2.1. The role of producers

Every day, on social media, millions of people produce and share information, in the form of written posts, images, videos, or reviews of any imaginable product or experience, often with no measurable return. From a cultural evolution perspective, this behaviour may appear paradoxical (Acerbi, Reference Acerbi2019b): information producers pay a personal fitness cost that allows social learners – information scroungers – to benefit from their exploration (Rogers, Reference Rogers1988).

An ecological approach to culture reverses the issue by asking what goals individuals pursue when they produce and circulate cultural items (Baumard & André, Reference Baumard and André2025). The distinction matters, because it shifts the explanation from the properties that favour the spread of a cultural trait to the incentives that shape its production. Traditionally, cultural evolution has explained the properties of cultural traits by considering characteristics of the receivers or consumers of the traits. The ‘producers’ standpoint’ (André et al., Reference André, Baumard and Boyer2023) suggests instead that the properties of cultural traits reflect (also) the payoffs that individuals expect to obtain. To illustrate, from a consumer-focused perspective, one might conclude that content about possible threats spreads widely because more likely to capture attention and be retained. A producer-based interpretation suggests instead that individuals often create or share material about possible threats because it is effective for their goals, such as signalling alertness or competence (Blaine & Boyer, Reference Blaine and Boyer2018; Boyer & Parren, Reference Boyer and Parren2015).

Attracting attention can be considered a meta-goal for any other activity. Attention is quantifiable on most platforms. Producers can see in real time how many people have viewed, liked, or shared their posts, and these metrics constitute a salient reward structure. Lindström et al. (Reference Lindström, Bellander, Schultner, Chang, Tobler and Amodio2021) showed that we are sensitive to these rewards, and we calibrate our posting activity to maximise engagement signals (for a similar approach, see Turner et al., Reference Turner, Gunschera, Subrahmanya, Salecha, Eichstaedt, Palminteri and Orben2025). Attention-oriented content typically displays universal features that, on average, facilitate engagement from the consumer point of view: high-arousal emotion, surprise, humour, provocation, or novelty, which have been often studied. The often-noted success of negative content fits this logic (Baumeister et al., Reference Baumeister, Bratslavsky, Finkenauer and Vohs2001). Brief formats, such as short videos, punchy statements, and concise images, are effective because easier and faster to process, and producers who quickly capture attention have the additional advantage of being likely rewarded by platform algorithms.

However, more specific predictions can be made. First, different platforms, and different communities within platforms, can reward different content (see ‘Different platforms as different ecologies’ section). Second, different content is linked to distinct goals and incentives. We will analyse some of them: reputation management, coalition building and identity management, participation in shared activities and coordination. These goals do not represent an exhaustive list, but they capture the main types of incentives that drive production in social media. Importantly, they also allow us to infer the features we should expect successful content to have.

From this perspective, cultural patterns on social media are not only, or primarily, the result of biases that govern how consumers acquire information. They are shaped by the strategies that producers adopt, and what becomes widespread is what best serves these strategies. Here, ‘production’ covers every action that can be performed on social media, with a different level of engagement: from creating and posting novel content, including comments, to sharing content produced by others, or endorsing it (with a ‘like’ or similar). In this sense, the distinction between producers and receivers is used as an analytical convenience rather than as a claim about distinct categories of individuals: all users are strategic agents. In fact, the case of social media is particularly interesting exactly because productive acts – like posting, sharing, or liking – are not only cheap and available to virtually all users but also persistent and potentially visible to large and heterogeneous audiences.

2.2. Transmission and replication

Drawing on what I discussed above, the primary aim of social media activity is rarely transmission. While this could be the case for specific goals, like participation in shared activities or coordination, transmission is often a by-product of other motivations rather than a goal in itself. For this reason, when we study informational dynamics on social media, we may want to shift our analytical attention from the mechanics of transmission to the goals that generated it.

A case that is probably familiar to readers is the use of social media by academics. When academics post information about their research, links to their papers and preprints, or conference talks, they do produce and transmit information, but they are not uninterested information producers. To understand this dynamic, we cannot focus only on receivers’ biases, but we need to know the motivations behind sharing, generally self-promotion. Even when academics share works from other researchers, producing useful information, their goal may be to build their reputation as competent scholars in a field and to let colleagues know that they are up-to-date on a particular topic. In most cases, they will share information that is consistent with their perspective, but they could gain extra reputation by sharing and commenting on different views. Generally, there will be a strong incentive to share true information, as the academic community will be able to assess its quality, and reputation damage could be costly.

It can be useful, as an exercise, to consider an illustrative case where visibility differences might appear to reflect receiver bias. For example, women’s papers tend to be less present or less engaged with on social media. A transmission-based interpretation would attribute this to audiences being less inclined to attend to or share work authored by women. Evidence from Peng et al. (Reference Peng, Teplitskiy, Romero and Horvát2025) points to a producer-level mechanism instead: women are substantially less likely than men to promote their own papers online, even after controlling for academic status and productivity. In this example, differences in diffusion patterns emerge not because receivers prefer certain authors, but because producers differ in how they use social media.

