These countless incidents of lawless action are not indirect or tangential effects of social network algorithms. Rather, they are the exact effects the algorithms are designed—and intended—to have. In 2014, Facebook learned how it could affect and alter its users’ emotional states. See Adam D. I. Kramer, Jamie E. Guillory & Jeffrey T. Hancock, “Experimental evidence of massive-scale emotional contagion through social networks,” Proceedings of the National Academy of Sciences of the United States of America (June 17, 2014); Gregory S. McNeal, “Facebook Manipulated User News Feeds To Create Emotional Responses,” Forbes, June 28, 2014. In a massive study on many Facebook users, cited above, researchers Kramer, Guillory, and Hancock concluded that they could precisely manipulate users’ emotions by merely altering the content users were exposed to on their newsfeeds. The study explained this phenomenon—dubbed emotional contagion—as follows:
Emotional states can be transferred to others via emotional contagion, leading people to experience the same emotions without their awareness. Emotional contagion is well established in laboratory experiments, with people transferring positive and negative emotions to others. Data from a large real-world social network, collected over a 20-y period suggests that longer-lasting moods (e.g., depression, happiness) can be transferred through networks, although the results are controversial. In an experiment with people who use Facebook, we test whether emotional contagion occurs outside of in-person interaction between individuals by reducing the amount of emotional content in the News Feed. When positive expressions were reduced, people produced fewer positive posts and more negative posts; when negative expressions were reduced, the opposite pattern occurred. These results indicate that emotions expressed by others on Facebook influence our own emotions, constituting experimental evidence for massive-scale contagion via social networks.
Notably, the study expressly discussed the applicability of emotional contagion to newsfeed content generated by social network algorithms.
Because people’s friends frequently produce much more content than one person can view, the News Feed filters posts, stories, and activities undertaken by friends. News Feed is the primary manner by which people see content that friends share. Which content is shown or omitted in the News Feed is determined via a ranking algorithm that Facebook continually develops and tests in the interest of showing viewers the content they will find most relevant and engaging. One such test is reported in this study: A test of whether posts with emotional content are more engaging.
That same year, the Cambridge Analytica scandal revealed that users’ physical actions, not just their emotional states, could also be readily manipulated based on the curated content shown to them. See Carole Cadwalladr & Emma Graham-Harrison, “Revealed: 50 million Facebook profiles harvested for Cambridge Analytica in major data breach,” Guardian, Mar. 17, 2018.
Cambridge Analytica spent nearly $1m on data collection, which yielded more than 50 million individual profiles that could be matched to electoral rolls. It then used the test results and Facebook data to build an algorithm that could analyse individual Facebook profiles and determine personality traits linked to voting behaviour. The algorithm and database together made a powerful political tool. It allowed a campaign to identify possible swing voters and craft messages more likely to resonate.
In other words, social media networks could predict, with a substantial degree of certainty, which content would emotionally stimulate a user to perform a specific act physically. In the case of the Cambridge Analytica scandal, the specific act or output the company was allegedly seeking to elicit was a vote for Donald Trump. “The Cambridge Analytica Story, Explained,” Wired. Whether former President Trump’s 2016 victory can actually be attributed to the company’s targeted digital marketing may perhaps never be known, but what is known is that this type of tailored content is proven to influence a user’s feelings and actions. A 2019 study conducted by the Pew Research Center found that, when directed to their ad preferences page, a majority of Facebook users, namely, 59 percent, said the categories accurately reflected their real-life interests. Thus, scientific studies and the very users themselves have confirmed the accuracy of social network algorithms. Paul Hitlin & Lee Rainie, Facebook Algorithms and Personal Data (Pew Research Ctr. Jan. 16, 2019).
Many social media networks have had no scruples with these technological tools and capabilities. For example, in 2016, Facebook publicly introduced FBLearner Flow, a prediction engine powered by artificial intelligence and designed to deliver “a more personalized experience for people using Facebook.” Jeffrey Dunn, “Introducing FBLearner Flow: Facebook’s AI backbone,” Facebook: Engineering, May 9, 2016. In a 2016 post, Facebook touted that “[O]ur prediction service has grown to make more than 6 million predictions per second.” Id. Confidential documents leaked two years later revealed that Facebook launched a new advertising service that offered advertisers the ability to target users based on predictions generated by FBLearner Flow of users’ future behavior. By feeding in the collection of personal data points that Facebook has on each user—which sources indicate may range anywhere from 98 to 52,000 data points and may include a user’s location, device information, Wi-Fi network details, video usage, affinities, and details of friendships—FBLearner Flow predicts how the user will behave in the future, e.g., what the user will purchase. Sam Biddle, “Facebook Uses Artificial Intelligence To Predict Your Future Actions For Advertisers, Says Confidential Document,” Intercept, Apr. 13, 2018; Caitlin Dewey, “98 personal data points that Facebook uses to target ads to you,” Wash. Post, Aug. 19, 2016; Adam Green, “Facebook’s 52,000 data points on each person reveal something shocking about its future,” Kim Komando, Sept. 17, 2018.
