An analysis of Facebook’s success at enforcing its content rules by the social media giant’s employees has shown how the company only removes a fraction of posts that violate its hate speech rules, The Wall Street Journal reported on Sunday, citing internal documents reviewed by it. One of the reasons behind Facebook’s inability to regulate such content on its platform is the lack of training its AI has on foreign languages, the report said, adding that in March this year, the company’s employees who were gearing up for regional elections in India observed that hate speech was a major risk in Assam, where there is growing violence against Muslims and other ethnic groups.
“Assam is of particular concern because we do not have an Assamese hate-speech classifier,” according to an internal Facebook planning document. The legislative Assembly election in Assam was held from March 27 to April 6 and saw the incumbent BJP-led National Democratic Alliance retain power.
Earlier this month, The Indian Express had reported, citing from the complaint filed with the US securities regulator by whistleblower and former Facebook employee Frances Haugen, that the company was aware of posts being published on its platform that contained misinformation or hate speech but it could not take action or flag this content because of the “lack of Hindi and Bengali classifiers”. Classifiers refer to Facebook’s hate-speech detection algorithms. According to Facebook, it added hate speech classifiers in Hindi starting in early 2020 and introduced Bengali later that year. Classifiers for violence and incitement in Hindi and Bengali first came online in early 2021.
An e-mail query sent to Facebook did not elicit a response.
The WSJ report, citing employees and internal company documents, has also pointed out examples of content that Facebook’s artificial intelligence engines should have detected but missed. It said that Facebook’s AI cannot consistently identify “first-person shooting videos, racist rants and even, in one notable episode that puzzled internal researchers for weeks, the difference between cockfighting and car crashes”.
While the report claims that Facebook only removes “a low-single-digit percent” of hate speech content, the company’s spokesperson in its response to the newspaper has said that these percentages referred to posts that were removed using AI, and didn’t include other actions the company takes to reduce how many people view hate speech, including ranking posts lower in news feeds.
For India specifically, Facebook has said earlier that its human reviewers review content in 20 Indian languages — Hindi, Bengali, Tamil, Malayalam, Punjabi, Urdu, Kannada, Marathi, Gujarati, Assamese, Telugu, Oriya, Sindhi, Mizo, Marwari, Chhattisgarhi, Tulu, Maithili/Bhojpuri, Konkani and Meitei.
“Assam is of particular concern because we do not have an Assamese hate-speech classifier,” according to an internal Facebook planning document.
This story first appeared on indianexpress.com
