Scientists from Princeton University published a paper this month with a major reveal: they created a system to help stabilize fusion reactors.
Stable fusion power has always been one of the big dreams of science. It means recreating the reactions that power the stars. Done right, it's a near limitless power source, delivering all the energy humans need. It's a great dream. And it's theoretically in reach.
Up until now, there's been a big problem: Researchers can't fully control fusion reactions. Fusion plasmas are turbulent -- they have tiny, fast-evolving instabilities, sudden bursts of heat and particles. To control the reactions, researchers need to track these fluctuations. But until now, they haven't had tools that could measure them. Plasma instabilities span micrometers and happen in milliseconds. They're both smaller and faster than any diagnostic tool can resolve.
If scientists can't see these details clearly, they can't pinpoint what triggers the instabilities. They can't predict when and where disruptions will happen. And they can't achieve the big dream: keeping plasma stable and efficient.
That's what makes this new announcement so intriguing. The Princeton device, called a Diag2Diag attempts to fix this blind spot, delivering the missing data. There's just one problem: their system generates measurements from data that was never collected. It's entirely made up by AI.
The scientists behind the study (which included Columbia University and two Universities in South Korea) make no secret of AI's prominent role in the results. No one's trying to fool anyone. But that makes it more disturbing, because more and more, the scientific community sees synthetic AI data as a viable option.
The Diag2Diag tool has that strange name because of how it works: It constructs one diagnostic signal from another diagnostic signal, rather than measuring the actual fluctuation. In other words, the tool doesn't actually measure anything. The AI trains on moments when measurements of the fluctuations do exist, and creates patterns, statistical relationships. It reconstructs the "missing view" by combining partial measurements with AI-driven simulations. The results are impressive. You get a usable, sharper view of plasma behavior.
But built into this view are guesses. Good guesses, sure, but guesses nonetheless.
The scientists are transparent about all this, although they put it in language only understood by other scientists. Said one, "our objective is not to replicate the entire physics of Magnetohydrodynamics turbulence in plasma from first principles, but rather to learn empirically grounded correlations among multiple diagnostics."
Put simply? AI isn't recreating actual physics. It's pattern-matching. And because scientists don't ever see the real data, there's nothing to verify it against. The guesses can be wrong and no one would ever know.
In some ways, this isn't new. Scientific data has used AI systems before to enhance measurements through fancy techniques like interpolation or denoising. But the Diag2Diag tool crosses a line: it generates measurements that were never taken, based on correlations learned from different tools measuring different physical quantities. Like the name says, it goes from diagnostic to diagnostic. It's a bit of a leap.
Of course, if it was just a theory in a published paper, there wouldn't be much concern. But the research directly feeds into ITER, the International Thermonuclear Experiment Reactor. (Scientists just called it "ITER" now, avoiding the very scary full name). The $25 billion fusion reactor is under construction in France and will potentially use Diag2Diag synthetic data to manages plasma stability. The researchers conclude their paper saying they offer "a powerful tool for ELM control development in future reactors like ITER".
How big a risk is this?
If diagnostics in fusion reactors are relied on with incorrect AI data included, researchers may think they detect plasma behaviors that don't actually exist. This could result in misguided adjustments that do more harm than good. Phantom readings could destabilizing the plasma more instead of stabilizing it.
Even worse, acting on bad AI data could make disruptions more violent. In large reactors, plasma disruptions release enormous force -- the equivalent of a magnitude 7 earthquake. It could damage reactor walls.
When you hear things like damaged reactors, you might imagine Chernobyl. But fusion reactors don't risk a chain reaction and don't generate a radiation cloud when everything fails. It's part of the allure -- fusion power is much safer if you can make it work. But if plasma starts hitting the walls of the reactor like a freight train -- very possible if fluctuation levels are increased -- it would be catastrophic for the facility. ITER would be destroyed.
Not the end of the world, but a major industrial catastrophe.
And it doesn't stop with fusion power.
Synthetic AI data is used in other fields. The Princeton paper's conclusion makes their broader ambition clear: "The implications of this work extend well beyond the immediate application to magnetic fusion. The multimodal super-resolution capabilities developed here can significantly impact areas such as laser fusion data analysis, accelerator data analysis, and molecular dynamics research."
AI can fill in data gaps in other domains, where diagnostic data is incomplete. It does an amazing job of it. It fills in medical imaging of tumors between scanning intervals, closes gaps in structural monitoring of aging bridges and nuclear facilities. It makes guesses in gaps from early warning systems for earthquakes, tsunamis, or reactor failures.
Where real lives are at stake, AI is more and more at the heart of the process, often used to fill in the blanks with synthetic data. And the results are convincing. AI is detailed and confident. But when it's simulations based on simulations, AI can guess wrong. And the damage can be very real.