Building a better fediverse
I’ve been here since November 2022, and sometimes it can feel like one long succession of roiling debates about possible improvements to Mastodon and wider fediverse. These debates aren’t what most people come to online social media for (so much so, they’ve given Mastodon a bit of a reputation), and they generate anger, upset, and friction.
But they seem unavoidable. By design, the fediverse isn’t a top down project. And even Mastodon, where a few individuals have large influence (directly by developing Mastodon, and indirectly via Mastodon’s large user share in the wider fediverse), is built around an ideal of community-focussed decentralisation, and is a non-profit funded via donation, much of it from users.
That alone would create a need for discussion. The fediverse is about online social media, though, and that creates a need for discussion far beyond what comes from it being a decentralised, community-driven project. This is what this post is about.
Most of the major debates I’ve been here for (and certainly the most difficult ones) involve three different parts:
a) users or groups of users reporting on their experience of the platform
b) users voicing opinions and expressing preferences concerning future developments (based in good part on a), whether these be specific features, or the direction of an entire part of the fediverse (say, Mastodon) as a whole.
c) discussion about how we should actually build stuff (loosely, not just in the technical sense) to respond to whatever comes out of a) and b).
I have little to say about a) —by definition, these are other people’s experiences, not mine, so all I can (and should) do is take note of them and try to understand. I also have little to say in this post about b), other than that it didn’t take long here to see different people, and different groups of people, wanting very different things.
What I do want to talk about is c) and how it features in all of this. Crucial to this is that no single group, let alone a single individual, knows how to build great social media. That is, no single group knows how to do c) —even if everyone managed to come to an agreement on what, collectively, they wanted.
I have three reasons for strongly believing to know this, and want to briefly say something about each one (please bear with me!).
The first is that I started out as a lawyer. That seems far from online social media, but the law is ultimately a system for enabling, guiding, and structuring human activity. To do that individual rules interact and work together. You cannot just write a rule and expect it to solve your problem. It will interact with other rules, and may produce unexpected outcomes elsewhere. Even worse, a legal system isn’t just a bunch of rules, it’s a bunch of rules in a social context. As a consequence, the same rule or set of rules that works amazingly in one society can be a disaster in the next (for example, voting on important issues by referendum works well in Switzerland, but may be ruinous elsewhere). This means you can’t just transplant solutions, it means there will likely be more than one path for solving the same problem, and, ultimately, it is the entire system as a whole, including the social context, that will (or will not) create the desired outcome.
My second reason for believing that no single group knows how to build great social media is that the one thing being a behavioural scientist teaches you is that predicting human behaviour is incredibly hard: you can change one small thing about the context and people will change how they respond. That’s true even when trying to understand one human at a time, not trying to predict the behaviour of an entire complex system of interacting humans (for more on online platforms as complex systems, see here).
My third, and final, reason is that we have, to date, not managed to build “good social media”. Evidence suggests that what we have built has harmed democracy, spread misinformation that has fuelled massacres, extremism, conspiracy theories, and science denial causing death and misery, while further eroding our privacy, alongside the positives it’s had. So we need to try something new. But that means reaching forward and laying the tracks ahead of our moving train.
There is plenty of expertise to help with this (though I think it’s an illusion to believe we know enough to avoid huge amounts of trial and error, let alone guarantee success). Software developers, who ultimately have to create code, know what’s easy to implement and what’s not, and how that rests on interactions between different design choices. Then there are users, whose experience with the software (and other systems in the past) gives direct insight into what does and doesn’t work for them. Beyond that a huge academic literature can be brought to bear: from communication studies, through human-computer interaction, sociology, psychology, political science, economics, network science, media studies, cultural studies, all the way to philosophy (and I’m sure I’ve overlooked some as it’s completely impossible for one individual to read all of it and the literature grows every day). Finally, interacting with that literature, are the efforts of civil society and organisations monitoring online platforms, and the legal frameworks (e.g., the EU DSA, Canada’s Online News Act, the UK’s Online Safety Act) that are emerging worldwide to tackle the problems online social media have created.
Importantly, these are all very different types of knowledge that cannot replace one another. That’s not to say all of it has equal relevance or role. One can build a social media platform without consulting users, and one can certainly build based on user feedback without academic input. But there are many different puzzle pieces now on the table.
