Sound effects are a different problem from music, shorter, way more of them, and judged on a completely different standard: does it feel right at the exact moment it plays, not whether it's pleasant to listen to on its own.
Why SFX is actually the easier win
A game needs dozens to hundreds of short sounds, footsteps, hits, pickups, UI clicks, and almost none of them need to be memorable on their own, they need to be consistent and not annoying on the two-hundredth repetition. That's a much lower bar than a memorable theme, and it's exactly where current AI audio generation is strongest: short, functional, high-volume.
The repetition problem nobody mentions
A single AI-generated footstep sound, played every half-second for an hour of playtime, will reveal every flaw a human ear would've caught in an evaluation but missed in a demo. The fix isn't a better generation, it's generating three or four variations of each common sound and randomizing between them, the same trick game audio has used since long before AI, applied because AI generation makes producing variations nearly free.
Where it still falls short
Sounds that need to carry emotional weight at a specific story beat, a death, a big reveal, still benefit from a human sound designer's judgment about timing and restraint. And highly physical, tactile sounds, a very specific weapon reload, a particular engine note, are where generated audio still sounds slightly synthetic to a trained ear, even when a casual player wouldn't notice.
A workable split for a small team
Generate the high-volume, low-stakes library, footsteps, UI, ambient loops, impacts, with variation sets to avoid repetition fatigue. Keep a small budget, even a modest one, for the handful of sounds that carry real emotional weight. This mirrors the same split that works for AI voiceover and AI music: bulk coverage from generation, human craft on the moments that matter most.
The tell that gives away generated SFX
Not the sound quality itself anymore, it's the absence of variation. A game where the same exact footstep sample plays identically every time reads as generated, or at least as cheap, regardless of how good that one sample is. The fix costs nothing but a little extra generation, which is exactly why there's no excuse to skip it.