What is Reverse Alignment?

The other side of "alignment", which fits AI to humans: an effort to rebuild society's institutions, rules, people's skills and the ways decisions are made so that AI can be used safely and its benefits widely shared. Glen Weyl and others put it forward as an agenda in 2026.

When people talk about AI safety, the usual question is how to fit AI to human values. This research is called "alignment". But even a perfectly well-behaved AI could disappoint, or do harm, if it arrives in a society that isn't ready to receive it. Reverse alignment is an idea that focuses on this readiness on society's side.

In short

Reverse alignment means rebuilding society's institutions, rules, people's skills and the ways decisions are made, so that AI can be used safely and its benefits widely shared [1][2]. It is an agenda put forward in 2026 by the economist E. Glen Weyl and others.

Weyl's slides present the two as a pair: alignment brings the AI side in line with human values, while reverse alignment prepares the institutional side to receive it [3].

Note that another paper uses the same term in an entirely different sense. This is explained below under "What is debated".

For example

An essay in Noema magazine (July 21, 2026). Weyl, James Evans and Chris White published it in the magazine Noema, together with a group of 23 people from different fields, regions and political positions (below, "the coalition"). The essay opens with an event. The US government ordered the AI company Anthropic not to offer its most powerful AI to foreign nationals. Within hours, the company stopped offering it to all of its customers. The essay points out that the institutions for governing, absorbing and understanding AI have not kept pace with the growth of its capabilities [1].

Historical precedents. The same essay lists cases where the benefits of technology spread thanks to institutional innovation: the institutions of mass democracy that emerged after 19th-century industrialization (antitrust laws, independent regulators, labor unions and so on), the international money system set up after World War II, ARPANET (the forerunner of the internet), India's digital infrastructure India Stack, and Taiwan's digital democracy [1]. Weyl's slides sum this up as "the lesson of the dynamo": before a new technology truly pays off, it waits decades for institutions to catch up [3]. The coalition's website adds an example: for about 20 years, electrified factories kept the layouts of the steam era, and productivity did not rise [2].

The public lecture "Platform Originals" (August 29, 2026). Audrey Tang noted that when factories and mines first appeared in Britain, it "took one or two generations" before limits on child labor were written into institutions. With AI, she said, what we do not want is "the British factories again, where you wait one or two generations, a whole childhood spent inside the problem," before compensation comes [4][5].

How it works

Three failures if nothing is done

The Noema essay names three failures that occur when society doesn't prepare [1][2].

Productivity without prosperity — Jobs are reorganized faster than people can retrain or move. The gains that are created then fail to reach most people.
Execution without verification — Thanks to AI, the effort needed to produce text, images and computer code falls almost to zero. What becomes scarce is the human capacity to check whether what was produced is genuine, trace where it came from, and take responsibility for it.
Capacity without constraint — The same tools also lower the effort it takes for states and large organizations to watch people. Unless the mechanisms for objecting and for protecting freedom grow alongside, power concentrates in one place.

The essay argues that none of these is inevitable fate; each is the result of neglecting institutions [1].

Lopsided spending

The essay also takes issue with the fact that, compared with the money spent on making AI smarter, only a tiny share goes to the institutions, education, organizations and civic infrastructure needed to integrate AI into society [1].

Twelve challenges

The coalition's website divides the work into 12 areas [2]:

Identity (showing you are a real, unique human without revealing who you are)
Privacy, proof of provenance (checking who made a piece of information), and data value (payment for the human record used to train AI)
Collaboration between AIs, collective sensemaking (making sense of the situation together), democracy, and law and freedom
The workplace, research, education, and transitions in work

Many of these overlap with topics the book Plurality covers: proof of personhood, data dignity, bridging and broad listening.

The hypothesis that institutions resemble their tools

The coalition has a distinctive hypothesis: when a society absorbs a new technology, the institutions that succeed come to resemble the shape of that era's central technology [2].

In the industrial era, for example, bookkeeping in a standard form and the assembly line, where work flows step by step, gave rise to large companies where orders flow from the top down, and to central banks. In the internet era, communication systems that split the work into layers, each with its own role, gave rise to organizations that also cooperate in layers — such as the bodies that agree on the technical standards of the internet.

Neural networks (the basis of today's AI) find what people have in common without averaging their differences away. They can also trace how much each part contributed to a result. The coalition expects that institutions able to receive AI will imitate this behavior at the scale of society [2].

