The Preprint That Ran

That is where this story actually starts, and almost nobody tells it that way. Kwon Young-wan uploaded first, listing Lee Suk-bae and Kim Ji-hoon ahead of himself. Lee uploaded the second — a longer author list, no Kwon. Lee's account afterward was that Kwon had left the Quantum Energy Research Centre four months earlier, that the papers weren't finished, and that nobody had authorized the post. A Nobel splits three ways at most. Work out for yourself why a person hits submit early.

So before LK-99 reached a single outside skeptic, the first round of scrutiny had already collapsed — inside a small research centre whose members could not agree on whose name went where.

What they had was lead apatite doped with copper, Pb₁₀₋ₓCuₓ(PO₄)₆O, and a phone video of a gray fragment tilting off a magnet. Not floating. Tilting. Within forty-eight hours that clip was global. Prediction markets opened. People who could not define the Meissner effect on Friday were arguing about flux pinning by Sunday.

Then the labs started cooking. Zhiqi Liu's group at Beihang got a semiconductor with high room-temperature resistance. India's National Physical Laboratory got weak magnetism, tried again with purer precursors, and got a diamagnet. Ross Colman's team at Charles University in Prague ran the synthesis live. Bernevig and Schoop at Princeton found at least three distinct phases in a single sample. And then Wu Wei and Luo Jianlin at Beijing's National Laboratory for Condensed Matter Physics put a name to it: Cu₂S impurity, structural transition around 385 K, sharp resistivity drop. Not superconductivity. A contaminant with excellent comic timing.

Claim to consensus: about seventeen days.

The lesson everybody took away was that preprints route around peer review and hype fills the vacuum. The first half is true. The second half misnames the organ that worked. Nothing in those seventeen days was peer review. Peer review is pre-publication gatekeeping — assigned readers, a queue, an obligation to file, no contact with the material. What happened to LK-99 was replication: post-publication, voluntary, unassigned, carried out by people who made the compound themselves. Those are different jobs at different points in the lifecycle, and it matters which one you are praising. The referee's job is to stop a bad claim before it travels; the swarm only engaged after the claim had reached tens of millions. What the swarm has that the queue does not is speed and hands on the sample. What it does not have is coverage. No queue means no obligation means no guarantee anyone shows up at all. LK-99 summoned its own auditors because it was spectacular, and spectacular is not a property you can schedule.

Andrei Bernevig's verdict was that "the social media, memes, etc., have been completely detrimental to progress," and that he hoped nobody would "ever do science like this again." He is right about the noise. He is also describing the recruitment mechanism for the swarm that solved it. You do not get the seventeen-day correction without the thing that put LK-99 in front of ten thousand condensed-matter physicists over one weekend. Same pipe.

The pipe is the problem, and not in the way it's usually named. Amplification and correction share a channel but not an audience. The levitation clip reached tens of millions of people. The Cu₂S structural transition at 385 K reached the people who were already going to find out.

But speed is the smaller half of it. Claim and retraction have always traveled at different rates — that is ancient. What is new is that only one of them has an address. The preprint is 2307.12008: a permanent identifier, a permanent home, a citable object that will still resolve in thirty years. The refutation is a dozen scattered arXiv postings, a Nature news item, a Physics World writeup, a thread. It never got consolidated and it never got a name. There is no document called LK-99 Was Cu₂S.

So the record that gets ingested and inherited carries the claim with far more mass than the disproof — and it does so at exactly the moment when the primary reader of arXiv stops being a person. Ask anyone what LK-99 was and they will tell you it was the fake superconductor. Ask what Cu₂S is and you get nothing.

Underneath that sits the smaller, older problem: a preprint carries the full format of authority with none of the filter. Two columns, figures, references, an identifier that looks enough like a DOI to pass. To anyone downstream — a reporter on deadline, an aggregator, a model ingesting arXiv wholesale — a preprint and a reviewed paper are the same object. That was always true. It just never got load-tested in public before.

Which brings it back to Kwon. Nobody gets a Nobel for finding Cu₂S. Wu Wei and Luo Jianlin closed the case in seventeen days and their names are already gone, and that is the same asymmetry that put two papers on arXiv two hours and twenty minutes apart. Discovery is rewarded, disconfirmation is not — at the lab bench, in the journal, on the timeline, and now in the index. The credit structure that made somebody hit submit early is the credit structure that leaves the correction unnamed.

So the file on LK-99 got closed under the summer science lost its mind, which is roughly the opposite of what happened. It was one of the fastest and cleanest self-corrections on record. It was also a lottery. Most wrong preprints are not spectacular. They are modest and wrong, and they sit there indefinitely carrying full format-authority, because nobody is excited enough to spend three weeks and a furnace disproving a small claim.

The next one will arrive with better video. And the system that summarizes it for you will have read 2307.12008 and never once heard about the seventeen days.