While the European Commission officially asked Amazonwithin the framework of the European Regulation on Digital Services (DSA), to provide more information on the means put in place to protect its users, researchers from the non-profit organization AI Forensics and the company Check First looked at the recommendation algorithm of the American platform’s bookstore space.
In a report published Monday December 11the two entities highlight the fact that Amazon’s recommendation and search system “not only promotes misleading books on health, immigration, climate change and gender issues, but also traps users in such narratives”. The authors fear that “with 181 million Amazon users in the European Union alone, these risks [puissent] have disproportionate impacts”.
The investigators – who focused on the French and Belgian versions of the Amazon bookstore – took as their starting point a list of 2,000 bestsellers addressing highly debated social themes, such as health, ecology or ‘immigration. Over several weeks in October and November 2023, and using profiles with clean history to analyze how the algorithm behaved by default, they then looked at which other books Amazon was directing Internet users to from the product sheet, like the “Products frequently purchased together” section. In total, more than 60,000 book recommendations were analyzed.
The book, a legitimate source of information
The researchers find that “for 71.7% of Amazon France search results with the term “Covid”, the associated search queries contained books by authors known to spread misinformation”. On the same subject, “80% of the books in the top 10 question the existence of the pandemic, minimizing its effects on health or presenting it as a conspiracy. The proportion is even more alarming on the [mot-clé] “vaccine” since 90% of the results offer stories anti-vaccination » , details the report. On the subject of immigration, this time, researchers counted only 30% of the top 10 works on this theme “propagate negative views on immigration”.
Joined by The worldthe AI Forensics team believes these results are all the more worrying as “books represent a legitimate source of information in public opinion” whose contents would supposedly have more credibility than a publication on a social network. Furthermore, according to Buse Çetin of AI Forensics, these results would undoubtedly have been “deemed unacceptable if they had been concluded about other platforms”.
Failing moderation
For the director of AI Forensics, Marc Faddoul, as for the researcher Paul Bouchaud who worked on the investigation, there is no doubt that Amazon is much more than a simple and neutral Marketplace. “ Amazon’s algorithms shape not only sales but also public discourse, particularly through book recommendations. Our results highlight Amazon’s failure to put in place basic mechanisms to mitigate systemic risks, showing the limits of AI self-regulation., argues Mr. Faddoul. For the authors of the report, the results of their investigation show that Amazon is not in the DSA’s spotlight: “Under Section 34 of the DSA, platforms like Amazon are required to identify, analyze, and assess systemic risks, taking into account factors such as algorithmic systems and content moderation”they specify.
Worse, Amazon would, according to them, contravene its own terms and conditions. Pointing out a faulty moderation, the investigators noted that works with “sexually explicit content” (notably hentai manga of a pornographic nature) could appear in the categories of adolescent books or in general and general public categories such as “family and well-being”.
Contacted by Le Monde, the American company wants to be reassuring: “Our investigations are ongoing and we will continue to take appropriate action against products that do not comply with our policies. » Amazon specifies that it “shares with the European Commission the ambition to create a secure digital environment” and that she “invests significantly to protect [sa] store from bad actors and illegal content, and to create a trusted shopping experience.” The company further maintains that its shopping and discovery tools are not designed to generate results biased towards a specific point of view.