A story in the New York Times this morning suggests that the National Security Agency has been analyzing our social networks through email and phone call records, apparently accomplishing “large-scale graph analysis on very large sets of communications metadata” of American citizens and foreign citizens alike. This network analysis uses not only contact data but GPS tracking to understand not only how we relate but how we move in relationship to each other.
From the description in the article, the methods that the NSA uses seem to be very similar to those that political science is using, in Michael Ward, Katherine Stovel, and Audrey Sacks’ words, to locate the “holy grail” of “effectively analyzing the interdependence and flows of influence among individuals, groups, and institutions,” a sea-change in the field.
I’m not arguing that we as political scientists have culpability in this (these methods did not originate in our field by any stretch of the imagination). But I am interested – if network analysis does the cool things it does for our work, what does it do for the work of those whose job is to watch and monitor us?
We are getting used to reading network analysis about the networks that we find intellectually interesting in global politics (see, e.g., the discussion of Emilie Hafner-Burton, Miles Kahler, and Alex Montgomery) – where people and groups are relatively positioned and relatively understood. This has been applied to networks of nations (e.g., Zeev Maoz’s book) to the formation of trade agreements (e.g., Mark Manger, Mark Pickup, and Tom Snijders’ article) in ways that arguable radically improve the explanatory power of substantive theories in those areas, and at the very least provide both a very different look and significantly more available information.
So what of social network analyses of which we are the subject, rather than the researcher? If my reading of the utility of network analysis is correct, then intelligence will learn more, faster, and that learning will constitute more useful information about who people are and what they do than simple spying. As an abstract application, then, it seems efficient and appropriate. It is, however, the very non-abstractness of this that is interesting to me: the fact that it very well may be happening to any one of us. To me, as much as I enjoy reading network analysis of social, political, and economic network analysis of phenomena in global politics, I have less than no interest in there ever being network analyses of me. However one is politically disposed towards intelligence gathering, this seems to be a major development in it (or at least in our knowledge of what they do) that requires substantive consideration.
Laura, an entirely uneducated viewpoint: my basic sense here is that network data is only as good as the theories available to interpret it. With cellphone data, one can construct a fairly good *description* of one’s social world: the heterogeneity of one’s ties, the centrality of particular individuals, a particular individual’s degree of isolation / number of strong and weak ties. Without some strong understanding of diffusion and influence, however, it’s not clear how useful this information would be to the government. Obviously, there could be a blanket rule re “talk to a known person of interest” once / be two degrees away from such individuals, but this would still mean most data would be thrown away. Still more since known persons of interest should be smart enough to use burners. Certainly this could be a powerful tool if one already has a suspect, but, if one is just groping in the dark, trying to figure out who the suspects might be, I’m not sure network data will help you too much. Unless you can come up with really good predictive models … and I would think this would become harder — not easier — as the number and diversity of one’s network nodes increase. But I’d be curious to hear what people who actually know network data think.
See finding Paul Revere for a sense of the power of network analysis of metadata. https://kieranhealy.org/blog/archives/2013/06/09/using-metadata-to-find-paul-revere/
Love this! — and good for teaching. I think my point wasn’t clear (and I downplayed my basic understanding: I’ve taken network theory courses, been at network-oriented institutions, and spent more than a fair amount of time with network scholars). Note that Paul Revere emerges as a central figure, perhaps an individual to investigate, because he connected to organizations that we already know have potentially subversive content. One could do the same with cellphone records, provided one has a good sense about which nodes / which clusters of nodes (organizations? geographic areas?) have potentially subversive / criminal content. But many of those who are critical of the NSA seem to be suggesting that the data could be used in a far more expansive way: that it could be used to troll through and identify suspicious relational patterns, largely independent of out-of-source information. I think that’s tricky. Further, even though a *description* of a network can show you who is central / who may potentially have the greatest influence or mobilizing reach, one still needs theories of action and influence in order to assess the probability that any one node in a network is “infected”, to put this gracelessly. When we’re talking about disease spread, as in Peter Bearman’s sexual networks work, we have good understandings about how a type of contact is linked to a particular outcome. But when you’re in the world of mobilization and social influence, even when you’re talking about something that individuals don’t necessarily have incentives to conceal, diffusion of a particular behavior becomes more complicated — and outcomes (e.g., does person A vote if her neighbors do?) can become more difficult to predict. So we need some understanding of how a particular behavior spreads, which nodes are most and least receptive, etc. If we use an overly simple or frictionless model of influence (esp in networks with high connectivity, like a strongly integrated immigrant or religious community, for instance), then the whole network turns red very quickly — and investigation will get too expensive (not to mention costly in terms of rights abuse). There are also questions about network structure and the behaviors the NSA is interested in: does centrality & integration = danger (as the Paul Revere model might suggest) or does isolation (low connectivity) = danger? etc. And the Gould question: how do you deal with the layeredness of networks (multiplexity)? … There’s a lot of power here, but we shouldn’t get carried away about how quickly and substantially this will transform surveillance and investigation. I’m more curious about how they could build up predictive models from the data & how much they / we know about recruitment into criminal and terrorist networks.
