Manoel Horta Ribeiro (@manoelribeiro),
This post is intended to address some of the comments and criticism w.r.t. my latest paper: Auditing Radicalization Pathways on YouTube. The paper got an unreasonable amount of attention, which involved media coverage, angry e-mails and upset YouTubers. Individuals associated with the communities in the paper were critical of aspects of the methodology. I believe most of these critics:
Misunderstood methodological aspects of the paper;
Don’t understand how doing research with large datasets work;
Criticize things that the paper does not say;
Largely ignored sections of the paper methodological caveats were addressed.
Regardless, it seems plausible that my paper will play some role in shaping the debate around YouTube, and thus it is important to address some of the critics. I will do that generically, addressing points that I have seen being raised rather than responding to individuals.
## What The Paper Does Not Say
First and foremost, it is important to stress that the paper inspected two different directions:
Is there evidence for user migration from the I.D.W./Alt-lite to the Alt-right?
Is the recommender system responsible for this?
We find very strong evidence for (1), and don’t find significant evidence to answer (2). Overall, the paper says that migration from these communities did happen, but does not really answer why this happen. There are several reasons why this could be, such as: (i) the recommender system, (ii) something related to the demographics that these audiences target, (iii) something about the messages shared by the I.D.W./Alt-lite.
Secondly, the paper uses aggregated data, and thus it is tricky to say (insert member of the I.D.W. here) is radicalizing people. This phrasing assumes that we found a causal relationship between a specific type of content and radicalization, and is much stronger than the statement “individuals who commented on I.D.W. and Alt-lite channels consistently migrated towards Alt-right videos”.
## The Controls
A funny story here concerns the channels we called controls. People are very used to traditional controls in contexts such as randomized blind trials. There, we can’t simply