It’s an idea that is expansive in its simplicity: once you hear it, you see the threads everywhere. It goes like this: when a new technology lands, the payoff doesn’t simply appear in its wake. The promised gains — whether productivity, inclusion, or innovations unlocked by that step change — only reveal themselves when the broader organizations, institutions, and governance it operates within adapt to address the way that technology shifts our capability.
This is what Reverse Alignment offers: a meta narrative and a frame for cross-sectoral coordination, the beginning of a human-side agenda to move from threat to flourishing, from subjugation to steering of super intelligence. The work has begun with a founding coalition of 27 experts across research, technology, government, philanthropy and civil society, with an initial co-authored paper published in Noema, which defines the program as “the ambitious creation and redesign of institutions, norms, skills and governance frameworks so that societies can harness AI safely and share its benefits widely.”
At RxC we are grateful for the leadership that E. Glen Weyl has demonstrated in proposing this agenda and in bringing together this extraordinary cohort to explain the stakes and first steps in making this transition. We’re honored to contribute on the democracy dimension of the 12 grand challenges, and also hope to contribute further by drawing on our research around data value, data dignity or labor rights.
Reverse Alignment could galvanize communities and concentrate the efforts of organizations around the world. We invite you to explore the Reverse Alignment website, to view the RxC Salon recording, and to discuss the ideas with us. If you can help to land a pilot or collaborate on a prototype of what Reverse Alignment looks like in practice, we would love to work with you. This is just the beginning of an agenda that we hope our work will align with and amplify in years to come.
About the Salon: introducing Reverse Alignment
On August 24, RadicalxChange convened a Salon around this ambitious agenda: six co-authors — Glen Weyl, Galen Hines-Pierce, Melissa Valentine, Jess Scully, Tim O’Reilly and James Evans — walked our community through the thinking behind Reverse Alignment, before we broke into small groups to work out what to build next.
Glen Weyl opened the event with the frame the whole paper hangs on. Capital expenditure on AI is now on the order of the railroads and the internet, and arguably bigger than Apollo and Manhattan combined — yet, as Glen put it on the call, “the social investments in adapting to these technologies are minimal.” Alignment makes machines consistent with human values; Reverse Alignment is the inverse move: repatterning human institutions so the tools we’re building can be useful, helpful and harmless at scale.
This ambition has precedents in economic history: “History shows that such investments are not optional supplements to technological change; they are prerequisites for its success,” the researcher Carlota Perez argues in her 2002 book Technological Revolutions and Financial Capital. Without alignment on the side of our institutions and governments, we get spectacular benchmarks and barely-there productivity gains — the meagre 0.66% projected decade-long rise in U.S. total factor productivity (Acemoglu) against Apollo-scale spend.
Galen Hines-Pierce laid the theory under that claim, drawing on Brynjolfsson and Rock’s “productivity J-curve,” Catalini’s work on verification, and Bullock, Hammond and Kreer on institutional constraint. The Noema piece frames this as a fork in the road: left unaddressed, “our default road is toward a world of productivity without prosperity, execution without verification and capacity without constraint” — but “this outcome is not inevitable.”
Galen’s deeper claim is that institutions take the shape of their era’s dominant cognitive technology. Industrial-era hierarchies matched centralized statistical tools (the Federal Reserve, Bretton Woods institutions); networked-era standards bodies matched layered protocols (IETF, ICANN). We’re now in the neural era — distributed, attention-weighted, gradient-assigned — and our institutions haven’t caught up.
| Era | Dominant cognitive technology | Institutional architecture that recapitulated it | Representative cases |
|---|---|---|---|
| Industrial / mass production | Standardized accounting; assembly lines; centralized statistical aggregation | Hierarchical, divisional bureaucracies; centralized fixed-rate regimes; chartered multilaterals | Federal Reserve; Bretton Woods (IMF, World Bank, GATT); the M-form corporation; Progressive-era regulatory state |
| Networked / packet-switched | Layered protocols; distributed routing; end-to-end design | Layered standards bodies; open RFC-style governance; member cooperatives mirroring network topology | IETF / W3C; SWIFT; ICANN; India Stack |
| Neural / learned representation | Distributed embeddings; attention and bridging; gradient-based credit assignment; continuous feedback | Institutions that learn: continuous deliberation; data-attribution rails; verifiable credential graphs; surprise-weighted evaluation | Polis / Community Notes bridging; collective-management organizations for data; W3C verifiable credentials; metascience funding on collaboration graphs |
Contributions from the authors
The paper names twelve “grand challenges” this mismatch creates; five of the co-authors on the call led one each.
