Social Media Horror Show: How they are wrecking our brains by 'warping our generative models'

Insights from a new model of brain being developed by the cutting edge research in Neuroscience

Look at these graphs first:

Fig. 1: Average PISA scores for OECD countries over past two decades. Source: The Economist

Fig. 2: Indicators of poor mental health among U.S. girls and young women, 2001–2018 Source: Psychiatric Research and Clinical Practice Volume 2, Number 1 https://doi.org/10.1176/appi.prcp.20190015

By now there is no doubt that the performance of students started declining and the incidence of suicide, self-poisoning, major depressive episodes and depressive symptoms in the young started rising sharply around 2010-2012 in US and many other developed countries.

That, you guessed it right, is the time when the use of smartphones and social media exploded in the same age group in those societies. (And that explosion eventually followed in other less affluent societies over the following years as smartphones got cheaper and social media platforms became bigger).

Was that a mere coincidence?

Many researchers think so and point out that the studies so far have been largely unable to establish a definitive causal relationship between the use of smartphones and social media on one side and the decline in mental health and academic performance by the students on the other.

They (many of them begrudgingly) accept that there is correlation between the two but the leading clinical researchers working in this field have steadfastly resisted the suggestion of any casual link.

More broadly, the call from people like US Surgeon General Vivek Murthy and Jonathan Haidt, the author of widely acclaimed ‘The Anxious Generation’ to policy level restraints in social media use have been largely ignored.

I have read and listened to dozens of articles and interviews with arguments from both sides over the years. My experience from my own use of social media and that of people around me largely conforms to the skeptical outlook presented by Murthy and Haidt.

Being a person of science, though, I find it hard to push against the near-consensual belief in scientific community about the absence of any proven casual link between social media use and the decline in mental health.

Well, that is going to change now, at least for me. After reading this post completely, you will understand why.

…

Recently, an insightful interview in Sean Illing’s Gray Area appeared in my podcast feed titled ‘Your Mind Needs Chaos’. The guest was Mark Miller, philosopher of cognition from the University of Toronto. That interview led me to a ground-breaking paper recently published in the journal Neuroscience of Consciousness co-authored by Miller titled Digital Being: social media and the predictive mind.

My first instinct after reading this paper was to post the entire article here at THG as Oxford University Press allows me to do so under the Creative Commons Attribution Licence 4.0.

Eventually, though, I thought it would be more useful if I presented the highlights of the 11-page article and summarized it for you.

Synopsis of the article: while we may still lack the empirical evidence of the causal link between social media use and decline in mental health, a new theory of cognition and affect can provide a sound theoretical framework to explore how precisely our brain is impacted by the pervasive use of social media platforms in their current form:

While concerns around social media have become mainstream, little is known about the specific cognitive mechanisms underlying the purported correlations highlighted by these studies, and why some of us find it so hard to stop engaging with these platforms when things obviously begin to deteriorate. In what follows, we suggest that both the rise in Snapchat surgery, and the connections between social media, depression, and addiction, can be accounted for via a unified theoretical approach grounded in an emerging, and now highly influential, theory of cognition and affect—the active inference framework (AIF). We propose that the structure of some social media platforms constitute ‘hyperstimulating’ digital environments, wherein the design features and functional architecture of digital environments impact the machinery of cognition in ways which can lead to a warping of healthy agent-environment dynamics, producing precisely the sorts of pathological outcomes we see emerging today

If you want to understand the scope of this paper fully, I strongly suggest that you first listen to the interview here or in any podcast app by searching ‘Gray Area’.

Understanding the basics of AIF was a deeply meaningful and rewarding exercise for me. Here is how lucidly the authors introduce AIF to us:

The revolutionary move of the AIF is to re-imagine the brain as a prediction engine constantly attempting to predict the sensory signals it encounters in the world and minimizing the discrepancy (‘prediction errors’) between those predictions and the incoming signal (Friston 2010, Howhy 2013, Clark 2016). contd…

Here it is notable that this paradigm differs totally from the conventional understanding in neuroscience that our brain, like a camera is a (passive) receptive system—rather than an active prediction engine—waiting for signals from world before forming its mental picture.

