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.
What is your thought ?