My account was terminated after 5 years for ‘ manipulation of services’ no violations no warnings- trying to appeal and no response. Apparently Feb and March 2023 has another massive termination. Kdp used to be good… also all of my books followed Amazon TOS.
]]>Hi – thanks for your article – are you comfortable to share the contact details for the lawyers who sent through the letters
]]>But I was excited to make this notebook and my first publishing ever on KPD. I was awaiting approval.
Then one morning I wake up an email saying my account is terminated, because I have other accounts and this goes against terms and conditions.
I can be quite sure I do not have other accounts. I felt it was harsh…this was my first entry, I checked everything.
I have tried emailing and getting no where and they just repeat what was said before and not answer any of my questions.
The last email they said this is their final decision. I am, deeply upset. I feel penalised for something they are accusing of me of. That I have not done.
]]>Hi Ken – thanks for you thoughts. I completely agree that a lot of the decisions are driven by AI. Amazon is well known for using algos to do most of the heavy lifting on fraud detection, and they even mention it in some of their guideline pages. I was going to include a brief mention of that, and probably should have, but the article was already really long but I do plan on mentioning it a bit in part 2. Not to the deep extent you did above though, so I’m glad you posted that as it’s definitely informative!
I also agree that the algos are inevitable and the cost of doing business for Amazon, but no algo should be given the ability to make an account termination decision and to be completely honest, I have no idea if they do. That was another reason why I didn’t include that… I know the AIs are part of the decision but I don’t know if they make any final decisions at this level, and certainly hope they don’t. I’m pretty sure that they do at the customer level in terms of blocking reviews, as those are often wrong and almost always successfully appealed by customers willing to go through the hassle of getting someone to review their case.
Regardless of whether an account termination decision is done by AI or some random rep though, it doesnt’ change the fact that due to the seriousness of the consequences, they should always be reviewed by a much more senior and well informed human or committee before any decisions are made, and authors should be given the chance to appeal their case. I don’t think asking to pay for that is the right approach though.
I understand the argument that the appeals process could potentially be overwhelmed with bad actors appealing their case but quite honestly, I don’t think that would really be a huge issue. Most people that will ask for a human appeal would be doing so because they believe they are innocent. Most people knowingly doing something against the rules would have little to gain by asking Amazon to look into what they were doing even deeper than they already had, because all they’ll find is evidence of wrongdoing. And Amazon actually has a history of pursuing bad actors fairly aggressively, including suing them. All they would have to do to deter improper appeals is continue or step up that game. Ask them to look into your case when you were running some scam, you run the risk of them seeing just how guilty you are and suing you into oblivion.
I’ll talk more about most of this is part 2 next week though. Thanks again for your comments!
Craig
]]>My guess is that this is basically an example of opaque and poorly-designed machine learning algorithms at work. In many (though not all) senses, Amazon treats authors like any other sellers — and there are LOTS of sellers. So they automate the detection of violations. A machine learning algorithm computes an author’s overall probability of violation, and Amazon likely has some threshold for taking action (again, automated). The causes they list in the email (and in their policies) probably factor into their model (each via a numerical proxy of some sort) along with lots of other stuff. The model doesn’t come up with a “cause”, just an overall probability. Unless they use a linear model (which I very much doubt), they can’t even compute which factors contributed the most to the decision. This is one of the big problems with such models: the factors involved may be semantically meaningful but the result is opaque. There IS no specific cause that can be pointed to, and my guess is that this is why they don’t offer one. The rules aren’t vague because Amazon is being evasive, but because they ARE vague. In that case, Amazon just is describing some of the factors the model uses, not hard-coded “if-then” type rules.
Another reason they may choose to avoid publishing specifics is to avoid being gamed. If someone is indeed a bad actor, and is told that action X caused the ban, they could create a new account (or use another existing account, which they often have) and simply avoid X. The more insight they have into the operation of the fraud-detection algos, the less effective those algos become. It’s a cat and mouse game between Amazon — whose goal is to keep its customers, not authors or sellers, happy — and fraudulent sellers (of which there are many). You may think this doesn’t apply to authors, but Amazon does not distinguish the two. There are many sleazy third party sellers and a few sleazy authors (ex. publishing huge numbers of unedited manuscripts from project gutenburg under their own presses). Basically, authors are being punished by the algos for the prevalence of fraud in other markets.
