Against slop(py judgment)
Pangram can't tell you what's worth reading.
Disclosure: ChatGPT was unabashedly used to spar and synthesize.
—
Everyone is riled up about “slop.” But why, exactly? What even is it? And frankly, why do we even care?
TL;DR: “Slop” is a hackneyed label that collapses three distinct concerns — provenance, care, and quality — into one crude proxy: was AI used? Yet revealed preferences complicate the equation: when provenance is hidden or assumed, readers actually like and share AI writing. This suggests anti-slopers don’t always object to the output, but what authorship implies about effort and honesty. Care and quality are not inextricable, either: a person can make something bad on their own, and an AI-assisted work can be good or useful. Detection tools may approximate provenance (for now), but they can’t tell us whether something is good, resonant, and worth reading. And as human-AI creation evolves, that narrow function will become less useful. The durable response is not better detection, but better judgment.
Does anyone know what “slop” even is?
Ask ten people to define “slop” and you’ll get ten different answers. Ask why they dislike it — whatever it is — and you’ll find there’s no consensus. “Slop” is to creativity what fascism is to political ideology: amorphous, abused and trite.
I’ve found most people’s distaste for slop starts with a visceral “idk why, but I don’t like it.” They can’t quite put their finger on it. The reaction may be real, but you still need to articulate the accusation.
Is slop anything we have a hunch was generated by AI? Is it anything Pangram flags? Is it a harder-to-define quality entirely separate from the means of creation? Can a 100% human-created artifact be slop?
Surprisingly, I’ve seen many people who genuinely believe anything created with AI is de facto slop. That’s more or less also how Webster defines it. And if you’re one of those people, I’m unlikely to change your mind.
For the purposes of this piece, I’m going to focus on and attempt to steelman the most common, deeper objections I’ve heard or inferred:
First, people feel deceived. A human put their name on a piece, but they didn’t write it entirely using their own mind and hands — and they weren’t transparent about it.
Second, people feel AI generated writing is soulless. It feels hollow, generic, and disconnected. It’s empty mental and spiritual calories.
And lastly, AI writing lacks care. The writer requests more sustained attention from the reader than they were willing to put into the work. It lacks effort and thoughtfulness.
Said another way, I think people are primarily concerned about how a work was made (authorship, or provenance), the effort and attention that went into it (care), and the resulting artifact (quality).
These are all fair concerns. But they’re also three distinct attributes, and an AI detection tool can only (probabilistically) speak to one of them. When Pangram flags AI writing and we scream “slop!”, we collapse three considerations into one, replacing our own judgement with a poor proxy.
Revealed preferences
People will fervently say they hate slop, but when provenance is obfuscated — when they don’t know who wrote the piece and how — their opinion often changes.
Two examples make the point:
Repeatedly, some of the best performing Substack essays use AI. These are essays that have garnered hundreds of thousand of views and shares.
Earlier this year, The New York Times ran a quiz where they put (blinded) human and AI writing samples side-by-side and asked humans to pick their favorite. They found that 3/5 respondents preferred AI writing.
In short, absent the creator, people aren’t always averse to AI writing. In fact, they may even prefer and recommend it!
Of course, this doesn’t mean that AI writing, or these specific pieces, are automatically of the highest quality. Performance and quality shouldn’t be conflated.
But the widespread preference does still tell us something interesting: namely, that whatever “offends” us about slop or AI writing is not always present or detectable in the writing itself.
If a work was unchanged, but a change in your perception alone changed your opinion, what should we infer?
This isn’t to say that the truth of authorship is unimportant. But it might mean the objection was not purely about care or quality.
Care and quality are not inextricable
Care complicates the equation further.
Substack co-founder Chris Best recently defined slop as “something that nobody believes in.” This gets closer than most definitions, pointing towards conviction, attention, and care. But it also suggests that “belief” is the defining ingredient in good work, which I’m not sure is true.
A restaurateur can sink tens of thousands of dollars and hours into a sandwich shop and still make a shitty BLT. A novice writer can care deeply about their novel and every reader may still put the book down after the first sentence.
