What we're studying

Our flagship report, The Pattern Beneath the Symptoms, puts decades of data from fields that rarely compare notes on one timeline. Three measures sum up the direction.

None of these has a single cause, and the report says so. The problem isn’t technology itself. It’s a business model that profits from attention, division and data, and that now sits between people and most of what they read, buy and believe. One finding shapes this study more than any other. In a preregistered experiment with 1,256 people, re-ranking hostile posts on a live platform shifted hostility towards the other political side by roughly three years’ worth of national change, in ten days, and in both directions (Piccardi et al., Science, 2025). How a system ranks what people see isn’t neutral. Design sets the outcome. Not every study agrees on the size of the effect: experiments that switched users to chronological feeds found no change in attitudes (Guess et al., 2023). That disagreement is part of what we’re here to test.

The same goes for data. The business model most platforms run on collects behaviour beyond what the service needs and turns it into predictions to sell (Zuboff, 2019). And being watched changes behaviour: even knowing you might be observed makes people less willing to say what they think (Stoycheff, 2016).

Depth over reach. Instead of maximising how many people see what, Us limits it: a small set of daily matches that expire, identity revealed gradually as both sides engage, connections that fade if no one keeps them. We’ll look at whether members return because something is worth returning to, and whether the connections that form last.

Trust that’s earned. Standing on algn comes only from taking part. Credits can buy more use, never more influence, and governance rules are published as they’re set. We’ll look at whether trust between members, and in the platform, holds up as the community grows.

Nothing to harvest. Memory stays on the member’s device. Cloud processing keeps nothing. Matching uses numbers that can’t be traced back to words. We’ll look at whether people speak more freely when there’s no one to perform for, and whether the model pays for itself without selling anything.

AI pointed at the mind. Sera is a counterpart, not a companion: it’s built to help you see your own patterns and to challenge you honestly, not to agree with you. AI systems tend to affirm users far more readily than people do, and users rate that agreement highly in the moment (Cheng et al.). Then Us turns that self-understanding outward, towards other people. We’ll look at whether it leaves people more capable and more connected, or just more occupied.

This is the part most companies leave out.

There is no outcome data yet. The study begins with launch. The research sets out the reasoning; only members’ experience can show whether it holds.

Depth may not survive growth. Every platform that started small and thoughtful has faced the same pressure as it grew. Growing at a deliberate pace buys time to find out; it doesn’t guarantee the answer.

Trust is slow. A decades-long decline won’t reverse in a year or two. What we can measure early are the conditions trust depends on, not trust at scale.

Intelligence can get in the way. The people most drawn to algn are often the best at building convincing explanations of their own patterns, explanations that feel like insight and aren’t. We haven’t solved this.

Deep patterns change slowly. In a well-known study of intensive human support, fewer than half of the most deeply rooted patterns had shifted after 52 sessions (Kopta et al., 1994). A product shouldn’t promise faster.

Voice may only matter once there’s depth. The largest controlled trial so far, 981 people over four weeks, found no overall difference between talking and typing to an AI (Fang et al., 2025). Four weeks is too short to test what we’re testing, but the finding stands.

Three obligations we haven’t fully met, set out in What Builders Owe.

Continuity. People who build real depth with algn are owed something if the service changes significantly or ends. Beyond notice and time to export, we haven’t yet designed what that is.

Known limits, not just intentions. Saying what algn is designed to do isn’t enough. Members also deserve to know where it’s likely to fall short. We’ve started in the AI & Safety Notice: where agreement can slip through, where tone is misread, where distress might be missed.

The cold start. Members invest trust before they can judge whether it’s paying off. Onboarding should be clear about what that investment looks like, and what happens if the relationship ends early.

The method is simple to state and hard to do: grow at a deliberate pace, measure the things that matter (culture, trust, whether depth holds as the community grows, how the architecture copes under load), and publish what we find, including the parts that don’t flatter us.

Only members who opt in are part of the study. The flagship report brings the wider findings together, and this page will be updated as results come in.