Study says LinkedIn's A/B testing of its People You May Know algorithm, involving 20M users from 2015 to 2019, may have affected some users' livelihoods
A study that looked back at those tests found that relatively weak social connections were more helpful in finding jobs than stronger social ties. Tweets: @taylorlorenz , @joshelman , @martinsfp , @caseynewton , @taylorlorenz , @cfchabris , @eric_seufert , @michaelzimmer , @jeffjarvis , @moonalice , @rmac18 , @natashanyt , @natashanyt , and @fabiochiusi Tweets: Taylor Lorenz / @taylorlorenz : These tests run by LinkedIn are indicative of how such tweaks to widely used algorithms can become social engineering experiments with potentially life-altering consequences for many people. https://www.nytimes.com/... Josh Elman / @joshelman : There are so many ways to interpret this story. The most optimistic one is that LinkedIn has helped many people get new jobs and advance their careers and lives. And they continue to try to make it better and learned an important new way to help people https://www.nytimes.com/... @martinsfp : It's easy to stay this article is over-egging the impact a bit, but it certainly highlights that second-order impacts of A/B tests on social platforms should be considered more. https://t.co/dXKJpOYJXd Casey Newton / @caseynewton : LinkedIn also ran an experiment on me. It sent me 48,000 emails designed to get me to visit LinkedIn https://www.nytimes.com/... Taylor Lorenz / @taylorlorenz : “The findings suggest that some users had better access to job opportunities or a meaningful difference in access to job opportunities. These are the kind of long-term consequences that need to be contemplated...” https://www.nytimes.com/... Christopher Chabris / @cfchabris : In other news, LinkedIn ran thousands of experiments by constantly changing features of their site & product without scientifically testing them, which affected people's lives and made money for the company, but we know nothing about any of their effects https://www.nytimes.com/... Eric Seufert / @eric_seufert : I'm a bit alarmed at seeing a routine A/B test regimen characterized as a “social experiment” for which explicit, opt-in consent should have been obtained. https://www.nytimes.com/... Michael Zimmer / @michaelzimmer : LinkedIn ran experiments on more than 20 million users over five years. I comment on the ethical consequences of this type of big data research in this great article by @natashanyt in the @nytimes. https://www.nytimes.com/... Jeff Jarvis / @jeffjarvis : Demonizing A/B testing. Without this “social experiment” (loaded phrase) they would not know how to improve prospects for all. LinkedIn Ran Social Experiments On 20 Million Users Over Five Years https://www.nytimes.com/... Roger McNamee / @moonalice : This story is a reminder that Microsoft is major player in surveillance capitalism (they harvest data from childhood to adulthood: Minecraft>Xbox>LinkedIn> search). Now there is evidence they experimented on 20 mm LinkedIn users, potentially affecting their careers. https://twitter.com/... Ryan Mac / @rmac18 : This shady experiment run on LinkedIn using the “People You May Know” algorithm reminds me of a story I tried to report involving famous accounts that I could never prove out, but always wondered about... https://www.nytimes.com/... Natasha Singer / @natashanyt : A fascinating new study in Science details how LinkedIn ran social experiments on 20 million users over 5 years. It shows how algorithmic tweaks can become social engineering experiments with potentially life-altering consequences for millions of people. https://www.nytimes.com/... Natasha Singer / @natashanyt : Remember how Facebook experimented on 700K users for a weeklong study looking at how people's emotions spread on the platform? LinkedIn experimented on 20 million users over 5 years by tweaking their People You May Know suggestions. https://www.nytimes.com/... Fabio Chiusi / @fabiochiusi : “LinkedIn ran experiments on more than 20 million users over five years that, while intended to improve how the platform worked for members, could have affected some people's livelihoods, according to a new study” https://www.nytimes.com/...
Context & Ripple Effects
The study turns LinkedIn’s connection recommendations into an employment-market question: its finding that weaker ties were more useful for job discovery means ranking changes can alter access to professionally valuable networks. A parallel set of Meta-collaborative studies found Facebook’s algorithm influential without necessarily changing beliefs, underscoring that algorithmic effects depend on the outcome being measured.
First-order effects
- LinkedIn users assigned different People You May Know variants received different mixes of potential connections; because weaker ties were more helpful for jobs, those recommendation choices may have changed some users’ access to job-relevant contacts.
- LinkedIn’s product and research teams face a clearer trade-off: optimizing connection recommendations is not only an engagement decision when the network is used for career mobility.
Second-order effects
- Researchers and critics gain a concrete basis to scrutinize employment-related outcomes from large-scale platform experiments, rather than treating recommendation tests as operationally neutral.
- LinkedIn may need to evaluate connection-ranking changes against labor-market outcomes alongside conventional product metrics, since the study ties the value of weak ties to job finding.
Third-order effects
- As professional platforms mediate introductions at scale, experimentation governance is likely to be judged by distributional effects on opportunity, not solely by whether an algorithm shifts attitudes or usage.
- The pattern points toward greater demand for independent research access into how recommendation systems allocate economically valuable connections, though the supplied coverage does not establish a policy response.
The trend: Recommendation systems are becoming an increasingly consequential layer of labor-market access as platforms determine which social ties users are prompted to form.