Relatedly, even if digital transmission offers a cheap and virtually error-free possibility of replication that would nicely fit with Darwinian-inspired models of transmission, social media dynamics make explicit that replication fidelity has little explanatory power by itself. Transmissibility is effectively constant across cultural items, so the relevant question becomes why individuals choose to replicate one item rather than countless others that are equally easy to copy and reproduce.

Moreover, cultural dynamics on social media often result from transformation, reframing, and embedding in new contexts. This can be clearly seen by analysing the diffusion of internet ‘memes’, whose name suggests Darwinian-like replication (for a longer discussion of non-replication of memes, see Acerbi (Reference Acerbi2019b, Chapter 7)). In an early study on ‘information evolution’ on social media, Adamic et al. (Reference Adamic, Lento, Adar and Ng2016) considered Facebook memes from 2009 to 2011. At the time, Facebook did not have a built-in ‘share’ button, so text needed to be copied-and-pasted manually. Successful memes, replicated verbatim, existed: ‘No one should die because they cannot afford health care and no one should go broke because they get sick. If you agree, please post this as your status for the rest of the day’ was found in the exact same form more than 470,000 times. However, this was rather the exception: the ‘mutation rate’ of transmission was, surprisingly given the digital support, extremely high. On average, 11 out of 100 copies were different from the original.

These differences, moreover, were not random. Some were clearly deliberate innovations. Common variants of the text above included ironic takes (‘No one should be without a beer because they can not afford one’) or appeals to the opposite political spectrum (‘No one should go broke because government taxes and spends …’). Others were likely due to psychological or technical constraints, with shorter (but not too short) sentences being copied more faithfully, or differences and omissions being more likely to be found at the beginning or at the end of texts. Because of the prevalence of these transformations, the researchers considered as a ‘meme’ not a discrete cultural token but a family of variants held together by partial resemblance, similarly to the concept of cultural attractor (Miton, Reference Miton, Tehrani, Kendal and Kendal2025).

In sum, the spread of internet memes can hardly be modelled as a process of reproduction, random mutation, and receivers’ biases. Memes became associated with coalition building and identity management goals independently of their intrinsic features (e.g., ‘Pepe the Frog’). Others are templates for mimicry or remix (Shifman, Reference Shifman2013). Historical memes like ‘Grumpy Cat’ or ‘Jealous Girlfriend/Distracted Boyfriend’ owe their success to the possibility of endlessly editing their captions.

This does not imply that transmission is irrelevant in social media dynamics. What spreads must still be intelligible and appealing enough to be picked up by others, and many producer-driven contributions would fail without some degree of transmissibility. The point is rather that transmission does not operate in isolation. High-fidelity replication is only one of several pathways through which information can circulate, and often not the dominant one.

2.3. Different platforms as different ecologies

Rather than treating ‘social media’ as a homogeneous domain, an ecological view emphasises that each platform creates a different landscape of incentives for cultural behaviours. Platforms differ in audience structure, affordances, feedback metrics, and algorithmic distribution mechanisms, and these differences systematically shape the kinds of cultural items that are produced and how they change over time.

3.1. Misinformation

One of the biggest concerns that accompanied the diffusion of social media has been that they would provide a fertile ground for the spread of misinformation. More recently, many assumptions behind this idea have been challenged. For example, while some features of social media transmission dynamics could favour the spread of misinformation, misinformation is not a social media-specific problem (misinformation from traditional media and, especially, from public figures can be much more impactful). In addition, when compared with the total information present online, or on social media, flagged misinformation represents a tiny minority (Altay et al., Reference Altay, Berriche and Acerbi2023).

From the perspective defended here, however, the most interesting point is not the relative quantity of misinformation but the causal model that underlies most research on the topic. The standard narrative assumes a direct link from exposure to a piece of misinformation to a change in beliefs or behaviour. This causal link cannot be taken for granted. Research shows that ideas and beliefs, especially political ones, are entangled with social and economic factors, and that people rarely change their minds in response to propaganda messages; attitude change is more likely when individuals encounter convincing arguments within trusted relationships (Mercier, Reference Mercier2020). Moreover, if we take the producers’ point of view, misinformation does not spread because people are prone to believe false content, but because producing and circulating certain kinds of information can be useful for particular goals within specific online environments.

From the producer’s standpoint, many forms of misinformation are by-products of incentives that reward attention, identity signalling, or coalition maintenance rather than accuracy. Political misinformation, for instance, is shared almost exclusively by a small subset of individuals who already endorse the views expressed, who also share partisan news from reliable outlets, and who strongly dislike their political opponents (Osmundsen et al., Reference Osmundsen, Bor, Vahlstrup, Bechmann and Petersen2021). Emotional, sensational, or norm-violating content is more likely to generate engagement and therefore more likely to be rewarded with visible feedback, whether the content is true or false.