Quality may be more likely to follow from care, but it doesn’t necessarily follow.
This also ignores the reader’s “belief” in the work (see: revealed preferences above).
The inverse is also possible: something produced with little effort — or with substantial AI assistance — could still be high quality and resonant. I imagine humans made careless, derivative, and low quality works for millennia before AI entered the picture.
In short, the creator’s process and relationship to the work matter, but they don’t determine the quality of the output or the audience’s relationship to it.
The fear behind the detector
Back to the main story line, and to say it directly: my best guess is that, at least for some people, the intense, knee-jerk “slop” reaction is a fear of uncertainty.
They might be unsettled by the idea of (potentially) liking AI writing, or that they can’t tell the difference between human and AI writing, or that AI might write better than they or most people can (soon). Maybe they’re worried by the possibility of finding value in AI in their own writing, or what that usefulness says about the skills and identity they’ve spent years building.
On some level, I get it. The technology can feel alien in the way it abstracts so much of the process and decision-making — the way it blurs the boundary between creator and tool.
And if any of this is true, it raises hard questions about the nature of creation, our involvement in it, and those scary, existential questions about purpose and meaning.
What is a human writer if a machine can generate respectable prose in seconds? How much do I need to be involved in the process to call it “my” work — does a mere prompt count? And maybe most frightening, what does it mean if AI writing can move a person?
Rather than engage with those questions, it’s tempting to reach for an easy, socially acceptable eject button: slap the “slop” label on anything written with AI and keep moving.
Pangram might prove a stopgap for some. At best, it can offer some high confidence1 estimate of provenance.
Should you use it? If you want. But understand that it’s just a data point — it doesn’t tell the full story.
Should writers disclose AI use? In many contexts, it seems a fair ask — and maybe even essential in claims of lived experience. But I also wouldn’t expect it to be done uniformly.
And in either case, I’m not sure we should demonize any and everyone who uses it.
Most importantly, don’t use an AI detection tool to tell you if a piece of (AI) writing is good or resonant or useful. In doing so, you are ironically outsourcing your thinking to the same technology and with the same kind of laziness you’re accusing the writer of.
Nor should you expect detection tools to answer the stickier questions raised above — about creation, art, and meaning. Pangram can say with some certainty about how a sequence of words were created, but it can’t tell you if a work is worth your attention.
You can only wrestle with those questions yourself.
Exercise (your own) judgement
In short, you must maintain and exercise your own judgement. Now, but especially in the future.
AI detection tools might offer some near term — albeit, IMO misguided — “defense.” But we’re still so early. AI is only going get better and more elusive. AI-human hybrid writing will evolve and improve; the boundaries will get less clear. The technology will become more accepted and ubiquitous2. Detection can’t be the endgame forever.
I’m not saying provenance is entirely irrelevant — it’s especially relevant in particular contexts, as mentioned before. There is a line between fiction and falsification, and generating and presenting an experience as your own.
But provenance is an input into your judgement — it shouldn’t replace it. It is not a reason to wholesale disregard anything created with AI.
Shreyas Doshi said it well and simply:
We should develop our own sense, heuristics, and filters for quality — for writing that’s good, resonate, and useful.
These questions and assessment require more than Pangram’s verdict. Building a muscle for discernment is the only durable approach.
And if enough of us exercise that judgement individually, there’s at least some hope that our collective attention will reward higher quality works and weed out lower quality ones — AI or not.
It seems obvious why AI writing is performant: it approximates the mean of public human writing, heavily weighted toward the internet and modern era, and therefore has broad legibility and appeal. Combined with the rate at which the content can be generated, there is a concern I have about a sort of “recursive abundance” of mediocrity. If AI and humans are continually training each other on the “worsening” mean of writing, there’s potential for a vicious, runaway feedback loop (i.e. the median human trains the median model; the median model produces more median writing; people consume, imitate, and republish it; and that material becomes part of the future information environment). In this case, we run a real risk of instantiating the dead internet and speed-running Idiocracy. But again, this speaks to a need for a high bar for quality and judgement.