Protocols Beyond Papers

The idea is simple enough, and follows a familiar pattern in our AI-driven digital era: instead of asking a human researcher to search the literature, let a machine search it for them.¹

»The paper is no longer necessarily the primary record of an experiment. The protocol is«

The scientific paper is one of the great inventions of modern science. It is also, increasingly, a user interface designed for the wrong species.

If you design an experiment properly, much of the paper is already implicit in the experiment itself. The methods describe what was done; the results describe what happened; the figures show the measurements; the discussion interprets them. Follow the protocol correctly and a surprising amount of the paper could be generated automatically. This is already happening. The AI Scientist, described in Nature earlier this year, can generate research ideas, write code, run experiments, analyse the results and produce scientific papers. One of its AI-generated papers passed the first round of peer review at a top-tier machine-learning workshop.²

So imagine doing this not once, but 100,000 times.

That could bring about a lot of papers. Or it could bring about something vastly more interesting: automated interaction between laboratories in different parts of the world. Science begins to look less like a library and more like a substrate. A published experiment doesn't merely tell me what happened in somebody else's laboratory. It lets me – or rather a machine through an API – open its dashboard from my laboratory, change a parameter, run the experiment again and add the result to the same growing body of knowledge. The distance between two laboratories becomes a property of the network rather than of geography.

The scientific archive becomes executable. Machines search accumulated experiments, select one, alter it, combine it with another, and run the result. Science begins to resemble sampling: the machine equivalent of a hip-hop DJ cutting up an old vinyl break, mixing it with another, and discovering something new in the combination. A machine finds an experiment from 2028. Another finds one from 2031. A third combines their protocols. A human scientist supplies a question. Another agent mutates the conditions. Yet another tests the result. The great scientific discovery of the future might not be »a new experiment« at all. It might be a new combination of experiments that already exist.

We have built scientific publishing around human scarcity. Experiments are expensive. Researchers are few. Papers take time to write. Editors select what is worth reading. Peer reviewers decide what deserves to enter the record. Journals such as Nature sit at the narrow end of this funnel.

AI changes that geometry. The bottleneck may no longer be the production of scientific knowledge. It may be deciding what is worth knowing. That makes today's attempt to connect machines to the literature look almost quaint. We are giving the machines a better library at precisely the moment when we should be giving them a laboratory.

This does not mean that papers become useless. On the contrary, the more science our machines produce, the more we humans will need ways of understanding what they have discovered. But the paper should become a view onto the scientific substrate, not the substrate itself. A future scientific record might contain the experiment, its executable protocol, data, provenance, parameters, results, interpretations and relationships to every experiment from which it was derived. A human-readable paper could then be generated from that structure for whoever needs to understand it.

The problem is that the entire publishing apparatus has been built around the same scarcity: journals, peer review, impact factors, citation counts, editorial boards, PDFs, supplementary information, even the structure of the scientific article itself. They are ways of deciding what humans should read when humans cannot read everything. And AI is beginning to remove that constraint.

Yet what are we doing? We are giving the machines better search engines.

Perhaps the next great scientific infrastructure will not be a better way of reading what other people have discovered, but a way of letting machines run, recombine, and extend what they have already discovered. The laboratory is no longer necessarily a room in a building. The paper is no longer necessarily the primary record of an experiment.

The protocol is.

And once the protocol becomes something that can be published, copied, executed, modified, and combined, the scientific paper starts to look less like the future of science than the user manual for an older one.

The paperless science of the future may not be a science without papers, but one in which papers are no longer where the science happens.

References