As a consequence, misinformation mostly shares the same features of reliable information. For example, mainstream news online also gathers engagement through divisive and, up to a point, negative content (Fischer & Acerbi, Reference Fischer and Acerbi2025). In theory, a possible difference is that false content can be tailored to maximise attention, while true content is, by definition, more constrained by reality (Acerbi, Reference Acerbi2019a). Some research found indeed that, analysing US television broadcasters, inaccuracy and the degree of negativity bias are correlated (Soroka & Wlezien, Reference Soroka and Wlezien2026). However, content-related differences between news produced by reliable and unreliable outlets are likely to be too small to be used to distinguish them, exactly because producers have comparable incentives.

This perspective also changes how we think about interventions against misinformation. Almost all interventions tested so far adopt the consumer’s point of view. Debunking, pre-bunking, nudges, and digital literacy aim to make receivers more sceptical of unreliable sources or false claims. It has already been noted that these interventions have the by-product of making us sceptical of all news, which, given that reliable ones are the majority, can have a marginal negative effect (Altay, Reference Altay2022). However, what about the producer’s standpoint? Individuals recognise that sharing misinformation can be harmful for their reputation (Altay et al., Reference Altay, Hacquin and Mercier2022), and sharing reliable and useful information elicits epistemic gratitude (Karabegovic et al., Reference Karabegovic, Wang, Boyer and Mercier2024), a positive social signal that can strengthen trust and make future information exchange more likely.

Interventions, or even platform architectures, could therefore build on these incentives. Rather than focusing solely on reducing exposure or correcting beliefs, platforms could make reputational consequences for accuracy more salient, increase the visibility of epistemically valuable contributions, or design features that reward reliability and domain expertise. The spread of misinformation already varies across platforms and niches: for example, LinkedIn, where professional reputation is at stake, shows lower rates of political content and likely misinformation than platforms organised around anonymous or ideological audiences. This suggests the possibility of shaping environments so that the incentives for producers align less with attention at any cost and more with demonstrating reliability, competence, or usefulness.

Of course, this raises the question of who would actually implement such changes. Platform providers operate under their own incentive structures, primarily optimising for engagement and retention; there is no obvious reason why they would voluntarily redesign architectures to reduce the spread of content that performs well by their own metrics. Meaningful change at the platform level is therefore likely to require external pressure (e.g., regulatory mandates such as the EU’s Digital Services Act) or explicit costs, for example, when misinformation episodes become commercially damaging. A further complication concerns users’ reactions to these changes. If platforms become less rewarding in engagement terms, say, more accurate but less stimulating, users may simply migrate to alternatives that better serve their goals. This competitive dynamic may be as difficult to overcome as the platforms’ own engagement incentives, and is itself a prediction the ecological framework would generate: individuals, like platforms, move towards the niches that best serve their interests.

3.2. Radicalisation and polarisation

Concerns about online radicalisation often rely on the same causal model of mainstream research on misinformation: individuals are exposed to extreme content and consequently shift towards more extreme attitudes. Empirical evidence provides a more nuanced picture. Large-scale audits of YouTube’s recommendation system show that while the algorithm tends to supply ideologically congenial material, it does not systematically push users towards more extreme content. In a study using 100,000 sock-puppet accounts, Haroon et al. (Reference Haroon, Wojcieszak, Chhabra, Liu, Mohapatra and Shafiq2023) find that recommendations become more aligned with inferred ideology, and that exposure to problematic channels remains very limited in absolute terms. Complementary experiments using counterfactual bots demonstrate that relying solely on YouTube’s recommendations actually leads to less partisan consumption than following real user histories, and that sidebar recommendations ‘forget’ earlier extreme viewing patterns (Hosseinmardi et al., Reference Hosseinmardi, Ghasemian, Rivera-lanas, Horta Ribeiro, West and Watts2024). These results suggest that observed radicalisation dynamics stem primarily from users’ pre-existing preferences rather than algorithmic steering.

From an ecological, producer-centred, perspective, polarising and extreme content emerges not from passive exposure but from the incentives operating within specific niches. On YouTube, political creators compete for attention within ideologically coherent communities, and platform affordances may reward provocative and identity-affirming content. Munger and Phillips (Reference Munger and Phillips2022) describe this as a ‘supply-and-demand’ dynamic, where committed audiences create a market for fringe producers and the affordances of the platform facilitate production at scale. Similar logics apply across social media: in homophilous political clusters, producers benefit from content that marks boundaries, signals loyalty, or displays moralised outrage. Polarisation, therefore, reflects the selective pressures of particular online ecologies and producers’ interests rather than a universal susceptibility to extreme messages.

