A note before we start. I run Growth at Substack. What follows are my opinions and not an official company stance.

We are fast approaching the tipping point. People are starting to show up with their pitchforks like a mob descending on city streets chanting, “Hey-hey! Ho-ho! The AI slop has got to go!” In today’s post, we break it down: how prolific is the problem, when is it ok to write with AI, when is it not, and who should be the arbiter?

These questions have teeth now, too. As I write this in the first week of June 2026, New York’s synthetic performer law is taking effect, the FTC has stood up a dedicated AI enforcement unit, and a small industry of detectors will grade any paragraph you give it on how human it sounds. This is totally understandable. Last year, around half of all new articles published online were machine-written, and complaints about AI slop had gone from a niche annoyance to a wail the whole internet was making at once. The flood is real, but IMHO (in my human opinion), there’s more nuance than the visceral panic suggests.

Share

When scrolling on LinkedIn, you know what I’m talking about – the massive influx of self-proclaimed thought leaders posting self-congratulatory notes that are suspiciously packed with em dashes and fluffed with “this, not that” antithesis phraseology. An early 2026 conducted by Originality.ai (an AI detection company), analyzed 3,368 long-form LinkedIn posts published in 2025 from 99 influential profiles across 11 industries. 53.7% were classified as “likely AI-written,” a 189% surge after ChatGPT launched.

There’s something disquieting about it. Once upon a time, bots were obviously not people, but since when did people become bots?

My contrarian view

We all want less AI slop in our lives. But the go-to solution commonly cited is publicly shaming content when it is suspected to be AI-generated and suppressing it. I want to argue the opposite. I think the disclosure is aimed at the wrong person. A piece of writing that is trustworthy and engaging should not become less so when you find out how it was made. The disclosure the reader is demanding is an answer to a feeling rather than a fact. There is a real merit to having AI identification that would raise the quality of what gets published. But it belongs in front of the writer, not the reader.

The trouble is that “should you disclose AI” turns out to be the third question in a line, and we keep skipping the first two:

• Is using AI to write actually a bad thing?

• Whether or not it’s bad, can a reader tell?

Is it bad to use AI to write? The effort argument.

Imagine 2 high school students writing the same research paper, 3 times, 30 years apart.

• 1996. The assignment means going to the library. Both students go, where one reads the sources, takes notes on index cards, builds a bibliography, drafts an outline, and writes across a week of evenings. The other skims one encyclopedia entry, invents two statistics, attributes them to a journal he has not opened, and turns in something thin by Thursday. A teacher can easily tell them apart without knowing anything about how either was produced. The level of effort each student put in is obvious.

• 2011. The year I actually finished high school (congratulations, now you know my age). Instead of checking out books from the library, the good student used Google, opened real web sources, used CTRL+F to find the relevant passage on the page, checked the claim against a second result, and cited it. The lazy one pulled the first paragraph off Wikipedia and called it research. The barrier to looking informed had dropped to almost nothing, and the predictable thing happened: the expectation rose to meet it. A paper that would have read as diligent in 1996 read as ordinary in 2011, because now everyone had the library in their pocket, so the standard moved up. (It is worth remembering that this was a genuine moral panic. Schools and colleges spent 2007 and 2008 formally banning Wikipedia because it was “made research too easy.” Nobody footnotes Google now.)

• 2026. Let’s run the parable for a third time today. The good student writes a careful prompt, asks for research, reads what comes back with suspicion, checks the facts that matter, throws out the boilerplate, supplies the actual argument, and shapes it into prose a person would want to read. The lazy one types one sentence into a GPT, copies the output, and submits a gray slab of “it’s important to note that” and “in today’s fast-paced world.” One of these is good work and one is not. The tool did not change which is which. The constant across all three eras, and what has survived every collapse in the cost of producing words, is the effort behind them. A disclosure label cannot decipher this.

This is not the first time, either. Just as Wikipedia made research too easy, the pocket calculator made arithmetic a relatively brainless affair. The reactions to both were identical in shape. When affordable calculators reached classrooms in the mid-1970s, the fear, as one history of the period catalogs it, was that “students’ computational abilities would be ruined,” that they would grow dependent on the machine and never learn to estimate or carry a number in their head. Several states banned calculators from standardized tests. The tool always wins in the end, and the only open question is what you decide to value once it has.

Even if AI writing isn’t bad, is it deceitful? It depends.

Before we even get to disclosure, there is a prior question we keep stepping over, which is what the deceit actually means here. When someone says using AI to write feels misleading, they are pointing at a discomfort that has two completely different things inside it.

The first is the fear that the writer is coming off smarter than they are, and that the words make them sound more researched, articulate, and composed than the unassisted version of them would. I understand the discomfort, but I would gently point out that we have never once treated this as fraud in any other case. The 1996 student who actually went to the library came off cleverer than the one who made things up, and we called that being better at the assignment. Google made an entire generation sound researched on subjects they had encountered 5 minutes earlier, and we did not demand a disclosure line reading the following confidence was sourced from the second page of search results. Spell-check made everyone a competent speller. None of this was deceit, because the standard was never “did you do this the hard way.” The standard was whether the result held up. Coming off cleverer than your unaided self is not a crime, as it’s the entire point of every tool anyone has ever picked up.

The second form of deceit is deeper: the fear that you are passing off a robot’s thoughts as your own rather than merely its polish. The ideas themselves were also contrived by a robot and you are wearing them like a borrowed coat.