Melissa Valentine (Stanford, organization design) argued that org charts are now “a subfield of computer science” — a line from her mentor sociologist Jerry Davis that reframed her whole field. Her workplace research shows static org charts constrain the very problems AI-era teams are trying to solve; the fix is representing people by their actual skills and relationships, not their box on a chart, so teams can assemble around tasks and dissolve when done.
Jess Scully (RadicalxChange) took this into democracy: platforms reward outrage and hide where people already agree. The fix is using the tools of algorithmic bridging and mindset of Civic AI to make democratic processes more responsive and reflective of what wins assent across a divide: building from Polis and Community Notes and working towards the level of meaningful public participation that Japan’s Team Mirai is beginning to unlock. One of the enabling moves is to embed AI-era democratic organisation and feedback loops into bureaucratic decision making, as demonstrated by Taiwan’s Participation Officer Network.
Tim O’Reilly (O’Reilly Media) distinguished institutions from mechanisms. YouTube’s Content ID proved creators could be paid once someone built the mechanism, while AI labs have called the same problem impossible. His fix borrows from the models themselves: backpropagation already assigns credit by marginal contribution, and collective-management organizations (the music-licensing model) could route the payments.
James Evans (University of Chicago and Google) warned that AI deployed in science is closing down fields faster than opening them. Models like AlphaFold assume existing data is complete and interpolate, missing the surprising, boundary-violating findings that drive real discovery. His fix: fund informational surprise and silo-breaking synthesis, not citation counts.
Galen Hines-Pierce closed on labor transition: a degree is “a one-shot encoding of a person,” built for an era of single-employer mass layoffs, not today’s continuous task-level reshaping. Portable microcredentials could let people prove precisely what they can do. He also argued that the coming fiscal squeeze may finally force the “grand bargain” that produces New Deal-scale institutions.
One thread ran under every talk: credit assignment. Backpropagation, attention, and surprise-detection aren’t just model internals — they’re templates for paying data contributors, funding scientists, and bridging polarized communities. If institutions really do take the shape of their tools, the neural era’s tools already show us the shape to build toward.
Feedback and ideas from the breakouts
After the co-authors spoke, we split into four small breakout groups, asked what brought people there, what could become a pilot in the next 6–12 months, and what would help this argument reach further. Here’s an anonymised sense of what came back.
Trust came up from several directions at once. Different groups converged, independently, on the same design problem: when an AI agent stands in for a person — negotiating on their behalf, synthesizing their views in a deliberative process, mediating a community conversation — how do you keep it faithful to what that person actually thinks? No one had a full answer, but its surfacing unprompted in multiple rooms is itself a signal.
A second theme was fragmentation: promising democracy and governance projects — citizen assemblies, deliberation platforms, sense-making tools — already exist in parallel with limited visibility into each other’s work, and funding pressure can turn collaborators into competitors. Proposed fixes ranged from a coordinating body to vet institutional proposals toward implementation, to a cross-field unconference pairing AI-native projects with people who have a longer track record running citizen assemblies and deliberative polls.
A third theme sat in tension with the first two: not everyone wants an AI-forward answer. A minority but persistent thread argued for organizing power that doesn’t depend on advanced computing at all — school-district level organizing for fairer redistricting, for instance — treating near-term reliance on AI as a risk worth naming rather than assuming away.
There was rough convergence that cities and states, more than national or international bodies, are the most tractable near-term scale, with live projects already at that level, though one group left open whether cities, civil society, or regulators are the right home for pilots at scale. And on format: future sessions should go deep on one topic rather than range across all twelve challenges — democracy and education were the strongest pulls in the room.
Where to now: ideas from our community and we invite yours
The argument is public; building it is a portfolio of bets across sectors, with no single party positioned to deliver alone. We’ll follow up individually with the people already carrying these ideas forward, and fold the open questions our breakouts raised — on collaboration, on trust, on holding the tension between this community’s AI-forward and low-tech wings — into how we run the next Salon.
If any of this resonates, we want to hear it: join our monthly global community call (first Wednesday, 9am Eastern), or — if you have something ambitious at a draft stage that could use more minds at play, and in questioning or landing your research in communities — drop us a line and we can explore how we might think it through together at an RxC Salon. Read the full paper, meet the wider 27-author coalition, and see where you fit at reversealignment.ai.