Adrienne, I appreciate your response. While my post does not come from the perspective of doing a lot of network analysis, I have read a fair amount of it, and my perspective is not entirely uneducated. I am assuming that, like the intelligence community utilizes math theory to break code, it is not entirely ignorant of the tools of social theory of network analysis that could help with these endeavors. I also think that recent leaks suggest that “persons of interest” are significantly broader than many of us previously expected, and might include, for example, scholars who study terrorism – people who would have no reason to use burners or anything like that. I hope that you are right that this is just shooting in the dark for them – but I thought it was interesting enough to start a conversation about
Absolutely!
Couple of thoughts on this:
* First, there is nothing unusual about the overlap of intelligence analysis/intelligence work and social science, either substantively or methodologically. The two have been working hand-in-hand with each other since the 1940s, if not earlier. See Barry Katz’s book “Foreign Intelligence: Research and Analysis in the Office of Strategic Services, 1942-1945.” Many of the university professors who worked for OSS then either went back to academia after the war or transitioned to the new CIA later in the 1940s. Also see Klaus Knorr’s monograph on “Foreign Intelligence and the Social Sciences” from 1964. Or Richards Heuer’s 1978 book on “Quantitative Approaches to Political Intelligence: the CIA Experience” which was drawn from papers presented on panels at ISA earlier in the 1970s. You can see extensions of this in the recent (2011) work on intelligence analysis produced by the National Academy of Sciences focusing on the overlap between intelligence analysis and the social and behavioural sciences (to include the work of Philip Tetlock, Bruce Bueno de Mesquita, etc).
* Second, as Heuer describes in the introductory chapter to his book, intelligence analysis may be more difficult than aspects of academic research because (1) analysts can’t choose the questions as they are drawn from decisionmaker interests/requirements, (2) time frames and time horizons to achieve an answer to the question are much shorter in the intelligence field, and (3) the purpose of the research is different, focusing on individual events rather than general causes/explanations. With greater difficulty comes higher levels of uncertainty, and therefore additional work on corroboration and confirmation of inferences may be necessary.
* Finally, on this specific question of NSA’s use of social network analysis, I highlight the importance of the “foreign intelligence purpose” that is emphasized in that NYT article. The odds of NSA wanting to conduct a social network analysis of an academic researcher in this area are very, very slim, and would have to be linked to counterterrorism, counterproliferation, or cybersecurity. As for the use of social network analysis in law enforcement intelligence analysis or other domestic kinds of intelligence, check network terms in google scholar and you’ll come up with some interesting articles. As an example, see: Malcolm Sparrow. The Application of Network Analysis to Criminal Intelligence. An Assessment of the Process. Social Networks. 13/3 (1991). 251-274.
I suspect what they’re generally looking for is violations of a certain number of triggers. Essentially they’ll add up the number of flags such as “very short phone calls”, “calls to Somalia” etc and then check how many other flagged people they are connected to.
Basically some version of the credit card fraud detection systems that block your card if you use it in an unexpected way.