But to fill the cracks = to directly incorporate AI output into parts of the proposal - is definitely forbidden by ISO.

I'm interpreting your use of the verb, 'to incorporate', here as "to copy and paste".

This past month I was calmly arguing with a meditation teacher that opiate dependants should be allowed to bring their opiate substitute medication with them on residential silent meditation courses. I made several points which I feel were quite convincing, but I stressed my final point which I feel was the most meritorious:

"If you disallow them to bring these medications, it means that instead of declaring these medications, they will sneak them in and hide them in their pillow cases."

The situation we have here with large language models is very similar:

"If you disallow people to use large language models in preparing their papers, then they'll conceal the fact that they used a large language model in preparing their paper."

Now I know that there are a lot of honest, forthright people here on the mailing list who will follow the rules and won't use a large language model if it is disallowed -- death before dishonour and all that. And I'm not doubting that there are members of society who always follow all the rules and never put a foot wrong -- I count myself as one of them -- I mean if I was a taxi driver, I'd declare every single fare on my income tax forms.

But for those of us who live in the real world, who weren't born yesterday, and who didn't land on this planet yesterday, I reckon that less than 10% of papers produced this past year were composed without some copy-pasting from a web browser tab navigated to openai.com. Personally I've never done it myself and I never will, because my integrity forbids me to do so, but I reckon that many people do it.

And so a new human skill is born: The skill of providing the best input into a large language model in order to get the optimal output. Over the coming years, individual humans will hone these skills. And who knows, similar to how we can buy "organic vegetables" at the supermarket, maybe one day we'll be able to download or buy "organic software" which was written entirely by humans. Similar to the situation with the vegetables though, these 'organic' products will probably be smaller and have less features (features being analogous to nutrients here).

I'm not saying that the ISO should change its policy about the use of large language models -- I'm just saying that we should occasionally glance at the elephant in the room. People are using large language models for _e_v_e_r_y_t_h_i_n_g_ nowadays, including their marital vows and their loved one's obituaries. I was at a public event recently for a bereavement charity and they had trouble with the projector so I went over to try help, and I saw that the web browser had loads of tabs opening asking ChatGPT what to say at the event, and specifically which terms and phrases to avoid using. Everyone's doing it.