Who is behind it

The coalition's website lists both Weyl and Tang among its members [2]. There are also Japanese and Chinese-language versions of the site [6][7]. In September 2026, RadicalxChange, a nonprofit founded by Glen Weyl, announced that it would take on the democracy area among the 12 challenges [8].

What is debated

Beware a different concept with the same name. In an April 2026 paper, David Jacob Harrison discussed "the reverse alignment problem". This refers to a phenomenon in which, instead of machines being fitted to humans, human values and goals become simpler and people are made more machine-like. It points in the opposite direction from Weyl and colleagues' proposal to deliberately rebuild society's institutions [9]. This article uses the term in Weyl and colleagues' sense.
How effective altruists (EA) received it. Effective altruism is a movement that asks how to do the most good with one's donations and career. On the EA Forum, Tom Vargas summarized the core point this way: if institutions are wrong, adding advanced AI that is well aligned with humans will probably still make outcomes worse. Vargas found it plausible, and wrote that because it deals with the foundations of society it is very hard to carry out, but could change things far more if it works. In the comments, some said in effect that aligning individual models does not guarantee that the whole system using them is aligned [10].
"Who aligns to whom?" At the lecture, someone asked whether reverse alignment is about deciding who has to accommodate whom. Tang answered that past institutional responses to industrialization, electrification and the internet all answered the same questions: "If you are not satisfied, whom do you appeal to? How do you come to know? How do you opt out, or switch to another system?" For AI, she said, the hope is that it will look more like the internet: small-scale trials, and the ability to build better ways yourselves [4].

Where it fits in Plurality

The term "reverse alignment" does not appear in the book Plurality (English edition). It is a claim made in 2026, after the book, by Weyl and the coalition. Care is needed not to mix it up with what the book says.

Still, its substance connects closely with the book. In its conclusion, the Noema essay's prescription is cooperation that bridges divided communities, places trust not in central authorities but in diverse webs of human relationships, and finds common ground without suppressing disagreement [1]. That is the idea of collaboration across social difference at the heart of the book. Its use of Taiwan as an example also draws on the same material as Part 2 of the book (Taiwan's digital democracy).

The 6-Pack of Care and Civic AI, in which Tang is involved, sit close by. Both agree that fixing what is inside the AI is not enough on its own. The 6-Pack is a framework for designing procedures in the places where AI works, while reverse alignment is an agenda calling for investment in society's institutions as a whole [2][11]. For how this relates to debates about the future of AI, see the AI futures debate.

Sources

Weyl, Evans and White, "What Humanity Needs To Flourish In The Next Decade" (Noema, 2026-07-21) — https://www.noemamag.com/what-humanity-needs-to-flourish-in-the-next-decade/
Reverse Alignment coalition website — https://reversealignment.ai/
Weyl, "Humanity's Sociotechnical Grand Challenges" (slides, 2026-06-23) — https://reversealignment.ai/assets/documents/grand-challenges-deck.pdf
Platform Originals: The Other Side of Alignment (2026-08-29, English translation of the transcript) — https://github.com/audreyt/transcript/blob/main/2026-08-29-Platform-Originals-The-Other-Side-of-Alignment.md ［CC0］
The same lecture, "對齊的另一面" (transcript in Chinese) — https://sayit.archive.tw/2026-08-29-platform-originals-%E5%B0%8D%E9%BD%8A%E7%9A%84%E5%8F%A6%E4%B8%80%E9%9D%A2 ［CC0］
Reverse Alignment Japanese-language website — https://reversealignment.jp/
反向對齊 (Reverse Alignment), Chinese-language website — https://reversealignment.tw/
Jess Scully, "On Reverse Alignment" (RadicalxChange, 2026-09-10) — https://www.radicalxchange.org/updates/blog/on-reverse-alignment/
David Jacob Harrison, "The reverse alignment problem" (AI & Society, 2026-04-25, abstract) — https://link.springer.com/article/10.1007/s00146-026-03043-4
Tom Vargas, "Reverse alignment" (EA Forum, 2026-07-22) — https://forum.effectivealtruism.org/posts/eAq8R3hs7qg6DareE/reverse-alignment
Civic AI — https://civic.ai/ ［CC0］

We welcome error reports and questions ("What is…?"). Submissions become public issues.

Report an error in this article ・ Ask a question