By Mari Hansen Tonight, an AI model is being fed never-taken nighttime city photos on a developer’s laptop. No lens, no tripod, no late-night walk up a hill while holding a phone. The pictures only exist because they were created by another model. The odd thing is that this is now regarded as typical.
The statistics surrounding synthetic data have begun to seem almost too handy. According to industry estimates, artificially generated data accounted for over 60% of the data used by AI applications last year, and this percentage is rising. It appears that investors think this is the next big unlock. Its creators present it as the solution to data scarcity, privacy laws, and the awkward reality that the open web has, for the most part, been scraped clean. They might be correct. It’s also possible that, while very few people outside the field are closely observing it, we are witnessing a silent bubble form in real time.
| Field | Details |
|---|---|
| Topic | The Synthetic Data Bubble in AI Training |
| Estimated share of synthetic data in AI applications (2024) | More than 60% |
| Key concern | Model collapse, loss of human nuance, declining edge-case accuracy |
| Notable academic voice | Kalyan Veeramachaneni, MIT Laboratory for Information and Decision Systems |
| Industry stage | Early but accelerating; venture capital flowing into synthetic data startups |
| Common modalities | Language, image/video, audio, tabular data |
| Primary risks | Privacy leakage, bias amplification, recursive degradation |
| Real-world use cases | Self-driving simulation, medical imaging, financial transaction modelling |
| Human-labelled data | Still considered the gold standard for high-stakes domains |
| Market trajectory | Expected to keep growing across industries |
These days, if you walk into any AI startup, you’ll hear the same grievance expressed in a variety of ways. The useful data is no longer available. Llama, DeepSeek, GPT-3, and GPT-4 all drank from the same well, which is shallower than most people realized. Medical records, payment disputes, supply chain logs, and customer chat transcripts are typically locked behind enterprise firewalls. This type of messy, real-world content teaches a model how a hospital plans its night shift or how a São Paulo fraud team flags a suspicious wire. This information is costly, difficult to obtain, and frequently untouchable by the law. As a result, businesses have begun taking the next sensible step. They fabricate it.
To be fair, synthetic data is not a euphemism for fraud. A small sample of real data is used to build a generative model, which is then asked to generate additional examples that resemble the original’s statistical patterns. When done correctly, it protects privacy and fills in the gaps left by real-world examples that are too uncommon or dangerous to gather. Imagine using hailstorms in rural Pakistan to train a self-driving car. Waiting for the weather is not something you should do. You would mimic it.
However, some researchers believe that we’re getting ahead of ourselves. The slow deterioration that occurs when models continue to train on outputs from previous models, such as a photocopy of a photocopy of a photocopy, is known as “model collapse.” On the surface, every generation appears to be fine. Only in edge cases, subtle errors, and answers that sound confident but feel a little off do the flaws become apparent.
The irony in this situation is difficult to ignore. Another billion-parameter model is not what AI most urgently needs at the moment. Somewhere in a fluorescent-lit office, a weary human expert is ranking outputs and taking notes. Real human judgment—the compromises, the unwritten guidelines, the times when someone says, “We don’t do it that way here”—remains the foundation. That anchor can be circumvented by synthetic data. It cannot take its place.
There’s also a more subdued outcome. The value of real human information—a real conversation, a real decision, a real error—only increases as synthetic content proliferates. It’s possible that in the future, compute and capital won’t be the most valuable resources in AI. It’s someone who genuinely understands what they’re doing and is prepared to sit down and explain it. It’s still unclear if the market will fairly compensate for that.