In the modern world of artificial intelligence, machines are increasingly learning from the very content they create. This cycle, known as recursive training, happens when a model generates new data, and that new data is then fed back into the system to train the next version of the model. While this might seem like a way to endlessly expand knowledge, scientists have long worried about a phenomenon called "model collapse." In this scenario, the machine begins to lose touch with reality, gradually forgetting the true variety of the world it was meant to understand and instead converging on a narrow, distorted average. This is particularly dangerous in medicine, where AI systems are being used to generate synthetic patient records to help research rare diseases or protect privacy. If these synthetic records are flawed, the models trained on them could make dangerous errors, yet the tools used to spot these flaws have often been unclear.

A team of researchers from Norway set out to investigate exactly how this degradation happens when machines generate medical data over and over again. They focused on a specific type of error often called a "hallucination," which in the context of medical records means a patient entry that describes something impossible, like a person having a heart attack after they have already died. The researchers wanted to know if these obvious, record-level errors were the main problem, or if something more subtle was causing the data to fail. To find out, they built a rigorous experiment using real data from a large intensive care unit database. They took this real data and used four different types of AI generators to create new, synthetic patient records. They then took those synthetic records and used them to train new models, repeating this process twenty times in a row, creating a chain of generations where each step relied entirely on the output of the previous one.

The results revealed a startling truth: the data did not fail because the machines started writing nonsense stories about impossible patients. Instead, the data failed because it lost its diversity. As the generations progressed, the synthetic records became increasingly similar to one another, erasing the unique combinations of symptoms and conditions that exist in real life. The researchers found that within just five generations, the data had become so uniform that it could no longer teach a computer how to predict patient outcomes, effectively becoming useless for its intended purpose. The machines had not invented new, wild errors; they had simply stopped inventing anything new at all. They collapsed into a narrow, repetitive pattern that looked statistically safe but was clinically dead.

This finding challenges the way scientists and regulators currently look for errors. The study showed that the standard tools used to count "hallucinations"—which check if a specific record exists in a database of real people—were misleading. These tools often flagged thousands of records as errors simply because they were rare or unobserved, even though those records were medically plausible. In fact, the researchers found that the records flagged as "hallucinations" by these standard tools were often the only ones keeping the data alive. The real failure was invisible to these tools: the loss of the complex, intersecting patterns that make medical data useful. The machines were not lying; they were just becoming boring, and in doing so, they destroyed the predictive power of the data.

The researchers tested four different types of AI generators to see if they behaved differently. One type, a robust variational autoencoder, failed the fastest, completely losing the ability to distinguish between patients who would be readmitted to the hospital and those who would not. It did this by collapsing all its data into a single, average-looking profile. Other generators, such as those based on adversarial random forests and conditional generative adversarial networks, failed more slowly but still lost their ability to predict outcomes. They did this by drifting toward extreme values or inventing new, unrealistic patterns of symptoms. A diffusion model, which is a newer type of generator, behaved differently; it drifted toward the extremes of the data distribution, creating records with very high or very low values that were technically possible but statistically rare. Despite these different paths to failure, the end result was the same: the data lost its ability to represent the real world.

Crucially, the study found that the "hallucinations" that worried regulators and the public—records that were clearly impossible or nonsensical—were actually quite rare. In many cases, fewer than one percent of the synthetic records contained values that were medically impossible. The real danger was not the presence of these impossible records, but the absence of the diverse, realistic ones. The machines were not generating fake stories; they were generating a world that was too perfect, too average, and too uniform to be useful. This suggests that the current focus on catching individual errors in synthetic data is missing the bigger picture. The safety of these systems depends less on checking if a single row of data is factually correct and more on ensuring that the entire dataset preserves the rich, messy variety of human experience.

The implications of this work extend beyond just medical research. The researchers argue that the term "hallucination" is often the wrong word for what is happening in structured data. In the world of text, a hallucination is a made-up fact. In the world of medical data, the failure is often a loss of fidelity, where the data stops reflecting the true distribution of the population. The study suggests that we need to shift our focus from counting impossible records to measuring the diversity and structure of the data. If we only look for the obvious errors, we might miss the silent erosion of quality that makes the data useless for training future AI. The researchers emphasize that clear labels and provenance are essential. We must know when data is synthetic and monitor its quality throughout its lifecycle to prevent it from silently degrading the knowledge base of future systems.

In the end, the study serves as a warning against the assumption that more data is always better. When AI systems are allowed to feed on their own output without safeguards, they do not necessarily become more creative or more accurate. Instead, they risk becoming trapped in a loop of their own making, producing data that looks real but has lost the very essence of reality. The researchers conclude that preserving the diversity of data is not just a technical requirement but a fundamental necessity for the reliability of artificial intelligence. Without it, the promise of synthetic data to revolutionize fields like medicine could be undermined by the very systems designed to create it. The path forward requires a new vocabulary and new metrics that can detect the loss of diversity before the data collapses entirely, ensuring that the machines we build continue to reflect the complex, varied world they are meant to serve.