On the first of march 2018, Jack Dorsey published a manifesto setting out Twitter’s plans for improving the site.
At the centre of this thread are worries that incivility, disagreement, and closed-mindedness undermine healthy collective conversation (see also Dorsey’s testimony to the United States Committee on Energy and Commerce https://docs.house.gov/meetings/IF/IF00/20180905/108642/HHRG-115-IF00-Wstate-DorseyJ-20180905.pdf)
Dorsey sets out four metrics for conversational health, taken from cortico (a non-profit based at the MiT Media Lab).
- Shared Attention: Is there overlap in what we are talking about?
- Shared Reality: Are we using the same facts?
- Variety: Are we exposed to different opinions grounded in shared reality?
- Receptivity: Are we open, civil, and listening to different opinions?
In this post, I want to raise some worries about the way that Twitter and others are framing questions about the quality of discourse on social media sites, focusing on the idea of conversational health, how to think about the quality of conversations, and the extent to which we should care about good conversations.
Let’s start with the health metaphor. As Rachel Fraser persuasively argues, metaphors are not epistemically innocent. A metaphorical statement can activate a network of inferential connections and ideological commitments from the source domain and project them onto the target domain. Using health metaphors to talk about conversation sneaks in a number of ideas we might find troublesome. First, worrying about the health of a given chunk of discourse encourages us to look for some disease that is the underlying cause of our problems. It is not too hard to see terms like ‘fake news’, ‘polarisation’, and ‘echo chambers’ as picking out diseases of discourse. We often find environmental metaphors playing similar roles: it is common to talk about information pollution, which encourages us to look for polluters. Secondly, talking health encourages an individualist mindset, according to which the quality of a discourse is the responsibility of the participants in that discourse. When heard as public advice ‘stop eating fast food!’ and ‘Stop sharing fake news!’ play related roles, transforming complex public problems into issues of individual responsibility. In light of these considerations, I suggest that we’d be better off sticking with talking about the quality of conversation, keeping a clean sheet of associations.
In order to think about the quality of conversation it is helpful to start by asking what a conversation is. Linguists and philosophers of language have together produced a large body of theory that addresses this question. There’s a lot of relevant work here, but I’ll focus on the account of conversation and its aims offered by Robert Stalnaker and Craige Roberts.
The Stalnaker-Roberts model of conversation starts from the insight that conversation is a kind of co-operative activity: the activity of answering a question together. We typically start a conversation by asking a question which fixes what the conversation is about: what Roberts calls the question under discussion. We can think of this question as a set of possible answers, and of the goal of conversation as being to reach the correct answer. Progress through a conversation then involves attempts to answer the question under discussion, either by asserting relevant facts, or proposing other relevant questions. We do not start from scratch: we can usually assume that our interlocutors are taking certain propositions for granted. The set of propositions which all speakers taken for granted make up what Stalnaker calls the common ground of the conversation.
Exercising a degree of charity, we might think that metrics of shared attention and shared reality are getting at these features of conversation. It is a precondition of having a conversation that there is a shared question under discussion (shared attention), and that there is a fairly rich common ground of accepted claims (shared reality). When we fail to co-ordinate on either count, conversation breaks down, and we wind up talking at cross-purposes, or being unable to understand one another.
Although the Stalnaker-Roberts model is intended as a descriptive account, we could turn these features of conversations into metrics for conversational quality. One could measure the degree of agreement between two speakers by counting the number of propositions in their common ground, and one could measure cross-talking by considering the differences between the questions which speakers are addressing. Implementing these ideas would take a considerable amount of work, but it might provide us with measures of conversational quality that are based in an independently plausibly account of conversation.
A couple of observations about this idea.
The degree of agreement and co-ordination we will find will depend on what size of discourse we are looking at. If we were to look at repeated Twitter interactions between two agents, we would expect a rich conversational background and shared questions. If we were to take a sample of tweets by several hundred users, we would expect to find people talking about different things, and taking very different claims for granted. Some people want to talk about politics, but others are quite happy talking about Dr Who trivia, or ultra-running. This doesn’t impugn the quality of their collective discourse.
The common ground and questions under discussion in a conversation cannot be simply read off from what speakers say. To understand what speakers are taking for granted we need to understand presupposition triggers, relevant facts about the speakers, and their previous conversations.
Most importantly, it isn’t obvious that good conversations are always the kinds of things we should be aiming for. A good conversation is one that fulfils the constitutive standards of conversation, which I’ve suggested includes having co-ordination of a question under discussion and common ground. But there might good reasons to flout the constitutive norms of an activity. Moving a bishop diagonally is a constitutive norm of chess, but there might be good reasons to let a child flout this rule to forestall a tantrum. The constitutive values of conversation can also come into conflict with our other commitments. Having a good conversation with a white supremacist might require taking on a substantive number of her commitments. This would lead to a good conversation, but would be all things considered a bad thing to do.
Bad conversations might be a good thing for intellectual progress. José Medina argues that discussions between people with different theoretical backgrounds can lead to beneficial epistemic friction, helping both sides of the conversation to question their background assumptions, and understand the gaps in their perspectives. This kind of productive disagreement seems to be what cortico’s principles variety and receptivity are getting at.
In closing I want to note a puzzle. It looks like the desire for good conversations (understood along the lines above), and the desire for productive disagreements pull in different directions. We can either have good conversations with interlocutors that have shared background, or bad but productive conversations with interlocutors who do not have a shared background. I would suggest that cortico’s metrics of shared attention and shared reality will score highly in the first kind of case, but low in the second case, and vice versa for variety and receptivity. We can have good conversations, and we can have productive conversations, but plausibly not at the same time.
Camille Pissarro, Conversation (Google Art Project)