One slide in the document touts Facebook’s ability to “predict future behavior,” allowing companies to target people on the basis of decisions they haven’t even made yet. This would, potentially, give third parties the opportunity to alter a consumer’s anticipated course. Here, Facebook explains how it can comb through its entire user base of over 2 billion individuals and produce millions of people who are “at risk” of jumping ship from one brand to a competitor. These individuals could then be targeted aggressively with advertising that could pre-empt and change their decision entirely—something Facebook calls “improved marketing efficiency.”
Biddle, supra.
Is First Amendment Protection Applicable?
What implications does this have for the categorization of social media networks’ algorithmic speech? In short, it potentially strips such speech of the heightened First Amendment protection it would enjoy were it to be classified as commercial speech. To qualify as commercial speech, a social network algorithm must concern lawful activity, which the direct incitement of imminent lawless action can never be. If a social media network knows how to influence any user’s behavior effectively—indeed, if it has sponsored scientific studies and conducted countless experiments and real-life case studies to confirm this knowledge definitively—and a clear correlation can be demonstrated between the content generated and shown to that user and the user’s resulting lawless conduct. It follows that the content was very much directed to inciting the imminent lawless action that followed. The social media network knows that such lawless action will imminently follow and shows the content to the user in spite of this knowledge. Indeed, Justice Thomas’s concurrence in the recent case of Malwarebytes, Inc. v. Enigma Software Group USA, LLC, 208 L. Ed. 2d 197 (Oct. 13, 2020), suggests that the scope of the immunity provision in section 230(c)(1) of the Communications Decency Act (CDA) was never meant to insulate interactive computer services like Facebook from liability where they distribute or circulate content that they know is illegal, e.g., where a user has reported or flagged the content as defamatory. Justice Thomas’s concurrence effectively implies that the current judicial interpretation of the CDA’s immunity provision is misplaced and that this legal issue is ripe for adjudication by the Supreme Court.
Each post that any given user encounters on his or her newsfeed is specifically chosen for that specific user. One’s newsfeed consists of meticulously curated content based on hundreds, if not thousands, of data points that the social network algorithm has been given relating to that specific user. Based on these data points, a prioritized list of content uniquely tailored to that user is generated. If it must be anything at all, newsfeed content is pointedly directed at a user. That is, after all, the very appeal of the newsfeed feature itself.
This practice’s egregiousness is particularly acute where such content is shown to users whom the algorithms have identified or flagged as current or potential sympathizers of the violent content or hate speech being circulated. In what happens to be a striking comparison, Justice Holmes aptly articulated the reason for this egregiousness in Frohwerk v. United States:
[I]t is impossible to say that it might not have been found that the circulation of the paper was in quarters where a little breath would be enough to kindle a flame and that the fact was known and relied upon by those who sent the paper out. Small compensation would not exonerate the defendant if it were found that he expected the result, even if pay were his chief desire.
249 U.S. 204, 209 (1919).
In light of the foregoing, social media networks’ algorithmic speech would likely be subject to a rational basis review that nearly any legislation would withstand given the significant public policy considerations at play, e.g., data privacy, consumer protection, and public safety. The specific language of such legislation would heavily depend on the nature of the algorithmic technologies being used in the industry. For example, legislators must take caution to avoid the pitfall of drafting overly broad or all-encompassing language that would unwittingly restrict or prohibit the use of algorithms that do not tend to incite lawlessness. This goal may be more easily accomplished given recent technological developments such as Facebook’s 2019 announcement that it removed more than seven million instances of hate speech because artificial intelligence—rather than human moderators—was now detecting hate speech on the platform. Billy Perrigo, “Facebook Says It’s Removing More Hate Speech Than Ever Before. But There’s a Catch,” Time, Nov. 27, 2019. Despite drawbacks with this seemingly remedial measure, this is nevertheless a promising start that may help lawmakers better tailor applicable legislative language to curb the violent real-world consequences brought about by social network algorithms.