To make things even harder, expertise is really hard to identify, because contexts that superficially seem dissimilar may share deeper characteristics (e.g., the idea that one can’t evaluate platform features individually, without considering interactions and wider embedding in social context may receive support from many different fields).
To solve a really difficult problem -one that hasn’t ever been solved before- it seems to me one wants all of the puzzle pieces one can get. And these can’t be acquired by any simple strategy of ‘listen to the experts’, because there is no single, readily recognizable, sufficient, source of expertise.
That brings me back to fediverse debate, and the elements a), b) and c). The first two communicate expertise that is absolutely essential. But, personally, I wish we could find better ways to collectively take note of what people are reporting about their experiences and their preferences, collectively acknowledge those and then, as best possible, everyone together move on to c).
It’s not that debate never moves on. It does, and suggestions get modified and new features are built. But sometimes debate seems stuck on a) and b) in ways that might not ultimately be that helpful (while potentially also quite painful). There’s many reasons for that, but one, I suspect, stems from perceptions of the problem.
There are problems, say dividing up a cake in front of us, where the means are clear (eventually, we’re going to have to slice it). The issue is simply how big everybody’s slice should be. Arguably, our fediverse problem, though, is that we need to figure out how to actually bake a nice cake.
No amount of treating the problem of ‘figuring out how to bake’ as a problem of ‘dividing up the cake’ is going to solve it (even if some ‘cake dividing’ questions might need to be negotiated along the way). And the kinds of communication that work well for one type of problem are probably not those that work for the other.
A few months back, I posted what I thought was an interesting paper on political anger [1]. The author argues that the role of anger requires cultivating multiple, very different, networks. One involves like-minded individuals; here anger can help motivate, shape group identity, and call to action. But in heterogenous groups, expressions of anger can undermine the possibility of empathy, respectful discourse, reason-sensitivity, and coalition-building. The author suggests the (only) way to square that circle and harness the benefits of expressions of anger while avoiding these costs is to create different groups with different styles of communication: homogenous in-groups where anger can be valuable, heterogeneous networks where anger’s costs outweigh benefits so it’s best avoided, networks where one can just learn what political adversaries think by passively listening, and, finally, heterogeneous non-political groups that allow political communities to repair emotional divides by finding connection through shared interests (hobbies, sport, music etc).
It’s this kind of difference that I have in mind but with respect to types of problem. My sense is that ‘dividing up’ versus ‘how to bake’ problems have different criteria of success and each require different input as a result. If we agree on a particular way to divide the cake, that’s it. By contrast, no amount of agreeing that the oven has to be at 30 degrees to bake our cake will make it so.
For debate to be successful, I suspect it needs to honour that difference. The need for widespread debate in the fediverse isn’t going to go away. It’s not just a decentralised community project, it’s about building something we’ve never built before. Solving that problem requires many different perspectives. I don’t know how we can do that, but trying to make those debates even a little bit better seems worth it, and maybe being clearer about the nature of the problem is one of a number of things that could help.
[1] Palmer, A. (2024). How to get angry online… properly: Creating online deliberative systems that harness political anger's power and mitigate its costs. Politics, Philosophy & Economics, 1470594X231222539.
The failure of referendums
By providing binary choices to complex questions, referendums fail to exemplify democracy – instead they embed division and polarise opinionsJohn Hall (Yorkshire Bylines)
How might adding search change Mastodon?
with Michael Maes & Davide Grossi
1. Arguments about adding search to Mastodon
Mastodon presently does not support full text search: it is not possible to search for words that are not accompanied by # hashtag, that is, it is not possible to search for words that have not been intentionally made available for search. Users (particularly those migrating from Twitter) regularly lament this absence, leading for calls to its inclusion.However, the absence of full text search has, to date, been a conscious design choice: For example, Mastodon’s founder and lead developer Eugen Rochko noted in 2017:
“If text search is ever implemented, it should be limited to your home timeline/mentions only. Lack of full-text search on general content is intentional, due to negative social dynamics of it in other networks.”
In keeping with this, the Mastodon Project currently supports only limited functionality search:“Mastodon’s full-text search allows logged in users to find results from their own statuses, their mentions, their favourites, and their bookmarks. It deliberately does not allow searching for arbitrary strings in the entire database.”
