#amazonfail - an architechtural vulnerability?
Francis Turner
April 13th, 2009
Much of the world has been twittering and blogging over the weekend about how Amazon appears to have been removing certain works from its sales rank algorithms. The works are removed because they are (allegedly) “adult” but in fact the real issue common feature is that these books seem to be about homosexuality rather than other “adult” issues.
There is, of course, much outrage - particularly because Amazon’s “glitch” excuse simply fails to meet the usual definition of a glitch. This problem seems to have started in February if not last year and to have gradually ramped up over the last couple of days. All in all this doesn’t sound like a “glitch” so much as a major software failure that has been exploited by humans.
I’m not alone in this basic viewpont, it seems to be shared by Information Week’s Mitch Wagner amongst others. Indeed there are some livejournal posts that indicate that the explot is the work of either a single hacker or a small group.
However the issue is not so much whether it is a hack or a concerted group so muchas to shine a light onto a what may be an architectural weakness of Amazon’s site. Amazon basically relies on site visitors to provide tags and to flag things as objectionable. It seems likely that for the most part these flags and tags pass through the Amazon system with no human supervision. This is probably partly because there are simply too many for a single human to verify and partly because for legal liabilty reasons Amazon found it better to not have someone make a corporate decision on the matter of “objectionable material”.
If the problem is indeed that people are gaming the system then fixing it while retaining the benefits of the “wisdom of crowds” approach that helps Amazon scale so well is going to be hard to fix - indeed it may be impossible. I’m sure that Amazon programmers are busy rolling back the auto-objectionable code and that someone is about to permit “objectionable” material back into the Amazon ranking systems but that is a short term fix and doesn’t solve the problem that Amazon faces regarding people not wishing to see “objectionable” content.
This problem is partly caused by the fact that different people have very different ideas about what is “objectionable” so no large group of people will ever agree. However it is exacerbated by the fact that Amazon attempts to present a single storeface to all. The situation then gets huge visibility because Amazon is such a major reseller that books often depend on Amazon ranking and Amazon visibility for success. Thus removing certain works from the ranking scheme is bound to get those associated with the works in question very worked up because they can see their revenue disappearing into a rankless desert.
Tags: amazon amazonfail
April 13th, 2009 at 6:51 pm
It’s worth mentioning that Dear Author has established that the books removed were removed based on publisher-supplied metadata, not on user-supplied tags. The glitches in the removals (e.g. hardcover vs. paperback editions inconsistently treated) are exactly predicted by metadata glitches.
http://dearauthor.com/wordpress/2009/04/12/amazon-possibly-using-category-metadata-to-filter-rankings/#more-11485
April 13th, 2009 at 8:35 pm
Actually, I figure they will probably implement some form of “safe search” ala Google and just apply the de-ranking to those searches that are “safe searches”.
April 13th, 2009 at 9:14 pm
Thanks — this is very helpful and makes a good deal of sense. But I wonder, spurred by a (possibly spurious) post I saw that passed on an alleged mention from Amazon that they were creating a “Christian Bookstore” — that is, a restricted sandbox for a certain kind of user. Of course, they could just run such a bookstore under a different URL, but what if they are doing it within the ‘Amazon’ larger storefront? It makes sense, and is just an extension of what they already do in terms of suggestions, etc., to block or make harder to find books that some algorithm calculates are likely to be ‘objectionable’ to a given reader. This need not apply only to Christian/social conservative readers, but could include other categories. The combination of a ‘content’ characterization and a purchase/browse record could presumably, with some sophistication, increase not only positive targeting (”Readers like you have also bought…”) but also negative targeting (unless you try really hard, we’re not even going to show you.) But it would be easy for such a system to go awry, and overexclude, or to put users into exclusion categories they didn’t belong in — especially if the ‘content filter’ was largely automated (any book with the following words/phrases in its title/blurb).
Perhaps this is all to Machiavellian, though, and Occam’s razor suggests that a gaming/hacking of Amazon’s ‘objection’ machinery may be sufficient explanation for the ‘glitch.’
May 26th, 2009 at 11:50 am
[...] crucial, conversations by ignoring others - Amazon, for example, recently was slow to pick up on a storm of criticism created by its policy on homosexual themes because it didn’t take Twitter seriously [...]