My five cents; there are many issues with AI as a DM'ing tool including but not limited to:

• All current LLM (large language/learning models) are either built directly on training data that is at best used without explicit consent or at worst stolen (in the IP rights sense), or are developed from models trained on questionable data sources. Unless you're writing your own model from scratch and then training it only on data you have the explicit right to use for that specific purpose, you're going to run into this ethical concern.
• LLMs are hugely inefficient and use not just massive amounts of power to run, but also massive amounts of water to cool the hardware. This has direct negative environmental concerns, from increased carbon emission for power generation to disruption of local ecosystems when that hot water gets dumped out into the local supply. So unless you're running your model locally on a closed loop cooling system and powering it entirely through renewable sources, you're gonna hit this environmental concern.
• Current LLMs are non-comprehending affirmation driven systems. What this means is the systems are biased to give a response in the affirmative to any permission request ("Can I X? Can you X?" etc) without understanding the actual 'truth' of what they're saying. This is because they function simply via probabilistic sentence deconstruction and assembly. They break your request down into tokens, use that to find other statements that probabilistically match what you've said, and return a response that probabilistically matches what others have said in response to their matched queries. In other words, if you ask the LLM if you can do something, there's no way of knowing if it's responding with any degree of correctness, giving a significant accuracy concern.

For those reasons above I won't engage with LLMs (aka AI) wherever possible, and this would extend to DM'ing. Setting the first two bullet points aside, a good DM knows when to say "no, you can't do that" because that maintains the coherency of the game and it's world. I wouldn't want to play with a system that doesn't "understand" that and just keeps giving incorrect or inaccurate responses due to it's affirmation bias.

Quote from GnollitAll78450 >>

Ok, here is a novel concept about computers running stories.

videogames! Pick up a text adventure. Try some of the old DND inspired games like Ultima, or Akalabeth. Or even... Baulder's gate 3. There are tons of other things to turn to without resorting to the overhyped liarbox that forgets whether i am playing a man or woman, elf or tiefling, and can't figure out that i am trying to cast a spell, when you need a fantasy fix.

lets stop with the " AI DMs" crud and do something actually good. 
Hells smells, take up writing!

 AI DM's will come. It is inevitable. Of course such things are an abomination. But that won't change the advent of them. 

Now, will they be "true AI"? No, not in the beginning. At the outset, they will simply be more sophisticated code than currently found in the best video games that interact with humans. You can bet everything you own that companies today are asking LLM's (which are not really AI, but the intermediate step) to build code to emulate as close as possible a human DM.

Technology, in the long run, is on a hyperbolic curve upwards. The real question is what comes first: The tech perfected for nuclear fusion plants to provide the electricity for the first true AI farms, or the tech for said true AI.

Quote from Pantagruel666 >>

Quote from 6thLyranGuard >>

 Generative AI is amazing in how profoundly wasteful and inefficient it is when it comes to resource requirements.  It's like what you'd expect from a Captain Planet cartoon.

 Nothing beats bitcoin mining for stupidly wasteful, though.

 NFTs!