Jukebox AI is what a lot of people use these days, and as far as I know it's hit or miss with vocals. Sometimes it can sound convincing but super low quality like someone recorded the audio with a toaster. Other times it can sound like the souls of the damned screaming from the depths of hell. But neither times does it sound very coherent. It's a fun tool to use with trying to expand pre-existing snippets but I don't think it's very likely that this snippet was made whole-cloth from an AI. There's a few videos of MJ songs being continued by AI (there's even a Jukebox AI-generated continuation of this same snippet, found here:
https://www.youtube.com/watch?v=RY_k0UjcdTE) but I don't think the technology's there yet, especially in trying to simulate MJ. Will sure be scary when it is, though.
Also @dam2040 I made the video earlier today but never got around to sharing it. The acapella generator was a bit rough with the original audio, and so the quieter parts of it got filtered out. I still think, though, that there's enough clear vocals for anyone to analyze and make up their own minds. I'm no audiophile or sound engineer, but to me the acapella doesn't show any signs of snippet-stitching or sampling or any Cascio-like shenaniganery. This leads me to think that it's real, but anyone else can give it a listen. Here's the link (it includes the original snippet and then the acapella after):