AI systems are increasingly training on their own outputs, and researchers say the consequences are arriving faster than the industry anticipated
There is a moment in certain kinds of collapse where the warning signs were always there, visible to anyone paying close attention. We may be living through that moment with artificial intelligence right now.
Researchers at Oxford and Cambridge have published findings in Nature that are drawing serious attention from technologists, media strategists, and policymakers alike. The phenomenon they describe is called Model Collapse, and it is already underway, in real time, on the internet in use today.
The mechanism is direct: when AI systems train on content generated by previous AI systems, something irreplaceable begins to disappear. Rare ideas erode first. Creative outliers fall away. The full texture of human expression, its contradictions, its regional cadences, its resistance to easy categorisation, gets smoothed into something statistically comfortable and intellectually thin. What remains is output that is technically coherent but lacking in depth.
The internet, once the most extensive archive of human thought ever assembled, is being flooded with AI-generated content at a scale no curator can manage. That content will become tomorrow’s training data. The feedback loop has already begun.
The Fire Metaphor
Ross Anderson, co-author of the Cambridge research, offered an analogy that deserves wider circulation than academic citations usually allow. Large language models, he said, are like fire: genuinely useful, genuinely capable of causing harm, and capable of polluting everything they touch if left unmanaged.
The pollution Anderson describes shows up in the data. There is a gradual homogenisation of AI outputs, a narrowing of what these systems consider plausible or interesting, and a quiet erosion of the statistical long tail where genuine originality tends to live. When a model trained on human writing produces text, and that text re-enters the corpus, and the next model trains on the result, each generation loses something of what made the original source worth learning from.
No better algorithm resolves this. The problem is structural, a consequence of a particular kind of scale.
The Case for Human Expression
There is a version of this conversation that gets trapped in sentiment, a defensive position around human creativity that sounds more like mourning than reasoned argument. The evidence, however, points in a different direction.
The case for human expression rests on something more concrete than warmth or authenticity. Human writing is irreplaceable as source material. At its most distinctive, it contains precisely the outliers, the idiosyncrasies, the minority perspectives and difficult observations that give any training corpus its range. Remove those, and what remains is a model that has lost sight of the edges of what it knows. Edges, in any field of knowledge, are where most of the interesting work happens.
Writers, journalists, thinkers, and communicators who bring genuine specificity to their work occupy a position that AI systems cannot replicate and, as training data grows more synthetic, cannot even meaningfully approximate.
Scale Over Substance
It would be unfair to suggest that no one in the industry saw this coming. But there is a meaningful difference between acknowledging a risk in a footnote and building strategy around it. The leadership of the most prominent AI companies has, on the whole, moved carefully around that distinction.
Sam Altman’s public statements over the years have reflected considerable optimism: large investments, ambitious timelines, significant breakthroughs anticipated just ahead. The commercial logic was coherent, and in many respects compelling. The gap between the capabilities being promised and the limitations being quietly accumulated has, however, grown harder to reconcile. The repositioning now visible across the industry, the revised timelines, the renewed attention to questions once set aside as premature, suggest that a period of recalibration is underway.
The pattern extends well beyond any single company. It is baked into how technology monocultures tend to operate. When the incentive structure rewards scale above all else, the costs of that scale accumulate quietly until they demand attention.
What Clarity Requires
Understanding what makes human expression distinctive has become a practical necessity, one the AI era has made urgent rather than peripheral.
The leaders, editors, educators, and creators who navigate the next decade well will be those who understand, with precision, what AI systems actually do, which is pattern-match at scale, and where that capability reaches its limit, which is at the novel signal that the pattern-matching depends on.
Clarity precedes strategy. And the clearest observation available at this moment is that original human thought is appreciating in value: growing scarcer in the places where AI is most active, while the systems positioned to replace it quietly consume the very thing that made them possible.
The machine is eating its own tail. The question worth sitting with is what gets built while it does.
Awakened Dispatch is an independent publication committed to original analysis at the intersection of technology, culture, and human agency.
