Changing the feed reshaped the debate more than changing users
In simulations of Italy’s Twitter debate, changing what people saw did more to reshape the conversation than asking them to behave differently. A chronological feed—one that presents posts in time order rather than ranking them through a recommender—raised the diversity of sources users attended to and reduced segregation between audiences. Prebunking and accuracy prompts, by contrast, barely shifted segregation. All three interventions reduced engagement.
That is the central finding of a study by Tommaso Pisani, Pietro Gravino and Giulio Prevedello. It is also a reminder that social media interventions can have two different targets: the choices people make after encountering a post, and the system that decides which posts reach them in the first place. The researchers argue that the second target can change the collective shape of public debate in ways that individual-level experiments may not detect.
The finding comes from a simulation, not a platform-wide experiment. The model reconstructs retweeting in the Italian public debate from 2018 through 2022 and tests counterfactuals against the same recorded activity by prominent politicians and news outlets. Its results are useful for comparing possible interventions, but they do not establish that a real chronological feed would produce the same changes among users.
The Science
A social platform’s visible debate is not simply the sum of what users believe. It is also the result of a selection process: a stream of posts is filtered, ordered and recommended, and then users decide what to read or share. These steps are intertwined in practice. If a researcher sees that a person retweeted more or less after an intervention, it can be hard to tell whether the change came from altered exposure, altered behavior, or both.
The Exposure–Engagement model is designed to separate those mechanisms. Its exposure stage represents what users are shown; its engagement stage represents whether they retweet what they see. Six adjustable parameters govern elements such as the influence of similar users on recommendations, the weight placed on leaders’ posting activity, and the persistence of users’ past retweeting patterns.
The researchers fitted the model to Twitter activity in Italy across 55 monthly transitions, from February 2018 to December 2022. They treated politicians and news organizations as “leaders,” and the other accounts interacting with them as users or followers. Leaders’ posts formed the content supply; user retweets served as the study’s observable measure of engagement. The raw dataset reported in the paper contains more than 29 million retweets and nearly 14 million tweets.
This choice makes the question concrete, but also sets a boundary around the answer. Retweeting is not the whole of social media use: it does not capture every view, like, reply, private conversation or change in belief. Nor does the model forecast what leaders will post next. For each simulated month, it takes the leaders’ actual observed activity as given and models how users might respond to it.
The team first tested whether the model could describe the monthly data better than a simple baseline that assumes each user’s retweet pattern stays exactly as it was the previous month. The fitted model did better in all but one month. The exception was March 2020, when COVID-19’s arrival in Italy disrupted ordinary patterns—a useful illustration of how a sudden external shock can overwhelm relationships learned from recent history.
They then tested whether the model captured not just individual retweet patterns but properties of the debate as a whole. Engagement meant the average number of retweets per user. Attention diversity measured how broadly a user’s retweets were distributed across leaders: a higher value means attention is less concentrated on a small set of sources. Segregation was measured through modularity, a network statistic that rises when leaders share followers mainly within distinct groups and have less audience overlap across groups. In this setting, higher modularity is a structural sign of a more divided debate.
What They Found
The model reproduced some system-level patterns more convincingly than others. Across months, its simulated engagement closely tracked empirical engagement, with a Pearson correlation of 0.86. Attention diversity had a similarly strong correlation of 0.85. Segregation’s correlation was more moderate, at 0.61. These numbers describe how closely simulated and observed month-to-month values moved together; they are not measures of causal certainty.
How closely did the model track observed monthly patterns?
Pearson correlations between empirical metrics and model-simulated means across 55 months.
| Label | Value |
|---|---|
| User engagement | 0.86 |
| Attention diversity | 0.85 |
| Segregation | 0.61 |
There are important biases beneath those correlations. Simulated attention-diversity values were compressed: the fitted slope was 0.56, meaning differences in diversity in the simulation were roughly half the corresponding empirical differences, on average. The authors therefore warn that intervention effects on diversity may be understated. The model also tends to produce lower engagement and diversity and higher segregation than the observed data, although the fitted differences in engagement and segregation were not statistically distinguishable from a one-to-one relationship.
Simulated-to-observed relationship: fitted slopes
Linear-fit slopes for simulated means against empirical system-level metrics. A slope of 1 would indicate one-to-one scaling.
| Label | Value |
|---|---|
| User engagement | 0.82 |
| Attention diversity | 0.56 |
| Segregation | 1.2 |
One likely reason is a structural limitation: in the model, a user–leader relationship with no retweet in the previous month has zero probability of producing a retweet in the next. A connection can disappear from the simulated network, but cannot newly emerge. Over time, that rule can concentrate attention among established relationships and make audience boundaries look sharper than they are in real life.
With those limits in view, the authors simulated three interventions. The behavioral interventions were prebunking—helping people recognize manipulative techniques before they encounter them—and accuracy prompts, which nudge people to consider whether information is accurate before sharing. The team translated effect sizes from earlier studies into changes in model parameters governing engagement. The third intervention replaced the algorithmically curated feed with a chronological one, changing exposure rather than users’ propensity to retweet.