Automated models are becoming ubiquitous in all social media (as they already were in CC fraud detection), and to some extent this is unavoidable. With hundreds of millions of participants, it simply is not viable to hire an army of human inspectors to police them. Even the appeal process would be inordinately expensive if a large fraction of designated-violators appealed. The bad actors certainly would appeal, again trying to game the system since it would be no skin off their back.
From my perspective, Amazon is at fault in two ways:
(1) Its automated fraud-detection algorithms are very poorly designed. Sadly, most ML algorithms are. There are legitimate ways to build ML algos (and determine their limitations), but few people have the understanding or inclination to do so. Most people just throw a bunch of factors into an opaque “deep-learning” algo (aka multi-layer neural net) and call it a day. My guess is that Amazon does this as well. Why? Because it’s cheap. Doing things right is hard and expensive. This said, when it comes to ML the results ALWAYS are opaque, and there is no “right” answer. There is a tradeoff between false negatives and false positives. False negatives piss off customers and false positives piss off sellers. Amazon only cares about customers. Moreover, the lowest error-rate a model can attain based on the available data may still be quite high. Even for a sensible model, there are statistical limits on what can be achieved. The “best” answer (based on available data, etc) may not be very good. One thing Amazon should do is have a human give a cursory glance to each case which warrants serious punishment (i.e. account closure, etc) to make sure it isn’t obviously wrong. A simple sanity check could eliminate 80% of false positives.
(2) The appeal process is a trade-off between accessibility and cost (and dissuading bad actors from pursuing frivolous appeals). It certainly should not be impossible to get reinstated without clout (i.e. a huge social media following or connections or a good lawyer) or be inordinately burdensome. A good way to structure it would be to have a hierarchical system, much like the US court system. At the lowest level, a quick review would be granted after some dissuasive hassle (to avoid being overwhelmed by fraudsters). I suspect that this would catch most issues. Then there could be a second level for cases where the author pushes back reaaaally hard (which fraudsters probably wouldn’t deem worthwhile). Finally, there could be some sort of legal arbitration as a last resort.
I know this may incur a lot of blowback, but instituting an appeal fee may help as well — especially if that fee funds a quicker and more transparent appeal process. This runs the risk of creating an adverse incentive (“gee, it would be a pity if this nice book business got broken, better pay us a protection fee”) and also would substitute money for social-media heft as an accessibility criterion. However, it could be argued that, since the primary damage in such cases is economic, money is an appropriate currency for this process. They could refund the appeal fees on successful appeal (which also runs the danger of creating an adverse incentive for them). Sadly, nobody has come up with a fullproof system for such things. But Amazon’s is pretty awful.
Sorry — bit of a long comment here. Anyway, just some thoughts on how we got here and why it’s not likely to improve any time soon.
]]>You’re right, Dale – and that’s one of the biggest issues that can and should change. There is no reliable point of contact that authors can use to really get heard. Contacting front line support goes nowhere. When I post part 2 of this next week, that’s one of the major changes I will be suggesting.
In terms of how to lobby Amazon at all, I think the only way to do that is through enough authors understanding the issue and making noise about it. That’s why I posted this. My hope is that it gets shared around, or sparks some conversations elsewhere, enough so that authors start to make some noise about this on social media or wherever – enough noise that Amazon can’t continue to completely ignore it.
]]>If there were a real person, or even a committee, to discuss these things with a lot of authors may be interested in helping get these changes in place. But as it currently stands, we’re without a voice.
Having a plan or at least a few ideas on how to lobby for changes at Amazon is likely the first step in this process… and I don’t have a clue as to who to contact to get this ball rolling. Does anyone else know where to begin?
]]>