AI Slop and Sturgeon’s Law

When the training data is 90 percent crap, what should you expect?

Photo by Ben Hadfield on Unsplash.

Theodore Sturgeon was a prominent science-fiction author who may be best known for 1953's More Than Human. If it were written today, it might be characterised as a transhumanist novel. He is also credited with originating what has come to be known as Sturgeon's Law. When asked to defend the quality of his genre against critics who claimed 90 percent of it was crud, Sturgeon famously retorted, "Ninety percent of everything is crap."

He wasn't being cynical but realistic. Any content consumer today will know that most of what is produced is mediocre, uninspired, or just plain wrong. This is pretty much true across the board—literature, film, music, academic research, and particularly the endless stream of content on the internet. Only a small fraction, at most 10 per cent, is genuinely excellent.

For decades, this was just a pithy or curmudgeonly comment on the everyday creation of what we now call content. In a wonderful bit of serendipity, Sturgeon's Law has become a way to counter AI hype and provide a cautionary note about AI tools and a fundamental problem with them.

The Training Data Dilemma: An Ocean of Mediocrity

Large Language Models (LLMs), the engines behind tools like ChatGPT, Gemini, or Claude, are trained on datasets scraped indiscriminately from the Internet. They then extrude text, mimicking writing, reasoning, and creativity by identifying patterns in the text and images created by humans. This includes everything from digitised books and scientific papers to news articles, Reddit comments, and billions of social media posts. Here's the catch: if Sturgeon's Law holds for the Internet (and anyone who has spent time online knows that 90 per cent is a generous estimate), then these AI models are being trained on a dataset where at least 90 percent of the material is, for lack of a better word, crap.

When you train a system on a diet of mediocrity, you shouldn't be surprised when its output reflects that. The model learns to replicate the most common patterns it sees. It's a regression to the mean, a levelling to the lowest common denominator. The result is what many are now calling "AI slop"—content that is superficially plausible, grammatically correct, but ultimately bland, soulless, and often subtly inaccurate. It mimics the style of human writing without the substance, the spark of original thought, or the depth of lived experience.

This isn't just a matter of quality; it's a matter of truth. The "crap" in the training data includes not just poorly written prose but also misinformation, conspiracy theories, and deeply ingrained societal biases. The AI learns these patterns just as readily as it learns grammar rules.

The Consequences for Authentic Creation

The proliferation of AI slop presents a clear danger to creators, researchers, and anyone who values authentic human expression.

• The Devaluation of Original Work: As our feeds become flooded with cheap, instantly generated content, original research and authentic creativity fade into the background. It becomes harder for readers and consumers to find the genuine article amidst a sea of synthetic text. Why would a company pay a Canadian writer or journalist for a thoughtful article when they can generate a passable, keyword-stuffed equivalent for pennies? The economic foundation of creative labour is eroding.

• Poisoning the Well for Future Knowledge: We are creating a self-referential loop of mediocrity. Future AI models will inevitably be trained on the slop produced by today's AI. The internet's dataset is becoming increasingly polluted with synthetic, derivative content. This will make future AI models less reliable and make it harder for humans to find trustworthy information and untainted sources. AI models are crapping into their input funnels. The results will not be good.

• Privacy, Data Colonialism and the Human Cost: Let's not forget where this training data comes from. It's our collective intellectual and creative output—our blog posts, family photos, and late-night forum arguments—scraped and ingested without our meaningful consent to fuel the development of commercial products. This is a new form of data colonialism, where the raw material of our lives is extracted, processed, and turned into a product that benefits a handful of corporations, primarily in the Global North.

  • But the exploitation runs deeper. The "dirty work" of making these AI systems functional—the painstaking and often traumatising labour of data classification and content moderation—is frequently outsourced to workers in the Global South. These are the people paid pennies to view and label the very worst of the internet's content: hate speech, violence, and abuse. This is the hidden, human cost of our shiny new AI toys, a stark reminder of the global inequalities that underpin the digital economy.

• The Climate Connection: The computational power required to train these massive models is immense, contributing to a significant carbon footprint. Are we willing to expend vast amounts of energy and resources, contributing to our climate crisis, to generate a tidal wave of digital mediocrity? Is this a worthy trade-off?

Charting a Different Course🍁

The federal government has just appointed its first-ever Minister of Artificial Intelligence and Digital Innovation, Evan Solomon. This presents a pivotal opportunity to reject the Silicon Valley model of reckless development and instead forge a uniquely Canadian path. This is the moment to champion Public AI—AI systems built in the public interest. We can create open-source models trained not on the internet's sludge, but on high-quality, curated datasets from our world-class cultural and scientific institutions. Imagine AI tools designed to strengthen public services, not just to maximise corporate profit.

The new ministry must establish strong guardrails for sustainable AI. That means implementing policies that demand energy efficiency and transparency from data centres, protect workers from exploitative labour practices common in the industry, and ensure that the benefits of AI are shared by all Canadians, not just a select few. The new minister has signalled a focus on economic benefits over regulation, which makes it even more critical for activists and citizens to demand a framework that prioritises people and the planet.

We cannot allow the future of our information ecosystem to be dictated by a model trained on the worst of us. We need a public conversation about the digital world we want to build, led by citizens and not just tech corporations.

This means demanding transparency in how AI models are trained, fighting for policies that protect the rights and livelihoods of human creators, and investing in and creating spaces for authentic, high-quality information and art to flourish, free from the noise of the slop machine.

We cannot allow the future of our information ecosystem to be dictated by a model trained on the worst of us. We need a public conversation about the digital world we want to build, led by citizens and not just tech corporations. Sturgeon's Law was a witty observation. It was never meant to be a technical blueprint for the future of knowledge. It's time we focused on curating and celebrating the 10 per cent of excellence rather than building a global infrastructure that endlessly regurgitates the other 90 per cent.