For this instalment of the Security Lancaster Seminar Series, the University of Exeter’s Prof. Mark Levine came in to talk to us about his research examining
the ways in which the social psychology of groups and identities can contribute to the analysis of behaviour in the security domain. The broad thrust of Levine’s research has been
arguing against the idea that groups are bad for our psychological health and that, despite the claims of the bystander effect theory, groups
can be harnessed for pro-social ends. Where current group identity research is focussed on ways to remove member from groups (c.f. anti-radicalisation), the sources of the challenges posed by the combination of group identity and technological advancement are equally the sources of great opportunity.
The traditional view, said Levine, was that the presence of a group reduces individual control, which increases the chances of violence. However, he argued that our understanding of the bystander effect is more nuanced nowadays, adding that it does not appear to hold up in
dangerous or violent emergencies—I thought immediately of the stories of heroism that come out of every recent terror attack, such as Echeverría and his skateboard. Part of the issue, suggested Levine, is that it is very difficult to study violence directly,
in the raw. However, through analyses of CCTV footage we can begin to analyse the behaviour of bystanders in real life situations, which is what Levine and his team have been doing.
Across 219 clips of CCTV footage, they found that the rate of at least one person intervening in an incident was around 90%. They then began to plot the
trajectory of violence, or a decision tree of escalatory and de-escalatory acts, in order to try and determine how many acts it would take before you could reasonably predict the outcome of the dispute. They found that despite the traditional view of larger groups leading to more antisocial behaviour and less pro-social, bigger groups actually led to no increase in antisocial behaviour, but a large increase in pro-social. Examining the sequences, they found that the outcome of an incident could be predicted from the third act. They then began investigation how the composition of intervening actors affected this, finding that three interventions by the same person were the most likely to result in an anti-social outcome, whilst three interventions by three separate people was the most conducive to a pro-social outcome. This echoes a lot of what I remember from my door supervision course, which emphasised disengaging from customers and passing them on to a colleague as a de-escalation tactic.
Citing the work of Flack et al. and Krakauer et al., Levine divided interveners into
pacifiers who deal with one party and
policers who deal with all. Across 43 clips, they found 119 pacifiers and 64 policers. Through hierarchical multiple regression analysis they examined the impact on the incident outcome of the number of bystanders, followed by the number of pacifiers, followed by the number of policers, finding that the first and third predict violence well, whilst the second has little impact.
Levine then detailed a series of VR experiments he had been running in which Arsenal-supporting participants in a virtual bar began chatting to a fellow Arsenal fan, before a non-Arsenal fan at the bar picked a fight with the avatar. The video he showed was a bit like a far more aggy version of Façade. He found that the presence or lack of a crowd of onlookers in the bar had an impact on people’s likelihood to intervene, but was surprised to find that marking the bystanders as fellow Arsenal fans inhibited interventions, whilst out-group bystanders maximised it.
Finally, Levine detailed his research into detecting social identities through textual analysis, based on the theory that different contexts would make the characteristics of different identities more salient. Taking posts from users of the Mumsnet Parenting (12,700 users) and Feminism (10,000 users) boards, as well as users of both (3,000), he conducted LIWC analysis on 7 identified categories that he expected to differentiate users of each the most, such as
third- or first-person pronoun use and
inclusive language. He found that his classifier could accurately classify 70–77% of posts, or 3 our of 4, and effectiveness required as few as 25 words per post. To test the classifier experimentally, he then got 45 participants who identified as both parents and feminists and presented them with material to make one or the other identity more salient. Each wrote essays on three topics—one parenting-related, one feminism-related and one neutral. Levine found that his classifier was less accurate here, but still managed a 68% rate on the neutral essays.
Later research from Levine’s team has focused on detecting
business(wo)men on the Silk Road forum, and on clearnet sources such as Reddit, with a darknet success rate of 68–76% and a clearnet success rate of 62–68%. All of these results seem to suggest that, to some extent, individuals do shift their styles of writing when different social identities become more salient than others, and that we can predict to some degree their identity based solely on language use.