Modest AI recommendation boosts diversity, new study of 18,076 users finds

A new paper from Dini Wang and Ho-Chun Herbert Chang examines how algorithmic recommendation shapes information diversity on social media. Analyzing 18,076 users active from 2014 to 2018, they found that after platforms introduced algorithmic ranking in 2016, the topical diversity of shared content first increased and then plateaued, while inequality across users emerged. To explain this, they built a hybrid human-AI diffusion model where exposure is a mix of social-network propagation and algorithmic recommendation.

Simulations and qualitative analysis reveal a non-monotonic effect: modest algorithmic mediation—around a 50% mix—actually raises average diversity and reduces inequality compared to a purely network-driven baseline. But strong mediation reverses the gains, narrowing diversity and concentrating it among a few users. Fitting the model to four years of data shows mediation share grew from zero pre-2016 to approximately 0.50 by 2018—a level that stays within the diversity-enhancing range while exceeding the point of equal exposure. These results provide a unified framework for designing algorithms that broaden rather than narrow what users see.