Written evidence submitted by Professor Keith Hyams and Dr Jessica Sutherland (SMH0015)

Social Media, Misinformation and Harmful Algorithms

Introduction

1. Professor Keith Hyams and Dr Jessica Sutherland are part of the Interdisciplinary Ethics Research Group (IERG) in the Department of Politics and International Studies at the University of Warwick. They are currently involved in a Horizon Europe project Knowledge Technologies for Democracy, which looks at the opportunities and risks for democracy arising from AI and big data technologies. The work of the IERG focuses on identifying ethical issues that arise in areas such as democracy, social media, big data, and AI.
2. As our communication, democratic participation and social lives have moved increasingly online so to has our knowledge dissemination. Whilst increased social media use has brought with it benefits such as the democratisation of knowledge (almost everyone can access information in real time), it has also brought with it increasingly challenging risks such as the risk of manipulation and problems arising from the ultra-personalisation of content feeds.
3. This evidence focuses on the epistemic, individual and social harms that recommender algorithms can cause. We characterise epistemic harms under two umbrellas: epistemic pollution and changes in the epistemic terrain. The individual harms we discuss are related to changes in individual behaviour such as negative outcomes for health behaviours after the spread of COVID-19 misinformation and disinformation online. Finally, we discuss wider social harms such as the effects of algorithmic sorting and targeting on democratic participation, polarisation and cohesion.

The Use of Recommender Algorithms by Social Media Platforms

4. Because social media platforms aim is to keep us engaged for as long as possible, thereby generating revenue, their algorithms are designed to keep us scrolling. Put simply, social media algorithms are designed to sort our profiles and content into different categories and target our feeds with the content that is most relevant according to this sorting. Whilst what is relevant to us may take many different forms (e.g., different content types like videos, text or visuals, or different content subjects like politics, beauty or gaming), the aim is to provide us with content that we are most likely to engage with. Recommender algorithms are therefore increasingly being used by social media platforms to promote content to our feeds. Essentially deep learning algorithms predict what content you are most likely to engage with based on your user data and past behaviour such as likes, comments, accounts you follow, and your demographics, and fill your feed with this.
5. Whilst recommender algorithms are not the only type of content broadcasting means that social media platforms use - for example they also use information such as who we follow, and what we search for in their collection of algorithms that form ‘the algorithm’- they are now the most dominant algorithm and drive most of the content we see on our feeds.
6. Because they use deep learning and machine learning techniques, they are less interpretable than the algorithms social media platforms previously used. For example, many social media feeds used to be simply ordered reverse-chronologically for the posts of accounts you followed. Less transparency in how decisions on what content to promote are made means developers, users, policymakers and researchers are unable to scrutinise these decisions and understand why certain content is promoted over other content. In particular, it is often argued that recommender systems push divisive content since we are more likely to engage with content that causes outrage, albeit negatively.[1] Since the algorithms used do not take into account sentiment of engagement, all engagement counts. Because of this, links between recommender algorithms and the spread of hate speech and propaganda have been made.[2]
7. A related problem that has arisen in recent years is the promotion of content from ‘verified’ accounts over and above content from others. As part of some paid-for verifications, the algorithm will prioritise your content on other users’ feeds. This allows users for a small fee to gain greater reach than may have been possible before.
8. We cannot however attribute all of the problems that arise from social media use to algorithms and platform design. For one, recommender algorithms learn from our behaviour and promote content to us accordingly. Second, content of course must be created (or generated in the case of generative AI and bots) by someone. Social media algorithms simply amplify the effects of human behaviours.

