06 Dec 2022
(by Peter Ormosi[i])
Digitisation has entirely rewritten the way we consume audio and audio-visual content. Some of the market failures that characterised the analogue broadcasting era, have seemingly disappeared in a world where evaporating fixed costs mean that the most niche content is now available online. In this world, there are increasingly vocal arguments that there is no longer need for Public Service Broadcasting (PSB) if the market can also provide for even the most diverse audiences. But the large platforms which host vast amounts of digitalised content, may not quite be the land of milk and honey some had thought for consumers, and certainly not for the producers of content. The reason is to do with how the content is provided on the platform, through algorithmic recommendations, which are likely to suffer from biases. These biases have the tendency to create a world, that suffers from the same market failures (including the disproportionate underrepresentation of diverse, and niche content) that justified PSB in the analogue era. Because of these market failures, a new generation of PSB could have a critical role in ensuring the provision of online content that adheres to objectives, such as the provision of creative, high quality and distinctive output and services in a way that reflects, represents, and serves diverse communities.
Among the most frequently cited missions of PSB is to educate, provide impartial news, and in general to reflect, represent and serve diverse communities. One of the tests used when justifying the need for PSB, is to look at how much these provisions would be delivered by market-based broadcasting services, and whether PSB is capable of catering for those provisions that the market doesn’t. Put differently, PSB should provide the services that the market wouldn’t deliver or would deliver in a suboptimal or insufficient way.
At the time when broadcasting capacity was limited and broadcasting services were financed exclusively from advertising, there was a strong case for PSB provision. This was partly because of a market failure (preference externality) in analogue broadcasting, whereby content preferred by the majority squeezed out content preferred by viewer minorities. As a result of these externalities, a broadcasting system without PSB would not have catered for tastes outside of the narrow mainstream.
The Internet changed these dynamics. Consumption is shifting from linear broadcasting to online content provision, digitisation, and the massive fall in the fixed costs associated with making content available. This served to foster and significantly expand online content diversity.[ii] For a while it seemed like consumer heaven. A world where the market delivers all content that there is demand for, irrespective of its popularity, and irrespective of the preferences of the majority. Cheap access to unlimited content at people’s fingertips. Understandably, demands to cut PSB budgets became increasingly vocal, and for an apparently justified reason. It seemed that information, education, and entertainment of the most niche tastes was provided by the market, without regulation.
Recommended content – biased?
This could have spelled the end of PSB, but any such conclusion overlooks an important feature of how audio and audio-visual content is delivered to consumers on online platforms. These platforms host an enormous array of content that consumers would find very difficult (if not impossible) to fully evaluate and choose from in an informed way. To reduce search and decision costs, the platform deploys a recommender system (RS), which learns the consumers’ preferences, and recommends content that the RS believes the consumer would prefer. The consumer, influenced by their own behavioural biases, is very likely to engage with these recommendations (rather than ignore them). If the recommended content is the one that would maximise a consumer’s utility, then this would all be fine. A win-win for platforms and consumers. But the platform has limited information about its consumers (and sometimes about its content), which means that the recommendation is likely to be imperfect in a systematic way.
It is well documented in the computer science literature that the recommendations provided by RS tend to suffer from various biases (bias in this context means an imperfect recommendation, a deviation from the ideal recommendation which would maximise the consumer’s utility). Fletcher, Ormosi and Savani (2022) offer a lengthy overview of these biases and link them to their impact on competition.[iii]
To illustrate, let’s start with a simple setup, where a streaming platform, and its RS is user-centric, i.e. their goal is to maximise user utility when making recommendations. Because of the inherent information asymmetry regarding the consumer’s preferences, the platform is more likely to recommend content that it has more information on (content that more users have engaged with). This is often referred to as popularity bias in the computer science literature. Popularity bias happens even in user-centric setups, and it’s simply to do with the algorithmic design of the RS. As a result, the RS recommends popular items disproportionately over less popular ones. People who prefer niche content (content in the long tail), lose out, and more importantly, creators of such content also lose out. The cold start problem aptly illustrates this latter point, whereby new content has a disproportionately low (or zero) chance of being recommended, simply because the RS has no data on this new content. The platforms work hard to avoid these biases, but even the smallest bias can have enormous impact on the market, due to feedback loops (a biased recommendation triggers consumer engagement, which feeds back into the RS as preference, which in turn further biases the system).
