I think this is an interesting discussion and there are lots of good points being made all round - however I would remind folk to review the Forum & Community guidelines and make sure the discussion is respectful at all times, even when you strongly disagree.
In my view Construct is primarily a tool for human creativity and expression. I am not personally particularly interested in AI-generated games or content. Other tools can do that and people can use them if they want, but I don't think it's something we should do with Construct. Using an AI assistant to help solve problems, sort of like an in-app support advisor, is probably a more reasonable idea to discuss.
I think that the intense hype around AI makes this harder to evaluate. I suspect many companies are cramming their products full of AI not because it's been proven to work well and be useful, but because they have investors or shareholders to impress, and they benefit from the AI hype. We are an independently owned company and we have no need for such things. On the other hand, AI is proving to be genuinely useful in some areas and it probably is here to stay and will significantly disrupt some things. I still don't think it's clear though what things are going to change and to what extent. Were it to become the case that customers expected all software to have AI assistants and tools that didn't became irrelevant and ultimately failed, then I think we would be forced to add some kind of AI assistant to stay in business. I don't personally see that point having arrived though, and we can act on it if or when it becomes the case. I think there is still significant risk around whether the big AI companies will prove to be economically viable (there may be a big crash on the horizon); risk the pricing for AI services ends up increasing massively; risk that lots of existing AI integrations end up being removed due to being expensive/ineffective/unwanted, etc. So I'm personally still in a bit of a "wait and see" mode.
Even if we did want to do some kind of AI assistant, there are some difficult problems to deal with. Probably the biggest is where does the training data come from? Using customer projects is totally off the cards - there is no way we would ever do that. 500 or so example projects built in to Construct is a start, but while I'm not an AI expert by any means, I get the impression that is probably far too little to train an AI on. However much data we need to answer beginner questions, I suspect we'd also need 10-100x as much to answer intermediate questions, and 10-100x as much to answer expert questions. If it's even capable of answering expert questions - my experience of using AI for coding purposes as someone with ~20 years experience is they constantly come up short, so in the end I don't use them much. One recent example is I asked a question about WebGL 2 and ChatGPT fabricated a quote from the specification which did not really exist. It similarly makes things up too frequently for me to consider it a reliable resource. I've tried all the other ones and they all have similar problems. What if we add an AI assistant and it ends up giving bad advice frequently enough that it's not worth using? Perhaps even after all that work, for all its flaws ChatGPT still gets you a better answer.
Specifically on the point of documenting the file format, that is actually probably a difficult thing to do. At the moment we treat it as an internal detail and that gives us flexibility to change anything at any time. If it was documented and things depend on the specific details, then it makes it harder to change. Besides the way I thought AI works (again, not an expert) was it learned mostly from vast amounts of examples, and it doesn't necessarily need a specification. Nobody feeds AI models a specification of how English works, and yet they can use it fluently. AI models probably don't need access to the C++ specification to generate C++ code so long as they have enough training data. So it may be something that puts quite a big burden on us and yet does little to help AI tools.
What would really help AI tools? We already have hundreds of thousands of forum posts, 1000+ pages worth of documentation in the manual, and 1000+ user-submitted tutorials, which I think is a decent body of knowledge. If you really want to help AI understand Construct, I think probably the best thing is to contribute material about how Construct works in places AI will find it: post to the forum rather than using chat rooms, share your knowledge in tutorials, and upload interesting Construct examples to the web (GitHub is probably a good place for that). And best of all, that will all be useful for humans as well!
Other than that I think we could write a tutorial outlining an overview of how the Construct project file format works in broad terms, which might help LLMs understand projects, despite not being a full specification. We could also put a file like llm-context.md in every Construct project with a similar overview to help AI tools understand the Construct projects they do come across, rather than possibly seeing them as just a random collection of files. Perhaps that's not too controversial a thing to do?