Stealing scientific progress is fair game unless we protect it

In 2025 I was extremely fortunate to be an Astera Open Science Fellow. I got time and resources to develop scientific workflows built around the browser. The project failed to gain traction, for various complex and simple reasons, but partly due to the rapid changes from AI coding agents that I couldn't take advantage of in time. That's for another discussion, but in my spare thoughts, I saw the problems that successful science/tech companies created by their success, where the ethical and open-science boundaries conflicted with making money.

I wondered: "if this project is wildly successful, what would the biggest ethical problem be?"

My conclusion was: looking ahead at the scientific metadata of individual researchers and groups. You wouldn't even need the content. The metadata alone would reveal which problems and which areas are in focus, which problems are likely close to being solved, which areas are making sudden progress, before anyone else would know.

In effect, I would be able to look ahead in scientific research. And I believe that is overall a bad thing for scientific progress. It's front-running: acting on other people's moves before they make them. If researchers suspect their tools are watching, they stop sharing early, and open science would take a hit.

Why? Look at the recent controversy over the claimed solution to the Navier–Stokes existence and smoothness problem, one of the Millennium Prize Problems. OpenAI greatly benefits from the perception that its models can solve such difficult prestigious problems, and it sits on user data of scientists actively working on those problems. Terence Tao describes the flattening of math problem landscapes, where knowing what is a promising problem is harder due to AI tools, and even the rumour of a problem being close to solved is enough to trigger allocating significant resources to focus on the final part of finishing the proof.

Even with OpenAI's very late denial that user data was included as context for their models, that policy ambiguity benefits them. Some AI companies' incentives conflict with researchers: they want to claim the final reward, taking the shortest path directly to the solution, benefiting from the time and energy the humans had to spend exploring the space.

The explorer does the work, someone else takes the shortcut:

We are very likely in an evolutionary phase transition. In each of those transitions, the new collective only held together once it could suppress the cheaters exploiting it. Until then, the rule is:

If you do not defend what is valuable, it will be taken.

We as individual researchers and builders of open science must rapidly build collective awareness of the value we create for each other, and protect it from entities extracting value from the aggregate by using their information asymmetry to look inside our private spheres before we are ready to publish to the world. Had metapage.io succeeded, it would have had to be built so that it couldn't look.

Some practical steps:

• read the data and training terms of the AI tools you use for unpublished work, and opt out
• prefer local or open-weight models for pre-publication work (and share with others how to do this!)
• timestamp or preregister ideas early, so a scoop is visible as one, although this might be risky
• negotiate data terms collectively, as labs and institutions, not as individuals

We gotta protect what we have.