Needless to say, these biases can be made worse if the platform is self-preferencing (there is an emerging literature on self-preferencing platforms, Fletcher et al (2022) reviews some of these). A self-preferencing platform recommends items, that maximise the platform’s profit rather than the ones that maximise the consumer’s predicted utility (for example because the recommended item is their vertically integrated product). A related case is where the platform is more likely to recommend content if it receives a higher commission for doing so). Once again, because the RS learns from these non-optimal recommendations, it hardwires the biased choices into future recommendations through feedback loops.
Finally, there are also biases, where the RS learns in an active learning environment, or when the RS is combined with a human editor. If for example human editors are more likely to recommend male content, then the RS learns this as a preference, and its recommendations will suffer from gender bias.[iv] Similarly, if the RS learns from past interactions that reflected other biases, such as racial bias, or bias against foreign content, then the recommender will learn and amplify these biases.
A new mandate for PSB?
Because audiovisual content is increasingly delivered through recommender systems, and because consumers are likely to engage with these recommendations, the biases in the RS will have a welfare impact. When consumers do not engage with the content that would maximise their utility, there is a loss in welfare. But much more importantly, recommender systems have a strong ability to distort competition between content producers, which could negatively impact consumers in the long-run. Moreover, if recommendations are biased, and if certain content never gets recommended as a result, in the long run it can stifle innovation for niche content and could disincentivise the creation of such content and lead to more homogenisation.
These conditions can give PSB a new mandate. A world where the provision of audio and audiovisual content is orchestrated by an RS, seems to be characterised by the same market failures that justified PSB in the analogue era. Popularity bias is simply an algorithmic reflection of the preference externalities that had affected market-based content provision before digitisation. In both worlds, the market provision of content would result in the disproportionate underrepresentation of niche content as a result of these biases or externalities.
For this reason, it is far from obvious that PSB should be phased out as a result of digitisation and apparent proliferation of choice. On the contrary, because of the shortcomings of algorithmically recommended content, and their impact on the market, there may be a strong case for a new generation of PSB that can address these shortcomings. As part of this mandate, an important function of PSB would be to enable that online content provision is not affected by RS biases and does not lead to homogenisation of content and taste, to ensure that popular content does not get disproportionately recommended, and that the RS does not block niche content from featuring in recommendations. In practice, of course it can be difficult to identify specific instances of biases. For example, if a viewer is recommended the most popular item, it can be hard to distinguish whether this is the result of popularity bias or whether it is simply the most relevant recommendation for that viewer. There are, however, a variety of techniques available to platforms to evaluate RS for a consumer population. A/B testing could be used for example to generate real consumer interactions and can thus be valuable for assessing how real consumers react to recommendations that are more diverse or novel.
Against this backdrop PSB could have an important role in the provision of online content that adheres to objectives, such as the provision of creative, high quality and distinctive output and services in a way that reflects, represents, and serves diverse communities.
(Although the focus of this short post is on cultural content, it is worth noting that news RS are also susceptible to biases. It is widely documented that this can lead to filter bubbles, echo chambers, and opinion silos, and could enhance and cement political polarisation. It should continue being an important remit of PSB in the digital world to counter these tendencies with unbiased news provision.[v])
[i] The author of this blogpost is Lead on a UKRI Trustworthy Autonomous Systems grant (EP/V00784X/1), researching the impact of recommender systems on suppliers competing on platforms.
[ii] Waldfogel, Joel (2017). How Digitization Has Created a Golden Age of Music, Movies, Books, and Television. Journal of Economic Perspectives, Vol. 31, No. 3, pp. 195-214.
[iii] Fletcher, A., Ormosi, P. L., & Savani, R. (2022). Recommender systems and supplier competition on platforms. Available at SSRN.
[iv] Ferraro, A., Serra, X., & Bauer, C. (2021, March). Break the loop: Gender imbalance in music recommenders. In Proceedings of the 2021 Conference on Human Information Interaction and Retrieval (pp. 249-254).
[v] Raza, S., & Ding, C. (2021). News recommender system: a review of recent progress, challenges, and opportunities. Artificial Intelligence Review, 1-52.