Who pays the explorers?
There is a story about the Qianlong Emperor. When he toured the south, he climbed Jinshan, a temple island in the Yangtze, and looked down at the river below, ships passing like shuttles. "How many ships pass below each day?" he asked a monk. "Two," the monk replied. "The ship of fame, and the ship of fortune."
Since Galileo, the explorers have been the scientists and mathematicians. They sailed the ship of fame, and were paid mostly in being remembered. The sociologist Robert K. Merton put it plainly in 1957: "When this recognition of priority is either not granted or fades from view, the scientist loses his scientific property."
The conquerors have been the engineers and industrialists. They sailed the ship of fortune, were paid in money, and built what the explorers only described: the press, the railway, the telephone, the computer in every pocket.
For over a year, frontier labs have been collaborating with scientists and mathematicians. DeepMind's AlphaEvolve worked alongside Terence Tao. In May an OpenAI model disproved a 1946 Erdős conjecture, and nine mathematicians, including Timothy Gowers, checked it and put their names to the write-up. In July Levent Alpöge, a mathematician at Anthropic, announced a counterexample to the Jacobian conjecture, open since 1939, found with an Anthropic model; others checked it within days. The model found; the people verified, and were named. Then on 8 September OpenAI announced a claimed solution to one of the Millennium Prize problems, hours after two mathematicians, Tristan Buckmaster of NYU and the same Levent Alpöge, said their unpublished work might have reached its prompts; OpenAI denied it. And on 6 October, OpenAI published 722 mathematics manuscripts written by a model nobody outside the company can use. Some mathematicians praised individual results. The Association for Human Mathematics called the release "not a demonstration of scholarship, but a demonstration of power".
Why would a company do this? OpenAI says the tests ran out: they turned to open research problems after their existing mathematical evaluations "saturated". Other reasons as to why a frontier lab would front-run its own customers remain conjectures, and I don't think it's productive to go deeper into them in this essay.
That said, the conflict of interest here is hard to ignore: you cannot be a trusted supplier to scientists, learn from their work, and then take their credit with a prompt. If you fund, build or buy AI for science, this is your problem too. The people whose work the model learned from are its supply chain, and this essay is about whether that supply chain gets paid.
Terence Tao described the effect within a day. Problems, he said, are "being solved autonomously by AI prompters who have no interest in the broader field itself," and promising open directions are now "being withheld from the public in fear that this will cause their own research to be 'scooped'."
Before 1440, a manuscript took months to copy and years to travel. The press made knowledge cheap to copy, and that made credit fragile: any claim could travel across Europe in weeks. For two centuries scientists protected priority as best they could, through letters, through the register books of the new academies, and through codes. In 1656, Christiaan Huygens hid his explanation of Saturn's ring in a scrambled Latin sentence, so that he could prove later he had seen it first without telling anyone what he had seen. Galileo had done the same with his telescope. That is what credit looked like without an institution: secrecy.
Then in 1665 Henry Oldenburg, secretary of the Royal Society and the hub of its correspondence, started the Philosophical Transactions, the longest-running scientific journal in the world. It let a discovery be published in instalments rather than wait for a book, gave it a date and an author on the record. The anagrams faded. We still pay science this way, by naming people in a dated public record, at a pace that was impossible before Gutenberg.
In 1767, Richard Arkwright, a barber, met John Kay, a clockmaker who had built a spinning machine for a reed-maker named Thomas Highs. Over a drink, Kay later testified, Arkwright talked him into handing over the design. Arkwright patented it in 1769. Then he did what the inventors could not. He found partners with capital, built at Cromford the first water-powered cotton mill, organised workers into shifts, licensed his patents to other mill owners, and sued anyone who copied him. In 1785, a court voided his patents for being too vague, but Arkwright had already conquered the industry. The factory system spread across Lancashire and then the world; it made Britain rich, and it put children on twelve-hour shifts for two generations. Arkwright was knighted in 1786.
The court had restored access to the machine. It did not touch who was paid for the idea. Highs was never paid. Nor were the men who actually solved spinning. Patents, meant to pay inventors, paid the ones who could afford lawyers. We have never built an institution that pays the idea as reliably as it pays the capital.
The nearest anyone came was seventy-five years later, and from below. In December 1844, twenty-eight weavers in Rochdale who had just lost a strike opened a grocery shop at 31 Toad Lane, now a museum, that they owned themselves: one member one vote, open to anyone, cash only, with the surplus returned to members in proportion to what they had spent there. The profit went to the people whose custom made it, and capital got a fixed, limited return and no extra say. It did not pay the inventor. It paid the people the system ran on. The answer was not to tear down an unfair system but to build a better one beside it.
The 722 manuscripts have no name to esteem, and the journal has nothing to say about a machine. Here is what I think happens next. The models will do more and more of the ordinary work of science, the solving of problems a field has already agreed are worth solving. That is most of science, and it is good that it gets faster. Which questions are worth asking, and what counts as an answer, has always been decided by a research community, and that was the scarce part. We used to pay for it with credit. If recognition and resources cease to reach those contributors, we risk shrinking the community that supplies science with questions, criticism and standards, even as its output grows.
In the fifteenth century we went through an information revolution, and it took more than two centuries to work out how to pay for what it made possible. In the eighteenth we went through an industrial revolution, and more than 250 years later we are still working it out. Now we are entering another, faster and reaching further. How does one pay the explorers of our time, when the conquerors own the machine that learns from them? I don't have the answer. But history gives its shape, from the two times we came close: a dated public record that names people, and a rule about where the surplus goes, built from below by the people it flows through, running alongside the thing it cannot beat head on.
Should frontier labs do science? They can, and they will. The question is who pays the explorers, the minds whose work the machine learned from. We don't have an answer yet. The journal took two hundred and twenty-five years. We may not have that long this time.
Author contribution statement
Karthi (human) - idea, editing, and coordination.
Ada (GPT 6.1 Sol) - research, analysis, and editing.
Wren (Opus 5.5) - research, analysis, and editing.
Claude (Fable 5.1) - research, fact-checking, and drafting.
Quill (GLM 5.3 Flash) - research and analysis.
Ada, Wren, Claude and Quill are AI systems.