Everyone’s worried about AI slop. The term went from niche internet complaint to Merriam-Webster’s 2025 Word of the Year in under two years. Online mentions grew 9x in 2025 compared to 2024, according to Meltwater’s social listening data. YouTube’s CEO named it a top priority. A New York Times investigation in March 2026 found roughly 40% of videos recommended to children appeared to be AI-generated filler.
The alarm is real. But the diagnosis is wrong. Slop didn’t start when someone opened ChatGPT. It started decades ago, in boardrooms and pitch decks and entertainment studios that had already optimized conviction out of the equation. AI didn’t create the problem. It gave the problem a megaphone.
What You’ll Learn
- Why AI slop is a symptom of a structural problem, not the cause of a new one
- How cultural stagnation and brand homogenization predate generative AI by decades
- What research says about how AI models default to the average
- Why conviction — not better AI filters — is the actual defense against generic output
- What Mark Fisher’s cultural theory reveals about the machinery behind sameness
What Is AI Slop, and Where Did It Come From?
AI slop is model-generated content that looks competent on the surface but has nothing underneath: factually thin, conceptually generic, or responsive to a prompt without responding to an idea. The term was coined by technologist Simon Willison in May 2024, modeled on how “spam” named unwanted email. His framing was specific: slop isn’t defined by its origin but by its practice — content “mindlessly generated and thrust upon someone who didn’t ask for it.”
The Johns Hopkins MINT Lab formalized this in a 2026 literature review, identifying three properties that make slop recognizable: superficial competence, asymmetry of effort, and mass producibility. The content appears fluent and well-formatted. It was cheap to make. And it exists inside a system designed to produce more of it.
But here’s the part the conversation keeps missing. That description (competent surface, empty center, produced at scale for engagement) isn’t a description of an AI failure. It’s a description of a market condition that existed long before anyone trained a language model.
Didn’t Sameness Already Exist Before AI?
Yes. The phenomenon of averaged-down, conviction-free output is decades old, and it shows up everywhere — in design, entertainment, food, retail, and brand strategy. AI didn’t introduce sameness. It industrialized it.
Start with design. The branding industry calls it “blanding” — the cross-market convergence toward identical minimalist aesthetics. Sans serif fonts, muted palettes, stripped-away flourishes. Luxury fashion houses rebrand with near-identical wordmarks. Cafes from Kyoto to Reykjavik feature the same exposed brick, Edison bulbs, and latte art. As Flux Branding documented in 2021, blanding removes associations and attempts neutrality. The visual language signals modernity and safety. It also signals nothing.
The logic is economic. Brands align with peers because algorithms group similar aesthetics together, boosting visibility. The safer the look, the wider the audience. But the cost is erasure. When every brand looks like every other brand, no brand means anything in particular.
Entertainment follows the same pattern. Sequels, reboots, and franchise extensions dominate because they carry less financial risk. The IP is pre-tested. The audience is pre-built. The New Republic’s Aaron Timms described it as a culture trapped in a loop — fashion resurrects old trends, musicians appropriate predecessor aesthetics, Hollywood produces live-action remakes of existing properties. The result is a present that looks like a curated version of the past.
Cultural theorist Mark Fisher named this condition before most of us could see it. In Ghosts of My Life, Fisher described a “slow cancellation of the future” — the collapse of the expectation that things might be genuinely different. Technology advances, capital flows, but culture reproduces what already exists instead of generating what doesn’t. The 21st century, Fisher argued, was trapped in retrospection, recycling past styles rather than inventing new ones.
This wasn’t a failure of imagination. It was a failure of structure. Post-Fordist economics destroyed the spaces where artists could take risks. Monopolized distribution channels narrowed what audiences encountered. The machinery of production, from venture capital to platform algorithms, rewarded what was proven over what was new. Fisher’s framework locates the problem not in any individual creator’s lack of ambition, but in a system organized to prevent the new from emerging.
How Does AI Amplify Existing Sameness?
AI models don’t produce slop because they’re broken. They produce slop because they’re working exactly as designed — and the design defaults to the center.
Researcher Andrew Peterson documented this in a 2025 study on knowledge collapse. Large language models generate output toward the center of their training distribution, systematically favoring high-frequency patterns over rare, specialized knowledge. This isn’t a bug. It’s the mathematical consequence of how these models optimize. The long tail, where the unusual and distinctive content lives, gets compressed. The average gets amplified.