At issue are multiple underlying concerns from protecting marginalised groups from intrusion and harassment to consciously anti-viral design (on Mastodon’s anti-viral design more generally, see Thompson, 2022).The absence of unrestricted text search is contested through regular requests for the inclusion of search through posts on Mastodon, specific Github requests for feature changes, but also attempts to simply circumvent these restrictions through the provision of alternative software tools.
Arguments for unrestricted text search typically appeal to individual freedom or individual ‘rights’ as a rationale. In keeping with this individual-focussed perspective, much of the discussion on ameliorating the impact of search focusses on consent. While opt-in to discoverability as a design-choice clearly addresses some concerns about search, it does not address the fundamental issue that “negative social dynamics” (real or imagined) are *system level properties* of the online discourse as a whole. This means they may potentially impact all users in some way, whether they consent or not. This raises multiple questions about the relationship between individuals on the platform, between individuals and the collective, and platform governance. We return to these implications below; the main goal of this blog post, however, is to make such talk of ‘system level properties’ non-mysterious in order to help promote better thinking and discussion about the design of online communication systems. For this, we present a toy example designed to provide some basic insights into online communities as ‘complex systems’ and how they might be altered by a feature such as text search.
2. Online communication as a complex system
Complex systems are systems characterised by large numbers of interactions between their (often simple) components that give rise to emergent properties of the system as a whole that are typically difficult to predict (e.g., Ladyman et al., 2013).Communicating individuals form social networks (such as the sample network seen in the figure below) in which individuals are nodes and their communication paths are represented by links between those nodes. Even though each individual may have direct connections with only a handful of other individuals, the interconnections between individuals may link them, collectively, into a much, much larger network. In this way, individual Mastodon users are linked not only to direct followers, but indirectly to those followers’ followers and so on.
Understanding online communication platforms as complex systems consequently involves trying to understand how information is propagated across such networks and what kind of emergent patterns this may to give rise to. For example, although it is individuals that read, write, and boost posts, we may think of their interactions combining to determine characteristics of the discourse across the network as a whole. Is the discourse predominantly friendly or hostile? Do focal topics that attract widespread interest persist, or is attention fleeting and fragmented? What kinds of community norms govern interactions?
To understand questions like these, researchers make use of computer simulations involving agent based models (ABM). Such models allow one to explore the behaviour of systems in ways that we could never do in real life. Figure 1 shows the interface of a simple model created to illustrate some basic aspects of online communication systems (a link to the model itself and instructions for further, hands on, exploration can be found below).
Fig. 1. Interface of the Agent Based Modelling software NetLogo with a social network.
The simulation creates a simple social network. It then simulates a contagion process across that network. Such models are widely used to study opinion dynamics or the spread of behaviour. They simply appropriate the notion of ‘being infected’ to receiving a message or witnessing a behaviour. Depending on what we are trying to model, it may make sense to think of the process of spread as either a ‘simple’ or a ‘complex contagion’. For a simple contagion, a single exposure is enough for ‘infection’. For a complex contagion, multiple exposures are required. This can be modelled, for example, by setting a threshold: an individual only becomes ‘infected’ once the number of infected neighbours exceeds the specified threshold. This is typically more appropriate than a single contagion for modelling the adoption of behaviours (e.g., Guilbeault et al., 2018).
The threshold value, like the number of agents in the network, or the structure of that network is a ‘parameter’: an attribute of the configuration of the model. The goal of agent based simulations is to gain insight into the patterns that emerge from agent interactions and understand how those patterns depend on the values of the model’s parameters.
For our toy example, let’s now think of the ‘infection’ as something like “anger”. Anger may be socially mediated in as much as others’ angry behaviour may make us angry in turn. So we will interpret the model as capturing the spread of anger via a diffusion process on the network: let’s assume, for example, that at each time step in the model, agents communicate a message and the message content reflects their current state as either not angry (black) or angry (red). The model is set up with an initial number of randomly selected ‘angry’ agents. Our interest then lies with the extent to which anger spreads.
Exploring our simple model reveals that anger will eventually spread to all members of the network over time, regardless of the exact number of initially angry agents, the size of the network, or its structure, if we set the contagion threshold to 1, that is, to a simple contagion.
