Your Brain is Wired to Love AI Slop

By Courtney Turrin · August 17, 2026 · 13 min read read

Category: AI

Human perception favors sameness, explaining why AI content and even website designs become uniform. LinkedIn is now flagging generic AI "slop.

LinkedIn added a button in July so people could report AI slop. The research that explains why so much AI writing and design already looks and sounds the same goes back to 1990, and it says the sameness is a feature of human perception, not a failure of the machine.

Sometime before 2000, a male white-throated sparrow in central British Columbia changed his song.

For as long as anyone had recorded the species, males sang a whistled introduction followed by a repeated triplet, the pattern birders wrote out as "oh, my sweet Canada, Canada, Canada." Then one bird, somewhere in the interior of the province, dropped the third note. Triplet became doublet. Canada, Canada, Canada became Cana, Cana, Canada.

Biologist Ken Otter spent two decades tracking what happened next with the help of citizen scientists who submitted more than 1,700 recordings. By 2004, about half the white-throated sparrows in Alberta, the province just east of British Columbia, had picked up the new song. By 2014, the doublet was the only version anyone could find there. The variant kept moving east across the prairies, over the Rockies, through Ontario, into Quebec, eventually covering roughly 3,000 kilometers. Geolocator data on tagged birds pointed to a mechanism: males from different breeding populations were wintering in the same grounds, where younger birds learned songs from whoever was nearby. Otter and his co-authors called it the first documented case of song culture shifting at a continental scale. By 2019, the traditional triplet survived only at the country's eastern edge.

A shared wintering ground turned one bird's variant into the only song a continent could produce.

Why LinkedIn Killed Its Own AI Writing Tool

On July 30, 2026, LinkedIn's Chief Product Officer, Hari Srinivasan, announced a new option buried in the three-dot menu on every post: "Seems like AI slop." Click it and the post disappears from your feed, with a note that the flag will help train the platform's ranking system. LinkedIn had already been working the problem for months. Laura Lorenzetti, the company's vice president and executive editor, wrote in May that the platform was building classifiers trained to recognize content that "may sound polished on the surface but lacks any real unique perspective or substance," and that flagged posts would see reduced distribution rather than removal. LinkedIn's own account of its early testing put the classifier's accuracy at correctly identifying generic content 94% of the time.

The company simultaneously pulled its own AI writing assistant, the one that had offered to "enhance" a draft, and replaced it with a plain proofreading tool that edits grammar without changing voice.

This is not an anti-AI argument. I run a marketing agency built around using these tools well, and the tools are not the problem. What follows is about a specific failure mode: what happens when a system trained to find the response that tests best applies that same response to everyone, at scale, regardless of who is asking. That failure predates generative AI, it is measurable in web design going back to 2007, and the psychology explaining why the pattern that tests best is also the pattern that starts to look and sound the same everywhere is three decades old.

Templates Flattened the Web Before AI Did

Researcher Sam Goree ran computer vision analysis across roughly 10,000 corporate websites captured in the Internet Archive between 2003 and 2019. He and his co-authors measured three things: color palettes, page layouts, and a set of visual features learned by a neural network trained to tell websites apart. All three measures told the same story. Website designs were most different from each other between 2008 and 2010, when larger, higher-resolution monitors gave designers more room to work with. Then, between 2010 and 2016, the differences collapsed. Average layout distance between sites dropped by more than 30% (Fig 1).

[datawidget key="web-design-convergence"]

Templates did that. Content management systems, component libraries, and shared frameworks arrived before any large language model did, and they narrowed the range of what a website could look like well before AI touched a single line of code. Anyone who has built websites since 2017, as I have, can see the newer version of the same pattern now: a site built with an AI coding tool like Lovable or Cursor is often identifiable within seconds, not because the tool is incapable of variation, but because most people using it haven't invested the time to push past the default. That is a fair caveat. Many of these builds are side projects or early prototypes where design was never the point. The pattern holds anyway: absent deliberate customization, the output looks like the tool, not like the person who typed the prompt.