Epistemic Harms

9. Since we use social media to acquire and share various forms of knowledge and opinion, including political opinions, news, and culture, social media platforms have quickly become an important form of knowledge technology. The way that content is provided to us have therefore become imperative to our understanding and perception of political and cultural issues.
10. As mentioned above, social media algorithms are often linked to the spread of hate speech and political extremism online. We can characterise this kind of harm as a form of epistemic pollution since harmful content such as fake news, misinformation and disinformation are introduced into our epistemic landscapes that may have otherwise not been if we did not use these platforms. This is only further perpetrated by the proliferation of bot accounts and tactics such as astroturfing that distort the flow of knowledge on social media platforms by targeting specific debates. So, even if both sides of a debate may be shown on our feeds, the arguments and support for one side may be inflated beyond actual public opinion. This does not mean these platforms do not also produce epistemic goods, such as the democratisation of knowledge sharing and giving a voice to those who may otherwise be silenced, but rather that the epistemic benefits of such open knowledge sharing may sometimes be tainted by the negative effects of such unregulated platforms.
11. However, even if the content itself is not particularly divisive, recommender algorithms can also alter the epistemic terrain by creating filter bubbles, echo chambers (although their existence disputed), and the ultra-personalisation of our feeds which simply changes what we see versus what others may see. This is harmful to society since it can drive polarisation as different narratives are spun, and this content does not necessarily overlap on our feeds. We may therefore only see one side of the argument or story, further driving us into these filter bubbles and echo chambers.
12. Other epistemic harms that can arise from recommender algorithms including a form of flattening of the epistemic landscape where minority or alternative viewpoints are not seen in equal measure to more extreme opinions as they are not as likely to be engaged with, and relatedly a kind of testimonial harm to these individuals who are excluded from debates as their content is not promoted to others.

Individual and Social Harms

13. If the ways that we acquire knowledge are becoming distorted and manipulated in the ways outlined above, then there may be effects on our individual behaviours. This was particularly seen in vaccine hesitancy behaviours surrounding COVID-19, and the spread of misinformation that led to the anti-immigration riots in the summer of 2024. These behaviour changes may be unconscious, for example by slowly changing our attitudes over time, or conscious because of a growth in mistrust online driving a distrust in institutions, experts and governments.
14. These changes in individual beliefs and behaviours further exacerbate polarisation in the public sphere. The impact of misinformation and disinformation could therefore be described as a cycle in which those who fall foul to misinformation may begin to further spread this misinformation or fake news both online and offline, further amplifying the misinformation and its associated views.
15. Relatedly, the ‘conspiracy loop’ – a feedback loop in which attention paid to conspiracy theories in public debate ultimately reinforces belief in the conspiracy theory in the conspiracists mind (and often distrust as outlined above) – also plays a role in, and is exacerbated by, polarisation online and offline.[3] When a conspiracy theory is shared and amplified online (either because of the algorithm promoting engagement, or individuals’ actions promoting the theory), these conspiracy theories then become part of political debates. The negation of the conspiracy theory may end up reinforcing conspiracies that call for distrust in political and academic institutions as an example.
16. Social cohesion may also therefore be impacted as people are driven towards the two extreme sides of debates, both by the algorithm itself and disinformation, misinformation and fake news. Again, we have seen this in the tension both before and after the summer 2024 riots between both sides of the political spectrum, and it is also seen around other issues as varied as climate change, healthcare debates, and election results. Essentially, when we are sorted into silos online – either because we seek them out (for example via out choice of social media platform or the accounts we choose to follow) or because the algorithm nudges us there – we are less able to communicate across the divide, and the gap between the two sides (either on political issues or otherwise) widens.
17. A final harm we would like to highlight is the impact that this polarisation and erosion of social cohesion can have on the ways we as individuals and a society participate in democracy. Whilst social media platforms may have initially been seen as arenas for political debate where anyone can participate, given the ways that both the algorithms and bad actors can influence and distort the direction of debate online, and has potential real-world consequences on individuals’ political views, the directions of political debates, and voting intentions.

Conclusion

18. The use of recommender algorithms by social media platforms has the potential to manipulate, distort and alter the content we see on our feeds. This is intrinsically harmful as it changes the content we have immediate and unfettered access to (thereby changing the epistemic landscape), and instrumentally harmful as it can produce negative social effects such as increased polarisation, decreased social cohesion, and negative effects to our behaviours.

17 December 2024