While the “Facebook Files” showed the many limits of the social network moderation tools, the latter publishes new numbers.
Le Monde
Some 1.8 billion false accounts deleted in three months, just like 777.2 million spam or 22.3 million hate messages … Facebook (now called “Meta”) presented on Tuesday, November 9, its report quarterly, for the period from July to September, on its moderation practices.
On Facebook As on Instagram, the company says, supporting numbers, having not only improved its ability to suppress the contents forbidden on its platforms (nudity, apology of terrorism, harassment …), but Also have progressed in their automatic detection. According to the data published by the company, almost all messages for suicide or violence, such as almost all of the prohibited content on the social network, are now automatically identified, without reporting on the part of Users.
But the main tool used by Facebook to measure the effectiveness of its moderation services is particularly criticized for several weeks, especially through the publication of “Facebook Files”, these hundreds of documents copied by the ex-employee of the Social Network Frances Haugen, to which Le Monde and several other editors have had access.
Prevalence and language problems
The main figure facing Facebook is the “prevalence” of bad content on its platforms, namely the frequency at which a user is confronted with a naked photo or a hateful message. Depending on the figures of the company, the prevalence of hate messages on Facebook has thus increased to 0.03%: on average, only three messages on 10,000 seen by users are hateful, a figure divided by three in one year. “We believe that the prevalence is the best step to assess our progress,” said Guy Rosen, Vice President of Meta, in charge of integrity (moderation, protection of users …), Tuesday, at a press conference .
But these figures are given only globally and are an average. They are not representative of the course of all users. Neither of all countries. As the “Facebook Files” show, in many places in the world, Facebook’s automatic detection tools do not work, or badly, and human moderators are very few. This is particularly the case for most dialects of Arabic. In many Arabic-speaking countries, the moderation of the social network is deficient, and therefore the measurement of the prevalence of prohibited content.
“The figures we publish are global and are developed from samples in several countries and in several languages, defended Mr. Rosen. We perform manual verification tests, and we do our best to understand the Prevalence of certain types of content also in countries who face specific risks related to news. “Why, in this case, do not publish country data by country and language by language? “This is something we may consider in the future,” Rosen said.
In its report, Meta made new data public, including figures on the messages deleted for harassment, with a prevalence evaluated between 0.14% and 0.15% on Facebook, and between 0.05% and 0 , 06% on Instagram. Harassment is one of the topics on which automatic detection systems remain less efficient, due to the difficulties they still have to understand the necessary context to find out if a message can be considered as such.