If you take one rule from this piece, take this one: never cite a source you have not opened. A model can suggest where to look. It cannot vouch for what it finds, and you cannot afford to either.
The verification pass
For every AI-surfaced reference, confirm three things before it earns a place in your bibliography. That the work exists, in a database you trust. That the authors and year match. And that the source actually supports the specific claim you've attached to it — not a similar one, the exact one.
Most hallucinated citations fail the first test instantly: there is no record of them anywhere. The dangerous ones pass the first two and fail the third — a real paper, correctly attributed, that simply does not say what you need it to say. Those are the ones that survive a careless read and detonate in peer review.
Build the verification pass into your process as its own step, with its own time on the calendar. Writing and checking are different jobs. Treat them that way and the retraction risk largely disappears.