This is actually rather easy to solve without ML/AI. They need to track the change delta (ex: with jsondiffpatch) and weight each part, presumably using levenshtein or text mining with weighted dictionary, and limit changes to 2 or 4 in 24 hours period. Then weekly compare those changes and treat the highest one as need to be reviewed by customer.

Additionally, each customer review should tied to specific change version, that they can easily compare with current one.

Easier, no editing. Create a new listing for any changes. Allow create listing from.. or linking to old listing for repprting. Show new version available text instead of buy again. Solved.

I'm going to be kind to them and assume they are using the AI to augment the obviously simple "rating delta over time" I'd do as the first pass solution.

Like "Over a 3 month period this product reviewed at 3.7 and the next 2.1" maybe look into why.

During a recent Amazon all-hands meeting, the company leadership talked about how there are something like 5,000 employees working on various marketplace frauds, including using ML on some of this.

I have no idea why it's so hard, but I don't think Amazon is being as complacent as it seems from the outside.

Frauds where Amazon loses money, or frauds where Amazons customers do not get the product of its purported reputation, or maybe not even the product they thought they ordered.

They were specifically talking about marketplace fraud, where Amazon customers are misled but no money is lost. Customer fraud (against Amazon) is a different thing, though there's a bit of overlap when it comes to fraudulent returns.