I think the tells of AI video are becoming more subtle, similar to language. It's no longer so much that the visuals are categorically impossible, such as mangled hands or impossible geometric arrangements of objects, but you can still see style and composition that is more frequent in AI video (but could be possible in a real video as well in principle).

Such as higher production quality, too beautiful people, a kind of stock photo sheen, etc. Of course if you use special LoRAs or prompts and input images, it's possible to leave the stock footage style, but most people don't bother with it, just like most people use stock ChatGPT in its default voice with its favorite trope-filled cadence etc.