Social media feeds 'misaligned' when viewed through AI safety framework (www.foommagazine.org)

Researchers from the University of Michigan, Stanford and MIT published a study showing Twitter/X’s timeline is “misaligned” with users’ stated values: the platform’s engagement-optimized ranking systematically amplifies posts scoring high on individualistic values such as stimulation, dominance and hedonism while downplaying collective values like caring and universal concern. Methodologically, the team recruited users, captured ~4,600 timeline posts via a browser extension, and used a set of 19 psychologist-validated “basic human values.” They applied GPT-4o to score each post on those values (validated against human annotators) and constructed value-optimized feeds per user. Value-optimized feeds were perceptibly different to users and statistically uncorrelated with the engagement-optimized ordering X disclosed in 2023, demonstrating that existing ML tools can readily prioritize human values but aren’t being used to. For the AI/ML community this is a proof-of-concept with broad implications: recommender systems—built on similar ML primitives as modern language models—can be technically aligned to human values, yet commercial incentives favor engagement. That gap weakens arguments that companies will voluntarily align more powerful models (or future AGI) if they won’t align easier, high-impact systems like feeds. The paper therefore reframes alignment debates: it connects everyday socio-technical harms of recommender algorithms to broader AI-safety concerns and stresses a need for governance, transparency, and value-aware optimization methods.