Deleting the misinformation didn't move the needle — because the network was already sorted
This cycle's throughline is that network structure and network position are already doing more explanatory work than the content flowing through them. The largest live-feed randomized experiment yet run — removing twice-flagged 'untrustworthy' sources from roughly 16,000 consenting Facebook and Instagram users for three months around the 2020 US election — found exposure rare on average but sharply concentrated in a minority of users, and found that removing it moved no measured belief, polarization or trust outcome. A companion empirical result helps explain why: a bounded-confidence homophily analysis of Reddit and Twitter/X data shows opinion neighborhoods are already sorted well beyond chance before any algorithm intervenes. Elsewhere, a landmark unified fruit-fly connectome shows biological neural control itself is distributed rather than centralized, a Nature Human Behaviour methods paper offers a new way to measure individual behavior-change thresholds directly from choice data, a large interdisciplinary review argues biological and social contagions should never be modeled in isolation, a two-decade US cohort finds friendship predicts later mood but does not buffer physical decline, and a computational study finds LLM agents left to interact spontaneously reproduce human network-formation biases. One honesty note up front: several of this cycle's strongest items are three to eleven months old rather than from the past two weeks — network science, unlike attention neuroscience, does not run on a dense weekly news cycle, and this digest would rather wait for verifiable, well-designed work than manufacture false recency.
share of the median Facebook user's 2020 feed from twice-flagged 'untrustworthy' sources — randomly removing it entirely (n≈16,000) produced no measurable shift in beliefs, polarization, or trust in media (Science Advances, 2026)
Where the human evidence is
Count by statusEvidence-stage map
Stage × promise (1–10)Network domain impact
Strength 0–10Why deleting misinformation from a feed didn't change what people believed
MechanismTwo of this edition's items fit together into one mechanism: opinion-sorted networks (Communications Physics, 2026) mean most 'untrustworthy' content exposure is already concentrated in a small, self-selected slice of users, so removing that slice in a randomized experiment (Science Advances, 2026) touched too little of what actually shapes belief to move it.
Complex contagion, thresholds and how spread is modeled
Individual thresholds for 'when I'll change my mind' can now be measured from ordinary choice data
Tănase, Algesheimer and Mariani adapt discrete-choice modeling — normally used to study consumer decisions — to estimate individual adoption thresholds (how much reinforcement from others a person needs before adopting a behavior or belief) directly from choice data rather than inferring them indirectly. A companion piece argues this lets social-change interventions target people closer to their actual tipping point instead of guessing at network-wide averages. This is a modeling and measurement contribution validated against existing behavioral datasets, not a new controlled experiment on human subjects — its value is a better instrument for studying complex contagion, not a new empirical effect.
Read the paperTreating each 'contagion' — biological or social — in isolation is the wrong model, a field-wide review argues
Hébert-Dufresne and a large international author list argue that studying pathogens, beliefs, behaviors and stories as independent, non-interacting contagions is a foundational error: these processes constantly interact within people (immune systems, prior beliefs) and across shared networks, producing discontinuous jumps and faster-than-expected spread that single-contagion models miss. They show mathematically that 'interacting contagions' can reproduce the same social-reinforcement patterns attributed to complex-contagion theory, blurring the usual biological/social divide. This is a perspective and synthesis piece, not new primary data — read it as a challenge to how most network-science and public-health models are built, not as a new empirical claim to act on.
Read the paperDon't expect a 'clean up the feed' intervention alone to change entrenched beliefs. The largest randomized test yet of removing flagged misinformation from live Facebook/Instagram feeds (n≈16,000, Science Advances 2026) produced no measurable shift in polarization or trust in media — plausibly because exposure was already a small, self-selected slice of what shapes belief.
If you're trying to understand or shift a social network's opinions, assume the network has already sorted itself by opinion similarity before any algorithm or intervention touches it. Bounded-confidence homophily analysis found opinion neighborhoods on Reddit and Twitter/X concentrated well beyond chance, more so for stronger ties (Communications Physics, 2026).
Treat 'the algorithm is polarizing us' as an incomplete diagnosis, not a settled mechanism. Misinformation exposure on major platforms was rare on average but highly concentrated in a minority of users, and experimentally removing it didn't move beliefs — the more interesting causal question is what makes that 23% self-select into high exposure in the first place.
When modeling how a health habit, norm or product will spread through a group, expect diffusion to hinge on individual adoption thresholds — how much reinforcement each person specifically needs — rather than on overall network connectivity alone; new methods can estimate these thresholds directly from choice data (Nature Human Behaviour, 2026).
Don't model contagions — biological or social — in isolation. Prior beliefs, existing infections and social behaviors interact, and a field-wide review argues single-contagion models routinely miss faster-than-expected or slower-than-expected real-world spread as a result (npj Complexity, Hébert-Dufresne et al., 2025).