We Are Not Doomed to AI Slop

In 2025, “slop” became Merriam-Webster’s Word of the Year. Not an acronym from the AI industry. Not a term coined by researchers. A word that ordinary people reached for because they recognised what was happening to the internet: a flood of AI-generated text, images, and video that looks competent but says very little.

The recognition was deserved. Nearly ten of YouTube’s hundred fastest-growing channels last year were AI-generated. Facebook feeds filled with shrimp Jesus and AI cat soap operas. Spotify removed 75 million spam tracks in a single year as generative tools made it trivial to produce fake music at scale. YouTube CEO Neal Mohan opened 2026 by listing slop reduction as a platform priority. The word caught on because the phenomenon is impossible to ignore.

But here is where most of the conversation goes wrong. The fear is that AI is inherently incapable of producing quality work, and that the internet will inevitably drown in synthetic noise. That framing misidentifies the problem. The issue is not that models cannot produce quality. The issue is that most people are still using them without any process at all.

What we call AI slop is, in most cases, the output of a missing workflow.

The Vending Machine Problem

When generative AI is treated like a vending machine, where a single prompt is expected to yield a finished artefact, the result is almost always mediocre. Hand a large language model a blog title and ask it to write the article. The result will be polished but bland. It reads like a summary of things you have already read, because statistically that is exactly what it is. The model plays it safe. It averages. It avoids sharp edges.

This is not a failure of intelligence. It is a failure of method. When context is small and interaction is single-turn, the model has no reason to take a strong position, no pressure to defend a claim, and no working memory to build judgement across a complex topic. In that setting, slop is the natural outcome.

Researchers have quantified what happens when this low-effort pattern scales. Ilia Shumailov and colleagues at Oxford, Cambridge, and Toronto published a paper in Nature in 2024 showing that models trained recursively on their own output undergo “model collapse,” progressively losing the tail of the original data distribution until their outputs bear little resemblance to the real thing. Train on slop, get more slop, in a tightening spiral.

But the research also points to the exit. Work by Julia Kempe at NYU and collaborators demonstrated that when synthetic data accumulates alongside real human data rather than replacing it, collapse is avoided. The problem is not synthetic content per se. The problem is synthetic content produced without curation, review, or process.

One Person, One Process, Production Software

The same pattern shows up in software, and the numbers are stark. The 2025 Stack Overflow Developer Survey, with over 49,000 respondents across 177 countries, found that 66% of developers say their biggest frustration with AI tools is code that is “almost right but not quite.” 45% report that debugging AI-generated code takes longer than expected. A VP of Engineering at Google was quoted saying “People would be shocked if they knew how little code from LLMs actually makes it to production.”

That is slop in code form. And it comes from exactly the same cause: treating the model as a replacement for engineering discipline rather than a tool that operates within it.

The teams succeeding with AI-assisted development are not skipping process. They are doubling down on it. They plan with the model, not just code with it. They produce specifications, phased implementation plans, and test strategies before a line of code is generated. Code arrives in controlled increments, reviewed and validated before the next step proceeds. Tools like Claude Code enforce this naturally: the agent opens a pull request, tests run automatically, a human reviews the diff before merge. That is continuous integration applied to AI-assisted development. The pattern is already production-real.

I can speak to this directly. inmydata Studio, our chat-based dashboard designer, includes authentication, billing, and production deployment. It is a product we had aspired to build for over a decade. Previous attempts required teams and timelines measured in years. With AI-assisted development and a disciplined process, I built and shipped it alone in two months.

That was not a fluke. We are currently building an AI expert interview platform that captures tacit knowledge through voice-based conversations, extracting the “when,” “why,” and “what to do when things go wrong” that documentation never contains. A week of planning. Two days to build.

Neither of those outcomes came from a single prompt. They came from structured workflows: planning with the model, decomposing problems, reviewing output at every step, and forcing the results through the same engineering standards we would apply to any production system. The model did not replace the process. The process is what made the model useful.

The Difference Between Prose and Meaning

The same lesson applies to writing, though it is harder to measure. A model can generate prose. Only a process can generate meaning.

