Daniel Sude, Ph.D.

A useful resource for building up intuitions about machine learning (in this case, supervised machine learning) https://healthcare.ai/visual-tour-lasso-random-forest/….

People in my field will try to categorize social media posts as civil {uncivil} or low {high} in deliberative quality using these techniques. The goal, there, is to train machines that can, with a high but not perfect degree of accuracy, sort through hundreds of thousands of posts. This can help not only generate a pool of ecologically valid stimuli for use in experiments but help peole to understand emergent crises.

Organizations like http://moonshotcve.com/ (CVE = countering violent extremism) will identify individuals who are being radicalized using similar techniques. They can then try introducing “alternative content” (content that challenges extremist narrative). According to one of their representative’s talk at New America’s 2020 Future Security Forum, these potential extremists do click on that alternative content. Note: These “low effort” tech solutions are not perfect. Some people are flagged for more effortful counter and deradicalization efforts.

These efforts succeed in part with the coopertion of social media companies. They can also be pursued independently (for good or for ill. Imagine extremists using machine learning to identify good targets for their messages).

However, my own view is that introducing new content is not enough. Content may serve as an exemplar – impacting perceptions of public opinion (within groups) as well as of normative behaviors. It may also, of course, serve as a source of {mis}{dis}information. Each of these functions can have a downside that is hard for the individual to correct.

Perceptions of public opinion and norms based on exemplars may in fact be quite inaccurate (as when participants in Sude et al. 2019 shifted their perceptions of public opinion after reading a single article). Exemplification is largely an automatic process, people are not necessarily aware of it and it can persist even in the presence of base-rate information (e.g. https://doi.org/10.1177/009365094021005003). When people “like” the revised perception of public opinion – e.g. when it implies that their opinions are popular – this can also lead to a motivation to embrace and defend these revised estimates, even in the face of counter evidence.

A social media company, aware of this impact, could consciously gather evidence about the “true” baserates (on their site, and, based on polling data, nationaly). The least it could do is provide base-rates that partially counter the exemplification effect. Alternatively, they could prompt users to “anchor” these base-rates on specific groups, both by presenting a finer grained profile of users that resemble that exemplar (individuating them) and by punctuating the social media experience with survey questions like: “What percentage of Republicans in your social network on Twitter do you think share this poster’s views?” accompanied by “What percentage of Republicans nationally do you think share this poster’s views?”: The goal would be to highlight the potential differences between the “answers” to these questions. It’s more likely that the exemplar provides meaningful information about your social network than about the national public.

Similarly, with regards to asking people to evaluate source and content quality before endorsing or sharing, putting the onus on the individual to be an “A student” is impractical. Affordances that facilitate external citations and links, and thus a stronger web of evidence, shift the burden away from the individual and towards a social media company which is better able to handle it. Importantly, the social media company could also categorize different types of sources, provide a transparent justification for why “mainstream” sources are more likely to be accurate (e.g. the legal and institutional processes that ensure higher accuracy), and then provide a more objective rubric by which individuals could evaluate alternative or non-institutional sources of information. This might actually be a good way for individuals who are doing really high quality work to get recognized: being scored consistently well on the more objective rubric could garner you a badge. Importantly, these rubrics would have to require providing concrete evidence from the text being evaluated (screen shots, for example).

From an economic perspective, of course, social media companies need to provide plenty of emotional rewards for putting the effort in. Making it easier to get an A is not enough, the A has to light up your heart (or the “addiction pathways” in your brain). Now that I’ve cleared my head of the ideas rustling around, Happy Sunday!