The strength of weak ties, fifty years on

Mark Granovetter published an article that became one of the most cited in the social sciences. His reasoning is structural: close contacts mix with the same people as we do and therefore hold the same information. What they know, we often already know.

Weak ties, by contrast, connect groups that do not overlap. They are what circulates new information, which Granovetter documented through a survey of managers who had found a post through a contact, later developed in his book Getting a Job, published in 1974.

The argument has a particular status: it says nothing about the emotional value of ties, only about their position in the structure. A weak tie is not a lower-quality tie but a bridge, and it is that bridging function that gives it its strength.

The sample was small and local, made up of professionals in a Boston suburb. The share of posts obtained through infrequent contacts was higher there than through frequent ones, a result all the more striking because it ran against intuition.

That result did not establish a causal effect: people with many weak ties also differ in social position and mobility. Later work, notably Nan Lin's on social resources, showed moreover that the position of the contact mattered as much as the weakness of the tie.

A team led by Karthik Rajkumar, Guillaume Saint-Jacques, Iavor Bojinov, Erik Brynjolfsson and Sinan Aral published in Science an analysis using experiments run by a professional network on its recommendation algorithms.

Those experiments, covering tens of millions of users, altered the type of contacts suggested, which produces a quasi-random variation in network composition and allows causal estimation. That is what had been missing for fifty years: a setting in which the weakness of ties is not chosen by the people themselves.

Moderately weak ties produce the most job moves, more than strong ties and more than very weak ties. The relationship is therefore not monotonic: it is the intermediate ties that carry the effect, probably because they combine a little trust with a little novelty.

The authors also note that the effect varies by sector, and is more marked in occupations where information circulates digitally. That heterogeneity is a reminder that the 1973 argument describes a mechanism whose reach depends on the market in question.

The study concerns one particular platform, in a skilled professional context, and measures job moves observed on that platform. Its reach beyond that perimeter is not established, in particular for low-skilled jobs or markets where hiring runs through other channels.

It also prompted an ethical discussion about algorithmic experimentation without explicit consent, a question distinct from the validity of the result but which occupied part of the debate.