The philosopher Agnes Callard offered a mechanism for this in a July 2026 interview with journalist Derek Thompson: comparison itself tends to flatten differences once it becomes possible. Her example: two school districts each run a high school with its own curriculum. Nobody compares them, because a student's home address decides which school they attend. Change the rule so any student can choose either school, and the schools start competing on the same measures: graduation rates, AP course counts, college placement. One school adds AP classes it never needed before, purely to keep pace. Neither curriculum failed. Putting both schools into the same evaluative field made convergence the rational move for each of them. A template library and a shared LLM do the same thing to websites and prose. Once everyone's output can be measured against everyone else's by the same metric, on the same platform, the range of what survives narrows.

Why the Prototype Wins

In 1990, psychologists Judith Langlois and Lori Roggman digitized photographs of faces and mathematically averaged them into composites. Adults rated the composite faces as more attractive than nearly all the individual faces that had gone into them, and the effect strengthened as more faces were added to each average. The finding held for male and female composites across three separate samples. Researchers later replicated the pattern well beyond faces: composite images of dogs, wristwatches, and birds were also rated more attractive than the individual examples that produced them.

Psychologist Piotr Winkielman and his co-authors identified the mechanism in 2006. Working with random-dot patterns and simple geometric shapes rather than faces, they found that prototypical stimuli were both easier to categorize and rated as more attractive. When they statistically controlled for how easy each stimulus was to process, most of the link between prototypicality and attractiveness disappeared. Ease of processing was doing the work. The technical name is processing fluency: a pattern the brain has seen the shape of before takes less effort to recognize, and that reduced effort registers as pleasure.

A large language model's default output is close kin to Langlois and Roggman's composite face. It is the statistical center of its training distribution, generated precisely because that center is the most probable, most fluent, most quickly processed continuation of whatever came before it. The AI-built website is not badly designed. It is often the averaged face: immediately likable, instantly legible, and nearly impossible to tell apart from the one next to it.

The complication is that liking and remembering are not the same thing, and the literature on which one attractiveness serves is unsettled. Some studies have found people recall attractive faces better; others have found the opposite, or no difference at all, once a face's distinctiveness is accounted for separately from its attractiveness. Fluency may buy an instant, favorable first impression. It has no guaranteed claim on what a reader carries away an hour later.

The Model Learned What Worked, Then Gave It to Everyone

Large language models are trained in stages. After initial training on text, most commercial models go through reinforcement learning from human feedback, a process where human raters (or another model standing in for them) score candidate responses and the system is adjusted to produce more of what scored well. Researcher Robert Kirk and six co-authors published a rigorous stage-by-stage analysis of this process in 2024. Comparing models at each step, from supervised fine-tuning through the full RLHF pipeline, they found that RLHF substantially improves how well a model generalizes to new situations, but at a real cost: it significantly narrows output diversity compared to supervised fine-tuning alone, both for a single prompt asked many times and across different prompts. The industry term is mode collapse. The model converges on the response pattern that scored highest and increasingly produces variations on that one pattern rather than exploring the space of good answers.

Economists Anil Doshi and Oliver Hauser ran a writing experiment testing this at the level of output a reader would actually see. Writers who used generative AI ideas to draft short stories produced work independent raters scored as more creative, better written, and more enjoyable than stories written without AI help, especially among less practiced writers. But the AI-assisted stories were measurably more similar to each other than stories produced by unaided human writers. Every individual writer got better. The population of stories got narrower.

Put the two findings together and the LinkedIn caption pattern, the three-word emphasis line, the "not because X, but because Y" construction, stops looking arbitrary. Somewhere in training or fine-tuning, that structure tested as effective, in the narrow technical sense of scoring well against a reward signal built from human ratings. Mode collapse then applied that single winning structure to everyone who asked the model for a post, a caption, or an email, regardless of what they were actually trying to say. What reads today as an unmistakable AI tell was, at some earlier point, simply good copywriting. It stopped being a differentiator the moment it became the default for millions of prompts at once.

Design research diagnosed the ceiling on this strategy decades before language models existed. Psychologist Paul Hekkert and his co-authors, writing in the British Journal of Psychology in 2003, found that people's aesthetic preference for industrial products depends jointly on typicality and novelty, and that the two factors suppress each other: a design gets credit for feeling familiar only up to the point where that familiarity crowds out anything new, and gets credit for novelty only up to the point where it stops reading as a member of its category at all. Researchers now call this the MAYA principle, for "most advanced, yet acceptable." A system optimized purely to maximize typicality, which is functionally what a mode-collapsed model does, is optimizing for exactly half of what actually produces a preferred design.

