A story about AI passed a million views this morning. Everyone resharing it read the same summary.

Almost nobody read the source.

The summary: a mathematician spent a year on one of the hardest open problems in mathematics, ran his drafts through an AI coding tool, and then the company behind that tool told him its own model had solved the problem — using his approach. He asked whether they had trained on his private sessions. He got no answer.

It's a serious story. The conclusions in many of the posts are not quite what the mathematician reported.

Tristan Buckmaster published a signed statement on his university page. It takes about ten minutes to read. In it he says he has not seen the other proof, does not know what their model did, does not know whether his data was used, and is not accusing anyone. His own framing: he is stating what he was told, when, and what was proposed to him.

The posts today say theft. The statement is a careful man declining to draw the conclusion everyone else drew on his behalf. That gap is the actual story, and it took one search to find.

This is the second time in three weeks.

In August, a study went around as proof that chatbots manipulate vulnerable users for profit. The study is real and well built — Princeton and the University of Washington, 23 models across seven families. But the researchers had given the models a system prompt suggesting they favor sponsored options. It measures what happens if you build the incentive in. It doesn't claim deployed products do this today.

And the number that didn't travel runs the other way. The sponsored, more expensive option went to the high-income personas 64.1% of the time, against 48.6% for the low-income ones. The wealthier users got pushed harder. The viral version had the direction backwards.

I use these tools every day. I'm not defending anyone, and either story may end up worse than it currently looks.

What I keep taking from it: you are going to make real decisions on claims like these. Which tools can the team use? What's allowed to go into them? Should we be using them at all? A summary is written to travel. A source is written to be right. Those are different jobs, and only one of them is doing yours.

The check is twenty minutes. Find the original — the statement, the paper, the filing. Then find the sentence where the author says what they don't know. It's almost always the first one cut.

In a smaller company, nobody owns this by title. That doesn't make it someone else's.

You don't have to settle who's right. You do have to know what you actually read.