We're in Nature with an opinion piece on how researchers should respond to #ChatGPT and conversational AI technology more generally!
It's been an interesting experience to reach consensus in an interdisciplinary team of scholars (2 psychologists, 1 computer scientist, 1 philosopher and me, an NLP-er).
We list 5 priorities:
1. Hold on to human verification
2. Develop rules for accountability
3. Invest in truly open LLMs
4. Embrace the benefits of AI
5. Widen the debate
Great work. Thank you!
You are spot on where you write: "This defies today’s binary definitions of authorship, plagiarism and sources, in which someone is either an author, or not, and a source has either been used, or not. Policies will have to adapt, but full transparency will always be key."
I think this continuum applies to the question of accountability as well (I developed that a bit here: https://sentientsyllabus.substack.com/p/silicone-coauthors ). There, I propose to leave the decision of co-authorship to the authors. It is certainly not deceptive to do that, which distinguishes it from gift-, ghost-, and guest-authorship. Transparency is key.
An unresolved implication is a desire to document process. That would be great, but adding another _dimension_ (progress) to linear text is conceptually difficult and I am not aware of technical approaches.
Your proposal for non-profit LLMs is interesting, but will ultimately run up against the same concerns as private sector LLMs - simply due to the need of significant funding for training and operation. An alternative might be public LLMs, modelled on our public library systems. I have not seen that discussed yet. Certainly very doable at EU scale.
Thank you for this contribution.
@boris_steipe Thanks -- those are interesting additions and alternative suggestions!
I think the difference in position about authorship disappears when we rethink what authorship will mean in the future. And I think I agree on public LLMs.
Not sure if I fully understand your point about documenting process.
We expect authors' transparency "to what extent" AI technologies were used. In my analysis I proposed some qualitative language, but whether that is enough transparency can be questioned. It would not make the actual flow of ideas explicit or verifiable - that is what I mean by "documenting process"; a small point, but central to the debate.
In the absence of verifiability, all we have is trust.
Realizing that has its own implications.
🙂