This is no longer true if we raise the contagion threshold to 2, so that an individual agent becomes ‘angry’ only when 2 of its direct neighbours are themselves ‘angry’. This means that ‘anger’ requires sufficient support among the direct neighbours of an agent: without that support, the contagion will stop in that neighbourhood, and may eventually come to a stop in the network as a whole.
This means also that anything that makes it more likely that multiple neighbours become infected will change the dynamics of spread. This includes increasing the proportion of initially infected agents. But it also includes changes to the structure of the network, for example, increasing the number of neighbours an agent has (i.e., increasing the number of links), or increasing the clustering whereby an agent’s neighbours are themselves neighbours. All of these now matter, and will interact to produce the patterns of spread, making that spread increasingly difficult to predict.
This brings us back to the topic of this blog post: unrestricted text search. How might we think about the effects of adding search? In effect, what search does is that it (dynamically and temporarily) rewires the network. Instead of seeing posts only from those we follow, we (and others!) can now see posts from arbitrary individuals as a function of message content.
So what happens in our toy model if we introduce some limited rewiring in the model as a result of hypothetical ‘search’? On a given time-step there is now a small probability (p=.02) that a randomly selected node receives three additional random links as a result of a ‘search’. All else about the model stays the same. Yet this simple addition dramatically changes the behaviour of the system. The complex contagion (with the familiar threshold = 2) now stands a decent chance of, once again, reaching all agents in the network. Figure 2 below shows the outcome of 1000 runs of the model (each with a different random starting configuration) both with and without this rewiring (‘search’). The x-axis shows the respective proportions of ‘angry’ agents in the population, and the y-axis shows the count of how many simulated networks ended up with that proportion.
Fig. 2 Histogram of 1000 sample runs by condition (‘search’/no ‘search’).
The plot reveals several important features of our model, and with it of complex systems more generally:
- the system’s behaviour varies as a result of random factors (e.g., which agents are randomly selected for initial contagion, and which are randomly selected for ‘search’).
- as result, there is no single answer as to “what happens” (e.g., when we introduce ‘search’). Rather, there is a space of possible outcomes, some of which are more probable than others.
- there are discontinuities in the space of possible outcomes: there are *no* model runs that end in 60% angry. Rather, once a large enough subset is reached, all agents are reached eventually.
- a minor change to the model can lead to very different outcomes: no run without ‘search’ saw spread to all agents.
- the impact of ‘search’ is not restricted to the agents that are rewired. On those runs where all agents eventually become angry, the vast majority end up in a different state than they would have without that intervention.
These points are echoed in Fig. 3 which provides further insight into the relevant mechanisms. Our stylised ‘search’ adds links to the network. To examine the impact of this, the plots show the correlation between final proportion of ‘angry’ agents and the number of final links and the number of ‘searches’ that took place (right hand), with the top panels (green) showing the results of the 1000 ‘no search’ runs, and the bottom panels (blue) showing the results of the 1000 ‘search’ runs.
We can again see clearly how much variability there is even under exactly the same parameter settings. We can see also that increasing numbers of searches (and with that additional links) increase the chances of spread, but that correlation is loose.
Fig. 3. Correlations between final proportion “angry” (‘infected’) and the number of additional links to the initial 200, and with the number of ‘searches’ in the simulation run.
3. Establishing an evidence base?
So what does this simple model tell us? It is nothing like real communication or the real scale of a network such as the one comprising all Mastodon users, and our ‘search’ and ‘anger’ are nothing like real search or real anger.The latter are just a label; we could have equally chosen ‘happiness’, or a wholly uninformative label such as ‘gleeb’. The meaning of the labels we chose to describe the model is exhausted in what they actually represent: state changes in a particular social network model. What we really want to know is what would happen *on Mastodon* if we introduced unrestricted text search.
This toy model cannot tell us that. It can still tell us plenty that is relevant, however. All of the general characteristics 1-5 above apply to complex systems more generally. Making a model more realistic will not generally make these features go away.
This means also that there are limits on what any kind of real-world empirical study could tell us. Even if we could run ‘experiments’ on Mastodon (or some other platform) that allowed us to look at the effects of introducing search, an individual ‘run’ of our real world network remains just a single point in a space of possible outcomes —outcomes that could have been different given random variation.