I think that fear rests on a confusion about what language actually is. At its core, a human thought is an organic event - a particular firing of synapses across brain tissue that belongs to exactly one person and exists in no words at all. Language is not the thought but the synthetic construct we invented to get the thought out of the skull. It’s a lossy compression format we have been refining for a few hundred thousand years, which is why “I can’t put it into words” is a sentence everyone has said and meant. So when a model predicts the next word in a sentence given your prompt, it’s just performing the encoding step, which you were always going to do imperfectly. The thought can be entirely yours even when the articulation is borrowed. Building upon the articulation of a genuinely human thought is a different act, morally and practically, from having nothing to say and letting the machine invent something to fill the silence.

Deceit is not using the tool to say your thing well, but having nothing to say and using the tool to disguise the absence of thought. One of those produces writing worth reading, while the other produces the slop. The good news, though, is that the slop announces itself.

The slop cutline

The reason I am not especially worried about the slop tsunami is that we are already extremely good at detecting it, and we got good at it the same way we get good at everything: by being annoyed enough times.

You have felt it. The opening. The short sentences in quick succession. The unapologetic em dash usage––. The machines that detect AI are already identifying this lingual texture extremely well. Pangram, one of the more accurate detectors, works by analyzing sentence construction, syntactic repetition, and the pacing across paragraphs. The detector can tell you the syntax and cadence resembles a machine, but it can’t tell you a human mind was not the source of every idea in it.

Leaning on those detectors as a public verdict is philosophically confusing, in my opinion. It’s an unjust in a way that should end the conversation about reader-facing labels on its own. A Stanford-led study found that seven widely used detectors flagged original human writing by non-native English speakers as AI-generated 61 percent of the time. Vanderbilt University turned off Turnitin’s AI detector in 2023 rather than keep falsely accusing its own students. Models like Pangram have improved substantially since, at accuracy levels of up to 98%. But even in a 1-in-50 false-positive scenario, an incorrect label shown to a reader who will read it as a mark of guilt does more damage than what it claims to protect against.

I believe we do not need the label, because we already have the filter, which is ourselves. Every consumer of writing carries a personal threshold below which the content reads as obviously machine-made, annoying, and not worth the time. Above that threshold, it reads as something a mind made for another mind. Those cutlines differ from person to person, but they will probably drift and converge over time toward a rough shared sense of what bot-written sludge feels like, the way we all eventually developed a shared nose for spam email and SEO content farms. Below the line, the obvious slop should go, and the market is already deleting it without any help from the state. But above that cut line, the answer to what “good quality” is, isn’t so cut and dry...

Aiming at the wrong person

A recent survey from Envato found that 58% of creative professionals had used AI in client work without disclosing it. Nearly half said they felt no obligation to list every tool in the process, which is to say the disclosure regime everyone is demanding already does not exist in practice and the sky has not fallen. Even the platforms that do mandate disclosure concede the point I have been making, because Amazon’s publishing arm draws a bright line between “AI-generated” and “AI-assisted” content and only asks you to disclose the former. The degree of AI’s involvement matters, and the institutions writing the rules already know that it does.

I believe a useful AI detection signal should not be a badge stapled to the work where the reader stands. It should be a mirror handed privately to the writer. Imagine the platform telling you before you publish, “this reads as machine-made,” the same way a coach tells you your swing has gone mechanical. The reason it has to point at the writer rather than the reader is the exact bias problem from two sections ago. A detection signal that is wrong part of the time is catastrophic as a public accusation and at worst a slight nuisance as a private note, because the writer is the one person who actually knows whether the flag is true and can take it or leave it accordingly.

Coincidentally, we’ve already run a version of this experiment with a different kind of label before. When Twitter began attaching warnings to disputed posts, researchers found a perverse side effect they called the implied truth effect: once some posts carried a warning, readers started treating every post without one as vetted, so the mere absence of a label got read as a clean bill of health. A flagging system that catches only a fraction of what it hunts quietly launders the far larger number it misses, because it has trained the reader to assume that unmarked means verified.

The lobotomy with a monthly fee

I don’t think this piece is fair without addressing the elephant in the room: Stupidification.

On his most recent show ‘Real Time’, Bill Maher called AI “a lobotomy with a monthly fee,” and the line lands because it names a fear the disclosure debate is too small to contain. The worry wasn’t only that you might deceive your reader, but that you would deceive yourself and offload the writing and the thinking to an LLM. Perhaps a generation handed a machine which produces competent, hallucination-prone prose on demand would simply stop developing the muscle that produces thought at all. This is the same fear as “calculators will ruin arithmetic” and “nobody will be able to write after Wikipedia.” The discomfiting part is that those fears were not entirely wrong. Plenty of people did get worse at long division. The real question is whether what we gained was worth what we atrophied. And that’s a question only you can answer.

The trust that matters here is in yourself to use AI without letting it use you. That’s a responsibility far heavier than a disclosure checkbox. It can’t be regulated, detected or stamped onto a post since it lives in the gap between the writer who reaches for the tool to say their thing better and the writer who reaches for it to avoid having to think of anything to say. The lazy AI user publishing the boilerplate slop is not going to get the audience they are hoping for, because the readers can see through it the same way they developed a nose for spam and a flinch for Buzzfeed listicle clickbait.

So in the long haul, I’m not that worried about AI ruining good writing and the need to shame those who use it publicly. It’s actually a way of avoiding the harder, introspective question, which is judgment about whether a given piece of writing was made by a human mind. I do all my own thinking, but I don’t always do all my own typing without some LLM refinement. Despite that, these posts still take several days to write and numerous rounds of revisions. I’m not convinced those are the same confession.

As always, I love to end every piece with a question for you:

Think of the last thing you read that genuinely moved you, that left you better than it found you: would it have mattered, at all, to learn a machine had helped arrange the words?