Algorithmic Cultivation: How Social Media Feeds Shape User Language

The study measures not just what users see, but how their written language shifts in response. The quasi-experimental design (comparing curated feed users against controls) moves beyond observational correlation to isolate algorithmic causality, which is harder to claim than most platform studies attempt.

This connects directly to the ConsumerSimBench work from earlier this week, which exposed gaps between LLM fluency and actual behavioral fidelity. Here, researchers are documenting the inverse problem: how platforms shape real human behavior at scale. Both papers share a methodological rigor (granular measurement, control groups, verifiable criteria) that contrasts sharply with the PARALLAX finding that many benchmarks leak ground truth. If language models trained on social data inherit algorithmic biases (as this paper suggests), then understanding those biases at the source becomes critical for anyone building RAG systems or consumer-facing LLM applications.

If the researchers release per-feed linguistic fingerprints (Science feed users adopt different semantic patterns than News feed users), that would confirm feed design functions as a latent training signal. If the effect sizes persist when controlling for user self-selection into feed types, that strengthens the causal claim. Watch whether major language model papers cite this in their data curation sections within the next 6 months.