonstrably discriminate. I’d rather we spent that energy making the writing process visible and valued, which is good pedagogy anyway, than perfecting our forensics, which both the evidence and our hearts say cannot work.
And for readers, which is all of us: maybe hold off on the verdict. I know I’m too quick to glaze over at Claudespeak, and maybe that’s a me problem. I’m a full-time author who also works extensively with GenAI, and even my hunch that a text might be AI-generated isn’t solid grounds for dismissing an entire piece. AI;DR is a fair response to slop. Aimed at a person, it’s a judgement made on shallow evidence, delivered with no right of reply.
The old contract between writers and readers rested on something nobody ever needed to say: a real, human person was here. That assumption is gone, and I don’t think any detector, watermark, or Wikipedia vocabulary list is going to bring it back. What’s left is the older, harder currency. A name, attached to words, by someone willing to answer for them.
Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch:
Leave a Reply