@elduvelle in my experience you need to find a thing they care about and then tell them about how "AI" damages that thing.
For example I've had success with a data-protection angle recently for people in the medical sector who do really care about the confidentiality of their clients
@aesthr
Thanks for taking the time to write this!
@elduvelle
I also take the approach of focusing on a thing they care about. The feeling "it's just not going through" is shared! Some reply "I know you are right, but I'll continue to use it nevertheless". But it is necessary and useful to explain once and again; like planting a seed, the effect is not immediately visible, and often it simply won't germinate. But sometimes it works.
A minor complaint on "LLMs are simply not capable of producing new knowledge". In mathematics, new knowledge is certainly available through the use of LLMs, and technicalities or stating that it was implicit in the sources won't convince my colleagues. What I do is try to show that the actual cost is overwhelmingly bigger than the benefit.
Could explain?
... In mathematics, new knowledge is certainly available through the use of LLMs...
@Davide_Sandini OpenAI (and also Anthropic, to a lesser extent) have released "drafts" of proofs or counterexamples of hundreds of previously open problems, often with a formal verification. It is obvious when reading the drafts that there was no mathematician expert writing them (they show no understanding of the issues and the importance of each part) but with effort they can be understood and most of them are correct. This means that we now know the truth status of certain questions that have been studied for decades without success, and this through the use of LLMs.
I still refuse to write that "the LLM proved it". There has to be a human prompting it cleverly. But the LLM was essential.