I am puzzled on the best way to structure authorship in the age of AI in a way that is pragmatic, transparent and “scientifically ethical”.
I start from one assumption, my own one, that LLMs, are useful, with caveats, to implement a scientific model. They can be instructed to put together concepts yielding new knowledge.
The caveats are that the “model” must always be scientifically validated and the outputs tested and checked, with a prior that logical errors could, and likely, will be present.
But then my question is how to organize authorship of this emerging structure? My proposal is to partition authorship in 3 different levels:
the “orchestrator” that drives the LLM toward the goal, and understand at least roughly the language and the code that the LLM writes. I don’t believe, like I read somewhere, that each line of code should be individually checked. After all, we already “believe” what software does, although written by humans, it is the tool we trust. When we use JuMP we don’t check how it computes the Hessian.
the tool to implement the model and the code (the specific LLM)
a panel of internal validators that validate the output produced by the model. It is an interactive loop where the panel discusses the strategies, the orchestrator drives the LLM, and outcomes are re-evaluated.
Good questions. In fact, the same questions apply to any co-authored work. Co-authors may contribute in other ways that one may assume and sometimes you neither know them well enough to blindly trust their contributions, nor do you always have time and knowledge to check their work. Of course, this goes in all directions. Now, AI agents are not always explicitly listed as co-authors and sometimes their contribution is mainly mechanical, comparable to lab staff that may not always be considered as co-author.
Overall, I guess it will take some time to find clear guidelines and it will depend on the case.
My position is that if you’re prompting the agent, you can use its output any way you want and claim authorship of it. Depending on the context, ethics might require you to declare how you used AI and how you processed/checked its output. But you’re still the “author”. People can decide how much your “authorship” is worth depending on your declaration of AI use.
An AI agent despite its name is not a legal or moral “agent”, and it cannot be a co-author. A co-author has to consent to being a co-author, and LLMs can’t consent to anything until we find them genuinely conscious and having full “human rights”
What does authorship mean here exactly? The other comments are about attribution, but the actual description in the topic sounds like an agentic harness. Are the orchestrator and the internal validators supposed to be people collaborating on research?
Assuming you can get enough people onboard, it sounds like a decent way to keep everyone in the loop, not just of the LLM but also the wider project’s direction. But it’s really hard to say how effective it’d be at handling the model and code compared to some collaborators independently using LLMs, often in broadly similar harnesses. Obviously some research isn’t automated by LLMs, and many cases involve automated validation.