That space of possible outcomes likely contains non-linearities and phase transitions (Sole et al., 1996). So even if we combine our best methods (experiments, simulations, observational studies) we will likely understand only something about the broad directions in which changing a parameter might push the system. It will remain the case that ‘just a little bit more’ could lead to a qualitatively wholly different outcome. And the interactions of multiple parameters will be even more resistant to our understanding.
While this means that we will not likely ever get a definitive answer on what adding search would do to Mastodon, it does shed light on some of the arguments and intuitions in that debate. First and foremost, the idea of restricting search as part of an anti-viral design stance is plausible, in the sense that it is plausible that search and the degree to which things may go viral are connected. Second, examination of even our toy model as a complex system shows that one cannot conclude that because a bit of something had little to no impact, a bit more will continue to be inoccuous: #hashtag search is already search, but that doesn’t mean that adding more won’t radically transform the system. A bit more can become radically different —so friction matters as a factor dampening the rate at which individuals take certain actions, here and elsewhere. Third, ‘minority actions’ can have global, system wide, effects far beyond those directly involved. That may seem mysterious in the absence of consideration of an actual complex system, but our simple example shows that it is not. This means also that “consent” to having one’s own posts included in unrestricted search does not solve all issues: one can’t consent on behalf of others to effects that *they* will incur.
4. Who decides?
The upshot of all of this is that decisions on system design for online communication platforms seem unlikely to occur in a context with an abundance of evidence that clarifies precisely what effects a particular change to the system would have. This makes all the more important considerations of who gets to decide and on what grounds.It is natural to try and cast these issues in terms of ‘rights’: individual rights of users, rights of those who have invested most in building the platform, and so on. But rights are never limitless, because exercising them typically touches on the rights of others —all the more so when we are dealing with public or collective goods. These raise a host of problems of their own (e.g., Reaume, 1988). And it seems extremely unlikely that those problems will magically disappear or resolve by virtue of decentralisation or federation. So Mastodon —as not just a piece of software, or a development company, but also as a community— might ultimately find itself seeking to develop governance structures to resolve such issues.
References
Guilbeault, D., Becker, J., & Centola, D. (2018). Complex contagions: A decade in review. In Lehmann, S. and Ahn, Y.Y. (eds.) Complex spreading phenomena in social systems: Influence and contagion in real-world social networks, Springer. accessed at: arxiv.org/pdf/1710.07606.pdfLadyman, J., Lambert, J., & Wiesner, K. (2013). What is a complex system?. European Journal for Philosophy of Science, 3, 33-67. link.springer.com/article/10.1…
Réaume, D. (1988). Individuals, Groups, and Rights to Public Goods. The University of Toronto Law Journal, 38(1), 1-27. jstor.org/stable/825760
Solé, R. V., Manrubia Cuevas, S., Luque, B., Delgado, J., & Bascompte, J. (1996). Phase transitions and complex systems: Simple, nonlinear models capture complex systems at the edge of chaos. digital.csic.es/bitstream/1026…
Thompson, C. (2022) Twitter alternative: how Mastodon is designed to be “antiviral”. Medium uxdesign.cc/mastodon-is-antivi…
Appendix: Instructions for model exploration
A link to the model is here. Clicking it will open a version of the model for running within a browser. All Netlogo models have three tabs (see Figure 1 above): the Interface, an Info Tab, and a Code Tab. The Interface Tab lets the user run the model via buttons and to-be-entered parameter values. The Info Tab contains a description of the model, and the Code Tab shows the computer programme itself.Pressing the Setup button will initialise the model, Go Once will execute a single time step, and Go will let the model run until behaviour stabilises so that there is no more change, at which point the run ends.
The contagion threshold is set by the number entered in the “threshold” box. Entering 1 into the “search” box enables search, 0 turns it off (alternatively, pressing the purple Search button will execute ‘search’ once with the pre-defined probability, regardless of setting). The topology parameters N and p determine the size and structure of the network, respectively (the graphs above were all produced with a network of N=100 agents, and a rewiring probability p= .19, giving rise to a so-called small world network; see the Info Tab for more detail).
For more thorough exploration, the model can be downloaded together with NetLogo (which is free), and explored using NetLogo’s inbuilt “BehaviorSpace” which allows one to define experiments involving many runs (for instructions on how to use BehaviorSpace).
Extract error messages to locale file by alpaca-tc · Pull Request #2162 · mastodon/mastodon
localize validation errorsGitHub