Detecting AI Text Is a Skill, Not a Given

Your experience, and mine, is that a LinkedIn caption reads as obviously AI-written within a sentence or two, while a long-form article can slip past unnoticed. The research on human detection accuracy backs that split, and it locates the difference in the reader, not just the text.

Researchers Jenna Russell, Marzena Karpinska, and Mohit Iyyer tested how well people could tell AI-generated writing apart from human writing across 300 non-fiction articles, including text that had been deliberately reworded to evade detection. Most annotators performed close to chance. But a subset of five annotators who reported using large language models frequently for their own writing performed almost perfectly, with their majority vote misclassifying only one article out of 300, without any special training (Fig 2). The researchers found these frequent users were picking up on more than vocabulary. They noticed originality, tone, and formality, qualities that automated detectors struggle to score.

[datawidget key="ai-detection-experts"]

Detecting AI writing, in other words, is not a fixed human capability. It is a skill distributed unevenly, concentrated in people who read and write with these tools constantly, which describes a lot of editors, marketers, and heavy LinkedIn users, and thinly distributed everywhere else.

Detection ability and disclosure penalty are two different things, and disclosure carries its own paradox. Researchers Benjamin Toff and Felix Simon found that audiences rate news labeled as AI-generated as less trustworthy on average, even when the articles themselves are not judged any less accurate or fair than unlabeled ones. Telling people a text was AI-assisted damages trust in it independent of whether the text earned that distrust on the merits.

When Polish Stops Signaling Anything

Marketing has had a name for the value of visible effort since long before generative AI existed. Researchers Tim Ambler and E. Ann Hollier published a study in 2004 arguing that perceived advertising expense, the sense that a company spent real money making an ad, raises persuasiveness indirectly, by raising perceived brand quality. They drew the framework from the handicap principle in evolutionary biology, the idea that costly, hard-to-fake signals (a peacock's tail, in the textbook case) communicate genuine fitness precisely because a weaker competitor couldn't afford to produce them.

Generative tools push the visible cost of producing a polished-looking post, deck, or website toward zero. That breaks the signal Ambler and Hollier described. Polish stops indicating investment once investment is no longer required to produce it.

LinkedIn's own accounting for what its slop classifier is trained to catch makes the same point from the platform side: content flagged as generic "may sound polished on the surface" while lacking any actual perspective. The surface and the substance have come apart, and LinkedIn built a ranking system around detecting the gap.

The numbers on what that gap costs are not settled. Both firms measuring it sell AI detection tools. Pangram Labs sampled roughly one million posts scrolled by opted-in users of its browser extension and found 41% of LinkedIn posts over 250 words were fully AI-generated. Originality.ai, searching public posts across a fixed set of topics with a lower 100-word threshold, found 53.7% of a 2025 sample and 81.2% of a July 2026 sample were "likely AI." The two figures are not tracking the same trend line; they are measuring different samples with different thresholds and different sampling logic, one drawn from what real users actually scrolled past, the other from public search results, and both are closer to a lower bound than a census.

Sampled by industry, AI-flagged LinkedIn posts in the "leadership and inspiration" category outperformed human-written posts by 75%. In healthcare, human-written posts outperformed AI-flagged posts by 44%. In government and public affairs, by 40% (Fig 3). Machine writing appears to win exactly where the category was already commodity content before AI existed, generic motivational posts that were interchangeable five years ago, and to lose precisely where a reader needs to trust a specific, accountable person.

[datawidget key="linkedin-industry-engagement"]

What Comes After the Doublet

Ken Otter's research team is already tracking what comes after the doublet. A newer variant, a modulated doublet with a distinctive amplitude shift on its first note, has emerged in a population west of the Rocky Mountains and has begun spreading the same way the original doublet did twenty years earlier: rare at first, then rapidly adopted once it crosses paths with enough birds passing through the same wintering ground.

Convergence was never the end state. It was the interval between one variant reaching fixation and the next one finding the same shared channel to travel through. A wintering ground works the same way whether the travelers are sparrows or sentence structures: it rewards whatever gets there first and spreads fastest, not whatever would have served the reader, or the listener, best in the long run. The species that keeps producing new variants despite that pressure is the one still worth listening to.