All three reduced engagement in the simulations. But their effects on the structure of the debate diverged. The two behavioral interventions slightly reduced attention diversity and left segregation essentially unchanged. The chronological feed increased attention diversity and reduced segregation in every month when fully deployed. The model therefore points to an asymmetry: nudging the response to content modestly changes activity, while changing the distribution of exposure can alter who shares an audience with whom.
This is not a claim that chronological feeds eliminate echo chambers, or that prebunking has no value. It is a comparison within this model, for this dataset and these outcomes. The abstract and reported results do not give a single universal effect size for each intervention, and the model’s diversity metric itself is compressed relative to empirical values. The appropriate takeaway is about relative direction and scale in these counterfactual scenarios, not a guaranteed percentage change in a real platform.
The authors also explored whether a recommender could be designed to improve several goals at once. Their proposed intervention increased engagement and attention diversity while reducing segregation. Its gains in diversity and segregation were about one tenth as large as those of the chronological feed. That smaller but multi-objective result matters because platform choices are often framed as a blunt trade-off: either preserve engagement or make the public conversation healthier. The simulation suggests that an intermediate design might improve multiple outcomes, though with more modest structural gains.
Scale of the Twitter dataset
Raw dataset totals reported for the Italian debate, 2018–2022.
| Label | Value |
|---|---|
| Followers | 2,713,443 accounts |
| Leaders | 583 accounts |
| Tweets | 13,996,467 accounts |
| Retweets | 29,131,371 accounts |
Why This Changes Things
Many interventions are evaluated through experiments with individual users. Those studies can answer whether a prompt changes an individual’s willingness to share, or whether a different feed changes what one person sees. They are less equipped to measure how millions of interactions collectively reorganize a public debate: whether audience groups become more distinct, or whether people’s attention is spread across a wider range of sources.
That gap matters because platform systems are relational. A user’s exposure can depend on what similar users engage with; one person’s choice can become part of the signal shaping another person’s feed. An intervention aimed at individual judgment may therefore have limited effects on the network’s overall structure, while a recommender change can alter many users’ exposures at once. The Exposure–Engagement framework puts those two pathways in a common accounting system.
The policy context is immediate. The European Union’s Digital Services Act requires very large online platforms to offer at least one recommender option that is not based on profiling. A chronological feed is one familiar version of that alternative. The study offers a way to ask not only whether such a choice changes a user’s experience, but what it could mean for the shared architecture of a debate.
The comparison also helps explain why individual-level findings can seem incomplete. A field experiment cited by the authors found that switching Facebook and Instagram users to chronological feeds reduced engagement without detectable effects on issue or affective polarization. That result does not contradict this study: the outcomes differ. One measures individuals’ attitudes; the other simulates changes in audience overlap and attention structure. A debate can become less segregated in its audience connections without each participant reporting a changed attitude.
The study’s reach is narrower than “social media” in general. It examines public Twitter retweets in Italy, with influential political and news accounts as the content leaders. Private groups, other platforms, different countries, and other types of interaction may follow different dynamics. The dataset’s raw scale—583 leaders, more than 2.7 million followers, and 29.1 million retweets—gives the model substantial observational material, but scale alone cannot make its assumptions universal.
Scale of the Twitter dataset
Raw dataset totals reported for the Italian debate, 2018–2022.
| Label | Value |
|---|---|
| Followers | 2,713,443 accounts |
| Leaders | 583 accounts |
| Tweets | 13,996,467 accounts |
| Retweets | 29,131,371 accounts |
Its most valuable contribution may be methodological. Interventions tested in different settings can be difficult to compare: one may report accuracy judgments, another sharing intentions, another engagement. Here, the researchers translate those findings into mechanisms in a shared model and assess them against the same three system-level properties. That does not erase differences between the original experiments, but it creates a more consistent basis for asking what might happen to the debate as a whole.
What’s Next
The next step is not to treat simulation as a substitute for experimentation. It is to use it to make experiments more informative. A platform or regulator could use a model like this to identify plausible outcomes, determine which metrics to track, and test how effects might vary as a measure reaches a larger share of users. The paper explicitly models intervention coverage by applying changes to randomly selected fractions of users, although the supplied results do not provide the detailed coverage curves or exact effect sizes.
Several uncertainties deserve direct testing. The simulated chronology intervention changes recommendation, but real platforms also shape what gets posted, what users learn to expect, and whether people stay or leave. The model holds leaders’ content supply to observed activity, so it does not capture those feedbacks. Its one-way rule for user–leader ties also makes it harder to represent people discovering a new source or returning to a dormant one. And the March 2020 exception shows that a system calibrated on recent patterns can struggle during major shocks.
Researchers could strengthen the framework by incorporating new relationships, testing it on other countries and platforms, and comparing simulations with real interventions that measure both individual responses and network-level outcomes. They could also investigate whether the proposed compromise recommender can preserve meaningful engagement under varied conditions, rather than only in the historical Italian debate modeled here. The authors’ results give a reason to ask those questions, not a final answer to them.
The broader lesson is that healthier online debate may depend less on persuading each person to make a better choice in isolation than on how platforms arrange the choices people encounter. Behavioral tools remain important, especially for resisting manipulation and misinformation. But if the goal is to change the collective structure of attention—who hears whom, and whether audiences remain sealed into separate camps—the architecture of exposure deserves equal scrutiny. The model offers a way to explore that possibility before a platform changes the feed for millions.