Finding an accessibility issue is only half the job. It doesn't even matter if it's you who found it, an automated tool or AI.
The fix can still be wrong.
This is especially true for AI suggestions. I'm sure AI can correctly spot that a button has no accessible name. But then, and I've seen this more than enough times to think it's not hallucinating, it will suggest a fix that looks tidy in code but fails in the browser and breaks translation.
That's not what you want, obviously. And review is important of course.
But reviews aside, how can you tell if an AI suggestion is useful?
Here's how I do it.
A useful AI fix should be specific. It needs to point to the broken element, explain the user impact and describe what needs to change.
It should fit your product. If you're using React, Angular, a vanilla implementation, already have a design system button or the fix is in your CMS, the AI should tailor the suggestion to that.
It should be testable. By you. You need to know how to confirm the fix worked. Can you tab to the control, check the accessible name, submit the form with errors and listen with a screen reader for the correct announcements?
And, my favourite, it should avoid lazy advice. "Use semantic HTML" may be true, but it is not enough to ship anything.
If the AI's recommendation creates more rework, then it's not worth it.