When you work with a model to explore an idea, test different angles, surface counter-arguments, bring in sources, shape a structure, and only then begin drafting, you get something qualitatively different from a one-shot generation. The model becomes a fast, tireless collaborator that helps you explore the space of ideas far more quickly than you could alone. The human remains responsible for judgement, taste, and intent.

This piece is a case in point. The argument was developed iteratively: researching evidence, testing the structure, challenging weak claims, and rewriting until the voice was mine and every assertion was backed by something specific. The model helped at every stage. But the decisions about what to say, what position to take, and what to cut were mine.

That workflow is not unusual among people producing serious work with AI. It is, however, invisible to the consumer, who sees only the output. The gap between a considered, process-driven piece and a one-shot generation is enormous, but the two can look superficially similar. This is why slop spreads: the cheap version is almost free to produce and, at first glance, almost indistinguishable from the real thing.

Almost. Not for long.

Platforms Will Not Save Us. They Never Have

There is a comforting version of this argument that says platforms will develop strong incentives to distinguish between raw generative output and work that has been reviewed and improved by humans. Trust, reputation, and economic value all depend on that distinction. The market will self-correct.

This is too generous. The dominant platforms optimise for engagement, and engagement rewards volume, novelty, and emotional provocation, not quality. This is not a new problem. It is the same incentive structure that gave us clickbait, content farms, and SEO spam a decade ago. We did not solve those problems cleanly. We are still living with the large-scale societal consequences of platforms that reward attention over substance: misinformation, political polarisation, mental health impacts, and an information environment that most people navigate with justified suspicion.

So no, platforms will not save us from slop. They never saved us from spam. They adjusted when the damage became commercially inconvenient, and not before.

But this is not the only force in play. Something genuinely different is happening in the tools themselves.

Anthropic, the company behind Claude, has built its entire business on the premise that safety and quality constraints make the product better, not worse. This is not marketing. It is a structural choice that runs through their model training, their product design, and their approach to deployment. Claude is the most admired LLM in the 2025 Stack Overflow survey. Claude Code is becoming the tool serious developers reach for precisely because it enforces the kind of discipline that produces reliable output. The ethical stance is not a tax on capability. It is what produces capability people actually trust.

That matters because it proves the model: you can build AI tools that embed process, that enforce review, that make good practice the default rather than the exception. The question is whether the rest of the industry follows that lead or continues to optimise for raw output speed.

A Growing Pain, Not a Terminal Condition

Slop is real. The damage it is doing to the information environment is real. But it is a growing pain, not a permanent condition, for a simple reason: nobody who cares about their work will tolerate being indistinguishable from everyone else.

The people flooding Facebook with AI-generated content are not trying to produce quality. They are arbitraging an attention economy that briefly rewards volume. That arbitrage closes as audiences learn to recognise and ignore synthetic noise, as platforms face regulatory and reputational pressure, and as the economic value of undifferentiated content trends toward zero.

Meanwhile, the people and organisations who care about what they produce now have access to tools that make quality dramatically cheaper to achieve. A structured AI workflow does not just match the old standard. It exceeds it, because it allows individuals and small teams to apply levels of rigour, research, and iteration that were previously only available to large organisations with dedicated resources.

The floor is rising. Not because the models are getting smarter, though they are. Because the processes and tools that surround them are maturing fast enough to make discipline the default rather than the exception.

The Word and the Work

“Slop” entered the dictionary in 2025 because we needed a word for what happens when you skip the work. AI-generated content produced without planning, review, or intent. Code generated without specifications, testing, or architectural thought. Output that looks finished but is not, because no one applied the judgement required to make it real.

The answer was never better models. It was always better process. The model collapse research confirms it. The developer surveys confirm it. The tools that are winning in the market confirm it. And the experience of anyone who has shipped production work with AI confirms it: the quality of the output is a direct function of the quality of the workflow that produced it.

That process is already here. It is embedded in the best tools, proven in production, and available to anyone willing to use it. The only question is how quickly the rest of the industry catches up.