I cannot for the life of me understand why Amazon doesn't already do this.
Ultimately, Amazon's goal is to maintain a strong brand and keep stockholder confidence. This presents a much stronger move without admitting as much weakness, compared to allowing the behavior to continue and trying to mitigate it with AI (since the layperson -- ie most stockholders -- has a poor understanding of the efficacy of ML techniques).
Additionally even if if they did, they are less incentivized to publicize it. Secretly diminishing the voting power of particular reviewers works just as well, if not better, if they don't know it's happening at all, and prevents creation of multiple accounts to avoid said penalties.
Unless they are imaging every item sold I don't see how that helps. Unscrupulous seller uploads high quality images of the real item, and delivers counterfeits.
Conterfeits are the biggest reason I don't buy more on Amazon. I can't say I've ever received a counterfeit item, but the stories are so widespread that I always have that fear.
I'm actually aware of some large FBA players that intend to create thousands of new 'brands' in the next year in order to push their wares.
Even still, they'll get their reviews, policy or no policy.
Fake reviews are fairly easy to spot because of the one star curve. Amazon can weight and sort search results accordingly, if they so chose. Amazon can add additional weights based on the user's history and using graph network analysis of that user's review distribution. Bad items don't necessarily need to be banned, but the negative weight can make the listings virtually invisible.
On the other hand, if the reviews are for a product other than what you are ordering, things get a lot harder. It is hard for a user to eyeball it, and perhaps equally as hard for Amazon to verify it at scale.
Prediction - AI plus high resolution imaging of the good at high speed.