An ecological perspective also explains why broad claims about algorithmic polarisation often overstate the empirical case. Evidence suggests that patterns of hostility or extremity observed online frequently reflect dispositions that predate social media use: individuals who behave uncivilly in political discussions do so across multiple online contexts (Mamakos & Finkel, Reference Mamakos and Finkel2023); those preoccupied with social status display hostile political behaviour both online and offline (Bor & Petersen, Reference Bor and Petersen2022); and increases in polarisation on platforms such as Reddit have been shown to depend largely on the arrival of already polarised users rather than on the radicalisation of existing ones (Waller & Anderson, Reference Waller and Anderson2021). Moreover, offline social environments are themselves far from politically diverse: everyday friendship networks and workplace interactions often exhibit ideological homogeneity (Gentzkow & Shapiro, Reference Gentzkow and Shapiro2011), residential patterns show marked political segregation (Brown & Enos, Reference Brown and Enos2021), and traditional media ecosystems also display strong partisan alignment, often as pronounced as, or even greater than, that observed on social media (Fletcher et al., Reference Fletcher, Cornia and Nielsen2020; Muise et al., Reference Muise, Hosseinmardi, Howland, Mobius, Rothschild and Watts2022). Much evidence, in sum, suggests that polarisation emerges from deeper social and economic factors and is only reflected online. In particular, polarisation appears most strongly in environments where user composition, platform affordances, and reward structures jointly favour antagonistic signalling, typically within already politicised niches, rather than across social media as a whole. Interventions that focus exclusively on limiting exposure overlook this dynamic. More promising approaches would target the incentives that make polarising content adaptive: reducing engagement rewards for moralised outrage, elevating signals of reliability or crosscutting engagement, or strengthening reputational pressures in anonymous or weak-tie environments.

3.3. Success

The arguments outlined thus far suggest that the content that becomes successful on social media may satisfy different constraints, so that no single formula can account for ‘virality’. On one side, as receivers are more likely to engage with content that is easy to process, emotional, negative, and group-relevant, this content will tend, on average, to proliferate online. Social media, after all, are very effective in bringing us the content we want to consume (whether this is the content we should, normatively, consume is another question). Moreover, with a potential audience of billions of individuals connected online worldwide, a viable strategy is to appeal to broad, universal, cognitive factors of attraction.

On the other side, content becomes successful not only because it attracts receivers, but because it affords other producers opportunities to pursue their own goals by resharing or modifying it. These dynamics become even clearer once we consider, as we did, that social media are not uniform environments but collections of informational niches. A content trait that fits the incentives of one niche may fail in another.

Variation across individuals interacts with platform diversity. Not all producers compete along the same dimensions, and different strategies can be simultaneously successful because they appeal to different audiences. Quantitative analyses of engagement patterns consistently show that predictive models of content that generates engagement online have small effect sizes (R ² ∼ 0.1–0.2 in studies that report it directly, e.g., Schöne et al., Reference Schöne, Parkinson and Goldenberg2021, Reference Schöne, Garcia, Parkinson and Goldenberg2023; and considerably lower in others, e.g., Bak, Reference Bak2026; for studies using other metrics finding small/heterogeneous effects, see, e.g., Brady et al., Reference Brady, Wills, Jost, Tucker and Van Bavel2017, Reference Brady, Rathje, Globig and Van Bavel2025; Fischer & Acerbi, Reference Fischer and Acerbi2025; but see, e.g., Rathje et al., Reference Rathje, Van Bavel and van der Linden2021; Watson et al., Reference Watson, van der Linden, Watson and Stillwell2024, for moderate effects). While these small effects can have a cumulative impact on millions of messages, they also suggest that heterogeneity arises because users optimise distinct factors of attraction. In political niches, highly moralised or confrontational posts may be rewarded; in informational niches, accuracy and clarity generate engagement; in lifestyle or creative niches, humour and narrative skill may be the most effective strategies. The same platform can therefore host creators who thrive through negative or provocative content and others who succeed through expertise, helpfulness, or aesthetic expression. This diversity in adaptive strategies explains why attempts to isolate a single, universal predictor of virality consistently yield weak or partially inconsistent results: success is relative to the goals of the producers and the ecological constraints of the niches in which they operate. A similar logic comes from research on fiction: narrative success depends on heterogeneous combinations of emotional, stylistic, and thematic elements, with no single factor dominating across works or audiences (Dubourg et al., Reference Dubourg, Thouzeau, Beuchot, Bonard, Boyer, Chevallier, Clasen, Boon-falleur, Fiorio and Fitouchi2024). Cultural success, in other words, emerges from many possible adaptive strategies, each shaped by how well a trait serves the goals of producers while fitting the dispositions and expectations of the audiences who consume it.