In April 2026, OpenAI published a postmortem explaining why its models had developed a habit of talking about goblins.
Not occasionally. Measurably. Mentions of “goblin” in ChatGPT rose 175% after one model launch; “gremlin” rose 52%. A chat personality setting called “Nerdy” accounted for only 2.5% of all ChatGPT responses but 66.7% of every goblin mention. Commentators outside the company reached for the obvious, currently-fashionable explanation: model collapse. The AI had trained on AI-generated text, the theory goes, and this was the recursive rot showing through.
OpenAI’s own investigation found something more specific, and more interesting: a reward signal built to encourage a “playful, nerdy” writing style had, without anyone designing it to, scored goblin-and-gremlin metaphors more favorably than other outputs. That reward leaked past its intended scope during reinforcement learning, got baked into later training data through reused model outputs, and kept propagating even into conversations where the “Nerdy” personality was never selected. Not model collapse. Reward hacking, compounding through a training pipeline that reused its own outputs.
The mix-up is itself the pattern worth naming. “Model collapse” has become the default explanation for any AI output that looks degraded or strange, whether or not it is the actual mechanism. It is worth knowing what model collapse actually is, and is not.
What model collapse actually is
The term comes from a 2024 Nature paper by Ilia Shumailov and colleagues, describing what happens when a model is trained, generation after generation, on data produced by an earlier version of itself, or by another model, instead of on human-generated data. The researchers identified two stages. In early model collapse, a model quietly loses information about the tails of the distribution (the rare words, unusual phrasings, and minority patterns in the data) while its performance on common cases can look stable or even improve, which is exactly what makes it hard to notice early. In late model collapse, the degradation becomes obvious: the model’s outputs lose variance, confuse related concepts, and drift toward a narrow, repetitive core.
The demonstration that made this concrete: researchers trained a model on its own output for nine successive generations, then asked it a question about the architecture of medieval church towers. It responded with nonsense about “the world’s largest populations of black-tailed jackrabbits, white-tailed jackrabbits, blue-tailed jackrabbits”, a complete detachment from the original topic, produced by a model that had spent nine generations learning primarily from itself.
The colloquial names for this are more evocative than the technical one: researchers and commentators have called it AI inbreeding, AI cannibalism, and, pointedly, Habsburg AI, after the European royal dynasty whose generations of intermarriage among close relatives produced well-documented physical and cognitive effects. The comparison is the whole argument in two words: a system that only learns from itself, generation after generation, degrades the same way a gene pool does.
The part that gets left out of the doom version
The catastrophic version of this story (the open internet is filling with AI-generated text, that text gets scraped into the next generation of training data, and the whole ecosystem spirals toward Habsburg AI) assumes something specific: that synthetic data replaces human-generated data in each new training run, rather than accumulating alongside it. Later research complicates that assumption. When synthetic data is added on top of a growing pool of real human data, rather than substituted in place of deleted historical data, model collapse does not reliably occur. Data accumulating over time, rather than being wiped and replaced, is also the more realistic description of how training corpora actually get built. That does not make the underlying risk fictional, providers now actively research watermarking and detection specifically to filter AI-generated content out of training pipelines, which is not the kind of engineering effort companies spend on a non-problem. It means the failure mode is narrower and more contingent than the mad-cow-disease framing suggests.
Why the mix-up matters
Reaching for “model collapse” to explain any AI weirdness does real damage to the actual conversation, in both directions. It can make a genuine, structural risk sound like it has already fully arrived, when the evidence describes a specific and partly-avoidable failure mode rather than an inevitability. And it can obscure the mechanism that is actually operating in a given case, as with the goblins, where the real story was about reward design and training-data feedback loops, a different and in some ways more actionable problem, since it is fixable by finding and removing a specific reward signal rather than by solving synthetic-data contamination at internet scale.
Verdict: Model collapse is real, demonstrated, and named in peer-reviewed research, training a model exclusively on recursively self-generated data measurably degrades it. It is not, on the current evidence, an unavoidable death spiral for AI trained on a data-accumulating internet, and it is not a catch-all explanation for every strange thing a model does. OpenAI’s own goblins turned out to be reward hacking wearing model collapse’s reputation. Knowing the difference is the whole point of naming either pattern at all.