…contd. In order to make apt predictions, these systems need to build-up a ‘generative model’: a structured understanding of the statistical regularities in our environment, which are used to generate predictions. This generative model is essentially a model of our world, including both immediate, task-specific information, as well as longer-term information that constitutes our narrative sense of self. contd…

The Miller interview in Gray Area elaborates the concept of brain as prediction engine with multiple examples of activities that we engage in to help our brains make better predictions. Many of those examples are related to the multitude of papers Miller has authored himself. Here, though they elaborate this theory with an illustrative example:

…contd. According to this framework, predictive systems can go about minimizing prediction errors in two ways: either they update the generative model to more accurately reflect the world, or they behave in ways that bring the world better in line with their prediction (Clark 2016). In this way, the brain forms a part of an embodied predictive system, which is always striving to move from uncertainty to certainty (Nave et al. 2020). By successfully minimizing potentially harmful surprises, these systems keep us alive and well. Consider the healthy and highly expected body temperature for a human being of 37°C. A shift in temperature in either direction registers as a spike in prediction error, signalling to the organism that it is moving into an unexpected, and therefore a potentially dangerous state. As long as the change in temperature is not too extreme, we could just sit there and come to terms with the changing temperature (update our generative model), or we might reach for a blanket or open a window. In these cases, what we are doing is acting upon our environment, sampling the world and changing our relation to it, in order to bring ourselves back within acceptable bounds of uncertainty.

…

Before delving deep into the insights from the paper, though, let’s examine the stakes of understanding our relationship with social media and the cost of not getting it right.

As I was listening to and reading Miller, two back to back hurricanes ravaged the Southeastern United States. Meanwhile the devastation of the disasters was being compounded by equally devastating conspiracy theories.

According to one, ‘the federal government deliberately steered Helene toward western North Carolina to make room for lucrative lithium mines’. Many believed that the hurricane was caused by humans, most notably the men and women who did weather forecasting and read the updates. Vicious death threats to them followed. Even one US Congresswoman went on to tweet that ‘Yes, they can control the weather. It is ridiculous for anyone to lie and say it can’t be done’.

Can you imagine the level of brainwashing through social media that these people must have gone through before having these beliefs?

If a wave of superiority complex passed through you while reading this for being saner than these people, please hold on. NOBODY including you and me is immune from this process given the media ecosystem around us. If not wholesome conspiracy theories, conspiratorial thinking and logic-building has seeped into the psyche of all of us.

So to me, there is a two-fold value in understanding the brain-social media intersection.

First, it gives us tools for informed introspection and can help us keep our sanity for as long as possible in this world of 8 billion warped generative models. Fighting a danger always starts with recognizing it.

Second, it helps us better understand the world including the direction which the collective humanity is taking. The stakes? Well, the outcomes of upcoming US election will be determined by 150 million or so of variably warped generative models.

Right now there is 50% chance that the people who believe that the proper response to escalating natural disasters like hurricanes, floods, wildfires and droughts is to kill the people linking them to climate change will rule the world’s richest and most powerful country. If they succeed, an unfathomable wave of added human misery will follow over the following decades as climate denial comes home to roost.

Therefore, understanding the dynamics which is shaping our psyches in these rapidly changing times is crucial for us to prepare ourselves for the future at the individual as well as societal level.

…

Finally, I will end this post with the most relatable excerpts of the paper from Neuroscience of Consciousness. After explaining the basics of AIF, the paper rightly focuses on the issue of our brain’s relationship with social media.

Here is how the paper starts:

A recent survey of young people online found that 48% had felt influenced or pressured by social media to consider having a cosmetic surgery procedure (Arab et al. 2019). Moreover, some specific social media influencers have spoken openly about how they undergo cosmetic procedures in order to ‘attract an audience online’, and to maximize the amount of positive engagement a post might receive (Truly 2019). While cosmetic surgery should not be thought of as intrinsically problematic, the potential link between social media platforms and a desire to undergo cosmetic surgery procedures has been dubbed ‘Snapchat surgery’ and represents the most recent addition to a litany of worries around social media, mental health, body image, and general wellbeing.

Here are the other highlights:

Social media can act as a spectacularly effective method for warping our generative models, as it often bombards users with bad evidence about both the offline environment around us and our place in it. Typically, in the offline world, our generative model and expectations are encoded with information incoming from an unfiltered environment, which means that most of the time, our generative model more or less accurately (or at least usefully) reflects the world. However, in cases of regular and heavy engagement with certain content on social media, incoming information about the world is very often carefully selected, curated, and digitally altered—we are potentially engaging with a fantasy. Moreover, apps that offer the use of filters also allow us to represent ourselves in carefully curated ways, potentially cultivating kinds and quantities of feedback and validation simply not available to us when we go offline. The space between being and appearing is potentially vast—with a few swipes, we can dramatically alter our appearance, or retake the same picture 20 times until our face exudes precisely the calm mastery of life we want to project. As social media platforms develop features which foster an increasing potential for inauthenticity, the more those platforms become potentially powerful bad-evidence generators, flooding the cognitive machinery of their users with inaccurate information, telling us that the world is full of incredibly beautiful, cool people, living wonderfully luxurious lives: social media platforms can act as a digital crowbar, prising apart our generative model from the offline environment. Instead, our model of the ‘real’ world comes to produce the expectations generated through the online environment, and the result is, potentially, increasingly unmanageable waves of prediction error which the system must now strive to minimize.

