The AI Journal

In 1970, Mark Granovetter surveyed 282 professional, technical and managerial workers in Newton, Massachusetts who had recently changed jobs through a personal contact, and asked each one a single retrospective question: how often, before the job came up, had they seen the person who told them about it? Most answered rarely. From that survey came one of the most cited claims in the social sciences: that acquaintances one sees only occasionally, not close friends and family, are disproportionately responsible for passing along the information that leads to a new job, because an occasional contact is more likely to belong to a different cluster of people than one’s own, and so to know something one’s close circle does not.

The claim travelled well beyond sociology, into management theory, career advice and the design of professional networking sites, precisely because it was counterintuitive: it told people that the friend they saw every week was, on this one specific margin, less useful to them than the colleague from a previous job they barely kept in touch with. The claim was elegant, structural, and for decades essentially untestable in the way the sciences usually mean the word. Granovetter’s own evidence was a retrospective survey of people who had already found a job through a contact, asked to recall after the fact how close that contact was, a design exposed at every step to memory and rationalisation: a person secure in their new position has every reason to remember the lucky tip from a near-stranger as more remarkable, and more causally decisive, than an unremarkable word from a friend. No one could randomly assign which of a person’s relationships counted as strong, so every study that followed, for and against, inherited the same structural limit. The literature grew for fifty years without producing a single result that could rule out the alternative explanation that people with more weak ties were simply the kind of people, better networked and more mobile to begin with, who would have found a new job quickly regardless of who told them about it.

The largest attempt to settle the question with real behavioural data, rather than recollection, came down on the opposite side from Granovetter. Burke and Kraut, working from internal Facebook data and a survey of thousands of users who had lost a job, found that contact with strong ties, not weak ones, predicted finding new work within three months, with people who heard more often from close friends roughly twice as likely to be re-employed in that window. The finding was widely reported as a direct rejection of the weak-tie hypothesis, and it had a real advantage the original survey lacked: it measured actual communication frequency from platform logs rather than asking people to reconstruct it afterward. What it could not do, because nothing about its design was experimental, was separate the effect of tie strength from the kind of person who maintains many close ties and also recovers quickly from job loss for unrelated reasons. Fifty years in, the dispute had better data on both sides and no way to adjudicate between them.

The adjudication finally arrived from a source with no stake in either theory. Between 2015 and 2019, LinkedIn varied its “People You May Know” algorithm so that different users, for ordinary product-testing reasons unrelated to sociology, were shown friend suggestions skewed toward stronger or weaker ties than they would otherwise have received. Rajkumar, Saint-Jacques, Bojinov, Brynjolfsson and Aral recognised that this variation amounted to a randomised experiment on tie strength at a scale no social scientist could otherwise fund or ethically run: the algorithmic treatment, not any property of the people receiving it, decided how weak or strong their newly formed connections tended to be. Across more than twenty million users, two billion new ties and six hundred thousand resulting new jobs, they could finally ask Granovetter’s question with the confounding stripped out.

The answer was neither side’s. Tie strength did not relate to job mobility in a straight line in either direction; it rose with weakness and then fell, an inverted U in which moderately weak ties produced the most job transitions, while the very weakest ties produced fewer than the moderate ones, and strong ties fewer still. Granovetter’s basic direction was vindicated against Burke and Kraut’s finding, in the sense that weakening a tie, from strong to moderate, measurably helped. But the theory as stated for fifty years had no room in it for an upper limit, no account of why a tie weak enough should start helping less rather than more. A connection so faint it barely registers carries almost no information at all, strong or weak; past the midpoint of the curve, weakness stops buying access to a different cluster and starts simply buying ignorance of whether the other person has anything relevant to say. The study’s further finding, that weak ties mattered more in digitally intensive industries and strong ties more in less digital ones, points to a mechanism, how much a sector’s hiring already runs through formal, searchable channels, that neither the original theory nor its harshest critic had built into their account at all.

The objection worth taking seriously is that the experiment answers a different question from the one Granovetter asked, however much it resembles it. His 1973 survey measured frequency of face-to-face contact among Boston-area professionals describing what they considered a meaningful relationship; the LinkedIn study measures messaging frequency and mutual-connection counts on a professional networking platform, and its outcome is a self-reported field change in an online profile, not a verified new job secured through a specific conversation. A platform-native “weak tie,” formed by an algorithm’s suggestion and maintained by the occasional message, is not obviously the same relationship as the semi-forgotten acquaintance who mentions an opening at a dinner party. Where the two map cleanly onto each other is not obvious, and the study’s authors themselves distinguish interaction intensity from structural bridging as separate, imperfectly correlated measures of the thing Granovetter had only one word for.

That gap narrows the claim rather than voiding it. What the experiment establishes is not that Granovetter’s 1973 relationship was exactly reproduced and exactly confirmed; it is that the first manipulation of tie strength ever run at any scale found a real, non-monotonic causal effect in roughly the direction his theory predicted, which is more than either the original survey or its most forceful behavioural rebuttal could establish on their own evidence. Fifty years of argument conducted entirely in correlational data had already produced confident claims in both directions; the one controlled result available says both camps were describing a real effect and getting its shape wrong, and that the shape, not the sign, is now the question the next fifty years of work on networks and hiring has to answer.