Every year I read several articles announcing that fusion is nearly here. Every year it is not. The joke is that fusion is always thirty years away, and like most jokes that survive fifty years, it is funny because everyone assumes they know why it is true. Almost nobody does.

In 1976 the Energy Research and Development Administration (the agency that became the Department of Energy two years later) published a program plan for magnetic fusion. It did not offer a date. It offered five, each attached to a price.

The United States did not choose one of the first four. It chose the fifth, and then it did not fund the fifth. Actual appropriations crossed below the fusion-never line around 1980 and stayed there for the next three decades, settling near $300 to $400 million a year while the plan's cheapest scenario called for triple that.

So the cliché is true in its outcome and wrong in its premise. Fusion did not miss a deadline. It hit the one we paid for. If you want the counterfactual, the moderate path put a demonstration plant on the grid in 2005, which is now twenty-one years ago.

That is the frame I would ask you to hold while reading any fusion headline, including the one below, which is a good result described badly.

Fusion is the reverse of the reaction in a nuclear plant. Instead of splitting heavy atoms, you force light ones together. Two hydrogen isotopes (deuterium and tritium) fuse into helium and throw off a neutron carrying most of the energy. No long-lived waste, no meltdown physics, and the fuel is effectively water and lithium.

The catch is the temperature. You need roughly 100 million degrees, at which point matter is a plasma, and no material container survives contact. A tokamak solves this with magnets, holding a doughnut of plasma suspended so it never touches the wall.

The number everyone quotes is Q, the fusion energy out divided by the heating energy in. Q above 1 means the plasma produced more than it was fed. This is where most coverage quietly goes wrong, because there are three different Q values and the press usually reports the friendliest one.

Here is the honest scoreboard as of this month.

Notice that only one line in that table is a near-term, specific, checkable prediction. Commonwealth Fusion has said two years, not thirty, which at least has the virtue of being falsifiable in 2028. Lev Artsimovich, who built the first tokamaks, is supposed to have said fusion will be ready when society needs it. Fifty years of budget data suggest he was describing a procurement process, not a physics problem.

DIII-D is the machine in this week's story, and its place in that table matters. It is a research tokamak, and the hardest unsolved problems there are not about temperature. They are about control.

On 6 September, ScienceDaily ran a Princeton release under the headline "AI can now control fusion plasma faster than humans can react." The system is called PACMAN, for Prediction And Control using MAchiNe learning. A national laboratory named its autonomous control software after a character whose entire occupation is moving through a maze eating everything while being pursued. Nobody in the release mentions this.

The underlying paper is Rothstein, Farre-Kaga and colleagues in Nuclear Fusion, open access, and it is worth reading. It is also not the paper the coverage describes. It is an architecture paper. It specifies a software framework and gestures at five experiments, most of which are written up somewhere else entirely.

That distinction is the whole fact check.

The science went public ten months before the press release announcing it as new. This is ordinary and nobody is hiding anything, but it is worth knowing that "a new paper" in a university release often means "a paper we have decided to promote."

The first is about failure. If any input carries an error, the controller issues no command. If two controllers request conflicting actions, the output block sends nothing and tells the operators. Then the paper does something press releases never do. It questions its own safety choice: this is done in an abundance of caution, the authors write, "but of note sometimes inaction can potentially be more dangerous than a bad action."

I have sat through a great many discussions of clinical decision support, and I have never once seen that sentence written down. The default when an algorithm is uncertain is almost always to suppress the alert, and the harm of the suppressed alert is almost never counted. A plasma physics journal put it in print because plasma does not accept an abundance of caution as an answer.

The second is about org charts. The paper sorts every algorithm into three levels, and then explains that the levels deliberately copy the DIII-D control room. Level 1 is the engineers who own the hardware. Level 2 is the physics operators who translate. Level 3 is the experimentalists who ask for things. A Level 2 algorithm can refuse a Level 3 request the same way a physics operator can tell a visiting scientist that the plasma they want is not available today.

The AI lives at Level 3, where requests originate. It does not live where the limits are enforced. That is not a technical detail. That is a governance decision rendered in C.