…

The seemingly extreme actions of seeking cosmetic surgery to look more like one’s online presence, then, are part of just one strategy for resolving this kind of prediction error. A recent survey found that more than half of cosmetic surgeons had patients ask explicitly for procedures which would enhance their online image, while many also reported patients using enhanced images of themselves as an example of how they would like to look (Hunt 2019). One filter, mentioned by prominent social media influencers, allowed users to preview the effects of specific cosmetic procedures, and while Instagram has now banned that specific filter, many perform very similar functions. While this may seem extreme, these actions make perfect sense when viewed through the AIF. If we become accustomed to our own doctored appearance, and to receiving all of the feedback associated with it, soon the level of validation available offline will be registered as a mounting prediction error that is likely to result in feelings of stress and inadequacy.

…

Returning to the case of Snapchat surgery and social media, note how high the stakes are in this scenario. Surgery might offer one way to attempt to resolve the mounting error, but if we are unable to resolve the error and continue to engage with social media, then this consistent failure is fed back to the system, eventually teaching it to expect its own failure and inability to act effectively in the world. This ‘pessimistic’ tendency in prediction bears a striking resemblance to the kind of scenarios now described by neuroscientists working on computational accounts of depression based on the AIF. Various forms of psychopathology, including depression, have now been described as a form of ‘cognitive rigidity’, wherein the system fails to adjust its expectations (including expected rate of error reduction) in line with feedback from the world

…

When the potentially enormous levels of social feedback do come, it is not immediately communicated to the user. Rather, we receive notifications in the form of a shining button or exciting sound which delays the discovery of the precise nature of the incoming content. The simple act of pushing a button to reveal information has been shown to trigger arousal and compulsive behaviour, and newly developed features on smartphones add further layers of anticipation (Atler 2017). The ‘swipe to refresh’ feature of the Facebook app’s news feed, for example, where users physically swipe the screen to generate a new stream of information, is a startlingly similar action to the pulling of a casino slot machine arm. In each case, users do not know for sure what kind of content will spring up until they swipe. This feature, coupled with the fact that Facebook’s feed is now effectively infinite has led to the app being described as ‘behavioral cocaine’ (Andersson 2018). One final layer of anticipation comes through the use of a smartphone itself, compounding the intermittency arousal of feedback and interaction: we have been so conditioned by anticipation of smartphone buzzing that ‘phantom vibration syndrome’—the erroneous sensation of our phone vibrating—now affects 65%–89% of people who use smartphones (Rothberg et al. 2010, Drouin et al. 2012). Crucially, these carefully engineered spikes in user anticipation mirror the anticipatory states known to underlie problematic gambling; in people who exhibit addictive gambling behavior, dopamine response has been shown to be most pronounced during phases of high anticipation (Hegarty et al. 2014). Rather than the reward itself, it is these highly arousing states of expectation of reward which have been shown to elicit the strongest dopaminergic response (Van Holst et al. 2012, Linnet 2014), and the designers of these digital platforms know this.

And this is how they conclude the paper:

The AIF has, in recent years, come to change how we understand a range of psychological phenomena, including addiction and depression. In this paper, we have used the theoretical tools of active inference to enter into an ongoing debate about the ways in which social media—and digital environments more broadly—have the potential to negatively impact our mental wellbeing. While deflationary accounts downplay the effects of the design and structural features of digital environments, this active inference account adds weight to arguments that there are engineered features of digital technology that can have profound consequences for our wellbeing. These arguments may have a wide ranging impact, given that these inherent features are deliberately implemented. As design guru Nir Eyal states, ‘Companies increasingly find that their economic value is a function of the strength of the habits they create’ (Eyal 2014). As it turns out then, the designers of social media, aiming to maximize engagement through design, may have a de facto interest in increasing the corrosive effect their platforms have on the mental health of users. Seen in this context, this emerging picture may lend significant weight to arguments that we should take digital hyperstimulants seriously as a threat to our wellbeing, and to voices calling for changes to the way digital technology like social media is designed, operated, and regulated.

After listening to Miller and reading his paper multiple times, I am convinced that this is one of the most important pieces of research of public importance that has to be publicized as widely as possible. I have started an explainer live video series in my Youtube channel (you can check the first three episodes here, here and here) but it is in Nepali and not useful to all of you but I will try to make at least one video in English.

If you came this far, I hope you’ll recommend—and forward—this piece to a couple of your friends. If you already subscribe to THG, you may consider pledging to pay for THG as soon as it goes paid here:

If you missed my earlier pieces related to neuroscience, you can check them here: