[liberationtech] YouTube doesn't lead to radicalization, study finds

John Ohno john.ohno at gmail.com
Thu Jan 2 17:57:26 CET 2020


Google sets (and checks) a perma-cookie for customizing youtube
recommendations & doing ad targeting -- so even without an account, a bot
following recommendations will have a customized element to rankings based
on history (unless that cookie is manually deleted). I recall a paper a few
years ago doing exactly the same thing (following recommendations with no
account) & finding that, from a marginally political starting point, all
roads lead to nazi videos, so I suspect that these results are probably
because of changes to the ranking algorithm designed specifically to
discourage fringe content (and encourage centrist content). (I don't recall
the name of that paper but I'm sure anybody paying attention to the subject
in 2016 and 2017 will recall seeing it.)

This kind of analysis is tough, not just because of the black box nature of
a constantly-tweaked ranking algorithm & the path dependency of a random
walk through recommendations (whose actual numeric rankings are not
visible) but because we know from experience that factors like location,
ISP, and browser type are taken into account, & spoofing these in a
controlled way without hitting anti-spoof measures complicates an
already-difficult analysis.

Regardless of any kind of tweaking Google might do, automatic
recommendation systems have an underlying problem (which all of them share,
aside from the rare experiment / art project): items are ranked and
recommended based on either existing popularity, hotness (which depends
upon popularity as a factor), or interaction count (which depends upon
popularity as a factor) -- which is to say, no matter the other factors
taken into account, recommendation systems drive people into groupthink by
nudging them in the direction of consuming the same things as other people.
As far as I'm aware, the only countermeasure for this (on a personal scale)
is the use of third-party randomizers (which select options totally
randomly, and sometimes also support filtering out anything with greater
than some maximum number of views). There's a business case against
eliminating the 'trendist' model of promoting already-popular things:
Google depends upon the handful of 'winners' on their platform as the
platform's face, & as fewer and fewer of them can make a living off
Youtube, the ones that remain make more & are more closely tied to the
platform. But, for subscription services, evening the playing field makes
more sense: Netflix makes the same amount of money regardless of how much
people watch (and, in fact, makes marginally more money if people watch
less but maintain their accounts), and so to the extent that they pay out
royalties based on viewership, they have no particular reason not to spread
out views more evenly through a more nuanced and granular recommendation
model.

On Thu, Jan 2, 2020 at 11:30 AM Niels Abildgaard <niels.abildgaard at gmail.com>
wrote:

> Freedom abhors accounts.
>>
>> Here are some accounts...
>>
>
> I have no idea what this is supposed to mean, or how it is relevant to the
> study. The study, concretely, didn't study how recommendations for a
> YouTube account evolved over time, as they were not logged in when looking
> at recommendations. Freedom might abhor accounts, and if you don't use an
> account on YouTube... good on you, I guess? But that's not what the
> discussions about radicalization on YouTube have been about. The study
> doesn't apply, is the point.
>
> Corinne's link has more good details :-)
>
> Den tir. 31. dec. 2019 kl. 20.47 skrev grarpamp <grarpamp at gmail.com>:
>
>> > an account
>>
>> Freedom abhors accounts.
>>
>>
>> Here are some accounts...
>>
>> https://en.wikipedia.org/wiki/IBM_and_the_Holocaust
>> https://ibmandtheholocaust.com/
>> infohash:20820F55D884C945154136689E436990107DD1E9
>>
>> --
>> Liberationtech is public & archives are searchable from any major
>> commercial search engine. Violations of list guidelines will get you
>> moderated: https://lists.ghserv.net/mailman/listinfo/lt. Unsubscribe,
>> change to digest mode, or change password by emailing
>> lt-owner at lists.liberationtech.org.
>
> --
> Liberationtech is public & archives are searchable from any major
> commercial search engine. Violations of list guidelines will get you
> moderated: https://lists.ghserv.net/mailman/listinfo/lt. Unsubscribe,
> change to digest mode, or change password by emailing
> lt-owner at lists.liberationtech.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ghserv.net/pipermail/lt/attachments/20200102/49f562b4/attachment.html>


More information about the LT mailing list