The Enduring Power of Slop // BRXND Dispatch vol 124
On how slop satiates our inherent desire for cohesive narratives
ICYMI—we announced our initial speaker lineup for BRXND NYC 2026 last week. We’ll have marketing, creative and AI leaders from Amazon, Microsoft, Bark, Rocket Money and more taking the stage with additional speakers announced soon.
$749 early bird pricing runs through July 31. Can’t wait to see many of you at the Times Center on November 5.
The Enduring Power of Slop
The terminally online are having a viscerally anti-slop moment. Gradually, then suddenly, the tech elite have decided they’ve had enough of sacrificing the written word at the syntax abattoir of “it’s not delivery, it’s DiGiorno.” Substack in particular threw down the gauntlet this week, partnering with Pangram to launch an AI detection feature across the platform.
For most of us working in and around marketing and AI, this feels like the proverbial “nature is healing” moment. But to most of the world, we’re just a bunch of whiny wordcels. Based on sheer engagement metrics, the people love empty calories masquerading as a healthy meal.
Outside the filter bubble, the proliferation of slop continues at an astounding pace. >40% of longfrom LinkedIn posts are now AI generated and shameless copy paste jobs regularly go viral on the platform. On X, one of the most shared pieces around the World Cup was this AI-generated essay hyping up the US. Soccer team. Even as someone who has Pangram wired into my consciousness, I read it and a small part of me wanted to run through a brick wall. Matt Freese punished me accordingly.
Slop, of course, is not an entirely new phenomenon and it long predates LLMs. NCIS…..ah whatever the hell city they’re in now… was the top watched drama of the 2010s. We have endured ten Fast and the Furious movies, not even counting the bizarrely excellent sloptacular known as Hobbes & Shaw. Plenty of human written content, especially when forced to take the form factor of SEO arbitrage, was just a more simplistic slop varietal.
Slop endures because it is ultimately a demand side problem. LLMs don’t even create our inherent demand for slop. They simply satiate it. Said another way, slop has product-market fit, at least when it comes to influencing humans. Paradoxically, I agree wholeheartedly with Noah’s take from last year that slop will be far less effective at influencing agents long-term than it is at hypnotizing people.
But while a vocal minority seems to be reaching their slop breaking point, why does AI-generated text maintain such a stranglehold over its consumers?
Slop wins because it conforms to neat narrative arcs that fulfill our desires for cohesive stories: At crime scenes, children make the best witnesses because they simply report back what they saw rather than editorializing into a natural story arc. The adult brain is hardwired to eliminate cognitive dissonance by effectively hallucinating events that didn’t happen to make our memories fit into a tidy narrative.
Does that sound like any of the AI tools you know? In many ways, a tragic flaw of LLMs is that they are even more allergic to cognitive dissonance than people are, hence the hallucination problem being such a tough one to squash completely. Moreover, any time there is even mild tension or paradoxical sentiment in an idea, models go out of their way to call it out ( “that’s the real sticking point”) and try to crush it when the more intellectually honest tactic would be to simply let the tensions be.
Insufferable vernacular tics aside, this makes AI writing more “comfortable” to read, especially in the passive consumption mindset that, well let’s face it, most people are in most of the time.
In many ways, the most interesting thing to talk about when it comes to slop is why it seems to those of us who spend the most time with these models that each new model, despite its increases in intelligence, seems to be more slop-y, not less.
Here’s Noah’s observation from earlier this month:
One of the things I’ve found most confounding over the last year is that as these models get better, at the same time, I feel like they’re getting worse at writing. Or maybe “worse” isn’t quite the right word, but I’m not seeing any end in sight for the slop.
I suspect this comes down to a confluence of circumstances: the nature of the model is to average things out, reinforcement learning with human feedback (RLHF) moves the models towards being a “helpful assistant,” and the move to coding and reinforcement learning with verifiable rewards (RLVR) means the models are further steered towards a model of the world with conclusive truths.But my guess on the biggest challenge is that good writing is fundamentally anti-consensus, and I’m not sure that’s a problem the models can really solve. As soon as the models all agree on a turn of phrase or style, it becomes “slop” and can’t really be used by humans anymore. As opposed to code, where (for the most part) a good answer is a good answer (you can verify the code works), the quality of writing is subject to other writing that exists around it.
At this point, I’ve abandoned using AI to generate any copy at all for public-facing work. Perhaps there’s some dogmatic pearl clutching here and I’ll re-evaluate with each new model release but for now, the trend of each subsequent model doubling down on a slop-y vernacular identity (and increasing the conviction with which it does so) is accelerating.
With heavy training on my voice, I used to work with ChatGPT and Claude for writing transition paragraphs, flagging and eliminating weasel words, and pushing me more towards brevity. I always found LLMs made my prose too detached and omniscient but older models felt more malleable. I now feel mostly powerless to stop the models from injecting stylistic touches that simply do not sound like me.
Ultimately, I suspect that we’ll lose more of the internet to slop simply because in more and more instances, slop will get the job at hand done. But as with any good or service, there will always be a market for artisanal prose, a market that will grow as agents seek out higher quality content and increasingly esoteric expertise. Thus, the inherent value of writing that requires real learned experience and imbuing contrarian thought will also go up. On balance, this is a better era to be a writer than the pre-LLM times.
But for all of its captivating power, AI-generated text will always have one fundamental flaw. Even if it satisfies the reader, the writer is a cognitive passenger princess in their own life.
If you have any questions, please be in touch. As always, thanks for reading.
— Mike
Great food for thought (except for the Fast and Furious ding; I'm with Nolan on that one).
Framing it as a supply + demand problem is fascinating; there's so much content, and yet there continues to be a demand for more, or rather more empty calorie content.
This: “Even if it satisfies the reader, the writer is a cognitive passenger princess in their own life.”