Peterson’s simulations showed that even a modest 20% cost discount for AI-generated content makes societal beliefs 2.3 times further from the truth. The models don’t lie. They flatten. They take the full range of human expression and return the most probable version, which is also the most generic.
A 2025 study on digital marketing confirmed this at the behavioral level. Researchers tracked restaurants using ChatGPT for Instagram content and found measurable homogenization — lexical similarity increased by 15%, syntactic similarity by 12%. When the restaurants lost access to ChatGPT, their content immediately became more diverse. The tool produced sameness. Removing the tool restored variation.
A meta-analysis aggregating 28 studies with over 8,000 participants reached the same conclusion from a different angle: AI augmentation improves individual performance on creative tasks but substantially reduces idea diversity across the group. People using AI produce more. They produce better. And they produce more like each other.
This is the mechanism that makes slop structural rather than incidental. It’s not that people are using AI carelessly. It’s that the models, by design, converge on the median. When the input has conviction, the output has something to work with. When the input is a prompt shaped by the same averaged signals the model was trained on, the output is the median of the median. That’s slop.
What’s the Real Root Cause of Slop?
Slop doesn’t begin when someone opens a generative AI tool. It begins when there’s nothing specific to say.
The question business owners keep asking (how do we avoid AI slop?) skips the diagnostic entirely. The problem isn’t the tool. The problem is what’s feeding the tool. When a brand has no clear answer to why it exists, what it believes, and how it shows up, every piece of content it produces defaults to the median. AI accelerates that default. It doesn’t create it.
The mechanism is conviction vacuum. A brand without a defined point of view has nothing to anchor its output. It doesn’t choose words; it accepts the most probable ones. It doesn’t frame ideas — it imports whatever framing the tool provides. The content that emerges isn’t bad because the technology failed. It’s empty because the brand was empty first. The AI surfaces that vacancy faster and at more volume than any prior tool.
Fisher’s unfinished final work, Postcapitalist Desire, arrives at this point from a different direction. The lectures explore why people struggle to want something genuinely different — why desire itself has been captured by existing structures. The way out, Fisher argued, isn’t to resist the present. It’s to be willing to want a different future. That willingness requires something capitalism’s incentive structures don’t produce: the conviction that something other than the current arrangement is worth building toward.
The parallel to brand work is direct. The companies that aren’t worried about slop aren’t the ones avoiding AI. They’re the ones that solved the underlying problem first. They know what they believe. They know why it matters. Their content has a spine because the organization has a spine. The tool, whether it’s AI or a blank page, reflects what’s already there.
How Do Brands Defend Against Slop?
The defense against slop is not an AI filter, a content policy, or a style guide update. It’s having something irreducible to say — and building the systems to say it consistently.
That means answering three questions before any content gets produced. Why does this brand exist? Why does it matter to the specific people it serves? And how does it show up, not just what it says, but how its signals hold together across every encounter?
Those answers function as a filter. When the brand’s point of view is clear, the content that flows from it has shape. When the point of view is absent, no amount of editorial oversight will prevent the output from regressing to the mean.
The market is sorting itself along this line. Meltwater’s 2025 analysis found that engagement with AI-generated articles dropped 40%, while human-generated content earned 5.44 times more traffic. Negative sentiment toward AI content reached 54% in October 2025. Brands like Aerie built entire campaigns around declaring they won’t use AI-generated images — and those campaigns outperformed everything else the brand published that month.
But the signal here isn’t that human-made content is inherently better than AI-made content. It’s that audiences can detect when content has nothing behind it. The emptiness is the problem — not the tool that produced it. A brand with conviction can use AI and produce work that sounds like itself. A brand without conviction will produce slop whether it uses AI or not.
As a general rule, slop is what happens when production outpaces conviction. The most reliable defense is not to slow production, but to build the conviction layer first — the brand’s worldview, its methodology, its point of view on what matters and why. Everything downstream of that layer carries its signal. Everything produced without it defaults to noise.
Conclusion
Slop is not an AI problem. It’s a conviction problem.
The question worth asking isn’t how to make AI produce better content. It’s why so much of the market had nothing specific to say in the first place. AI didn’t create the vacancy. It revealed it — at scale, at speed, in a form visible enough that we finally had to name it.
The brands that will survive an AI-mediated market aren’t the ones building better filters. They’re the ones that had something irreducible to say before the averaging began. The answer was always the same. It just costs more to ignore now.
