Grain — Case for Impact

This is the strongest theory of change. The mechanism is direct: Grain replaces an app with infinite algorithmic content with one where your friend feed naturally runs out.

The 31% self-control gap from Allcott et al. (2022) establishes that excess use is real and large. The 78% adoption rate for binding limits shows people want help. The design features driving excess use — infinite scroll, variable reinforcement, algorithmic content injection — are precisely what Grain removes by construction. The deactivation and limit-setting RCTs consistently show that reduced time improves wellbeing by 0.06–0.09 SD, replicated across seven studies and multiple countries.

The gap in the evidence is that no study has tested the specific intervention Grain proposes: a substitute platform that preserves social function while removing extractive design. The closest analog is the chronological feed experiments from Guess et al. (2023), which did reduce time spent. But Grain goes further than chronological ranking — it eliminates non-friend content entirely.

The behavioral economics framework is well-established, the mechanisms are well-documented, and the intervention logic follows directly. The uncertainty is about whether Grain can actually capture the social graph, not about whether the intervention would work if it did.

What this looks like at scale. Suppose 10% of U.S. social media users (~25 million people, per Pew Research) adopted Grain and it displaced half of their time on other platforms. The average American spends about 129 minutes per day on social media, and the Allcott et al. self-control gap suggests roughly 31% of that time is unwanted — about 40 minutes per day that people would prefer not to spend. If Grain’s finite, friend-only design helped users recapture even half of their unwanted time on the displaced portion (~20 minutes per day), that would aggregate to roughly 3 billion hours of recaptured time per year across 25 million users. Using the per-user welfare estimate of $100–150/year from the deactivation and limit-setting literature, that translates to roughly $2.5–3.8 billion per year in aggregate welfare gains. For context, this is comparable to the annual economic value of a mid-sized federal public health program — from a single design change to a single product category.

This might be the second-strongest theory of change after time recapture, and the one most distinctive to Grain.

The CHI 2025 screenshot study found person-to-person communication had the lowest regret of any smartphone activity. The deactivation studies found people socialized more with friends and family when off Facebook. There’s a plausible case that Grain, by centering direct messaging and friend updates rather than content consumption, shifts the ratio of social interaction to passive consumption.

The dating recession evidence is suggestive — in-person time with friends down 50% since 2010, declining confidence in social skills among young adults. Grain can’t fix this directly, but if the platform’s design encourages genuine back-and-forth (responding to friend posts, direct messaging) rather than passive scrolling, it might preserve or strengthen social ties that Instagram’s design lets atrophy. You scroll past your friend’s post, maybe drop a like, but never actually talk to them. Grain’s finite feed and lack of competing algorithmic content might increase the likelihood of genuine interaction.

The evidence gap is that no study has compared relationship quality across platform designs. We know deactivation increases face-to-face socializing, and we know person-to-person communication has low regret, but we don’t know whether a friend-focused platform produces more or better social interaction than a general-purpose one.

This is where the evidence gets genuinely complicated. The population-average effect of social media on mental health is small — Orben and Przybylski’s 0.4% of wellbeing variation, the meta-analysis showing negligible active/passive effect sizes. But the population average obscures substantial heterogeneity.Braghieri et al. (2022) found Facebook introduction increased depression by 7% and anxiety by 20% relative to baseline, and that was pre-algorithmic Facebook — meaning the social comparison dynamics that Grain would still contain were sufficient to cause harm. Facebook’s internal research found 32% of teen girls with body image concerns said Instagram made them worse, though those numbers have methodological issues.

The problem for Grain specifically is that some of the mental health harm comes from social comparison, which is inherent to seeing your friends’ lives — not an artifact of algorithmic amplification. A chronological friend feed might reduce exposure to influencer content and algorithmically surfaced appearance-focused content, but it doesn’t eliminate the “everyone’s life looks better than mine” dynamic. It might even concentrate it, since every post you see is from someone you actually know and compare yourself to.

The strongest case is for adolescent girls, where the evidence (despite being contested) is most suggestive of meaningful harm, and where the mechanism plausibly runs through algorithmic content — explore page, beauty content recommendations, engagement-bait body image content — rather than just friend posts. Grain’s friend-only model would eliminate the algorithmic exposure channel while preserving the social connection channel.

There is a genuine unresolved question about whether social comparison among friends is itself harmful. No study has isolated friend-only social media from algorithmic social media for mental health outcomes.

The headline findings seem supportive: Milli et al. showed engagement algorithms amplify partisanship by 0.24 SD, Jia et al. showed reranking reduced partisan animosity by 2 points. Grain eliminates the algorithmic amplification channel entirely. But the Meta 2020 election experiments are the biggest problem — across four large-scale studies, chronological feeds, reduced like-minded content, and removed reshares produced no significant effects on affective polarization or political attitudes. Nyhan et al. found reducing like-minded content by a third moved nothing.

The reconciliation might be that Grain’s mechanism isn’t about feed ranking within a political content ecosystem — it’s about removing political content exposure altogether. A friend-only feed has very little political content compared to Instagram’s explore page or X’s algorithmic timeline. You’re not reranking political posts, you’re replacing them with photos of your friend’s dog. That’s a different intervention than what the Meta experiments tested.

But Boulianne’s meta-analyses found social media has small positive effects on civic participation, and the deactivation studies showed reduced political knowledge. If Grain pulls people off platforms where they encounter political information, you might reduce polarization while also reducing engagement with democratic processes. That’s a tradeoff, not a pure win.

The direct experimental evidence (Meta 2020 studies) suggests feed-level interventions don’t move polarization. Grain’s mechanism is different — displacement rather than redesign — but the displacement pathway to reduced polarization hasn’t been tested. And the potential cost to civic engagement complicates the story.

Guess et al. (2023) found that removing reshared content on Facebook reduced exposure to untrustworthy news sources. Grain’s friend-only model has no resharing mechanism and no algorithmically surfaced news content, so exposure to misinformation through the platform would be near zero compared to Instagram, Facebook, or X.

But the honest question is whether this matters much in practice. Misinformation exposure on social media is concentrated among a small share of heavy news consumers. Most people’s Instagram feed is already mostly friends, food, and entertainment, not political misinformation. And to the extent that Grain displaces Instagram but people still use X or YouTube for news, you haven’t reduced their misinformation exposure — you’ve just moved it to a different app.

Grain could also create its own misinformation channel through DMs and friend posts. Friend-sourced misinformation (“my friend shared this so it must be credible”) can be more persuasive than algorithmically surfaced content from strangers. WhatsApp’s misinformation problem in India and Brazil is the cautionary example of what happens when misinformation spreads through trusted social graphs.