# Humane Radicalism

Source: https://informationaldemocracy.substack.com/p/humane-radicalism

Audrey Tang turns a familiar safeguard around. Instead of humans kept in the loop of AI, she argues for AI in the loop of humanity: machines that communities invite in for a question they chose, that answer through a named human who signs for them, and that are built from the first day to be sent home. This essay develops Audrey Tang’s provocation — From humans in the loop of AI to AI in the loop of humanity. Read the full provocation → Essential points in this essay: A click is not a mandate. Students would take money to leave TikTok alone but pay to leave it together. Mandate belongs where the relationships are: portable social graphs, as in Utah’s Digital Choice Act, and assemblies convened for one question, as in Taiwan in 2024. Every machine output needs a byline. A model may draft, a named human signs and can be called to account, the way a newspaper keeps a corrections column, a letters page and an editor who can be summoned. Civic AI is scaffolding. Its success is what a community can do once the machine has gone, so the hand-back is planned and budgeted on the first day, together with the rights to fork, merge, retune and retire. About the author: Audrey Tang is Taiwan’s cyber ambassador and a 2025 Right Livelihood Laureate. Read the biography → Share AI in the loop of humanity means machines that many communities can invite in, hold to account and send home. The Japanese word for automation is 自動化, literally “self-moving”. Toyota Motor Corp. writes it with one extra mark. To the left of the character for motion it adds 亻, the radical for a human, and “move” becomes “work”: 自働化. The company calls the principle jidoka, translates it loosely as “automation with a human touch” and traces it to the looms of Toyota Group founder Sakichi Toyoda, which stopped themselves when a single thread broke. The loom maker he founded, now Toyota Industries Corp., still spells its Japanese name without that mark: 豊田自動織機. The whole idea lives in that one mark. A machine that only moves keeps running until someone intervenes. A machine that works with people knows where its competence ends. When a thread snaps, it halts and calls a human, and the flaw is mended before it runs through the whole bolt of cloth. On Toyota’s assembly lines the same mark hangs overhead as the Andon cord. Any worker who notices something wrong can pull the cord. A supervisor comes to the station and, if the problem cannot be fixed in time, the line stops. The right to stop belongs to the people nearest the work, and the cord means something only when pulling it is safe. My first essay for this series named the stance that carried industrial logistics into public life: MaxOS, the maximisation operating system. MaxOS kept the line’s speed and dropped the radical. It lets people back in only as labour inside its loop. Yet even jidoka, at its best, keeps the human in the loop of the line. The line sets the tempo, and the human keeps pace. That essay posed a concluding string of questions: “Who sets a system’s mandate, who inspects its synthesis and who may reopen what it has closed?” My answer is an inversion I have been calling “AI in the loop of humanity.” People set the tempo. The loops are the ones communities already run as they listen, decide and correct themselves, and the machine joins them on three conditions: It is invited, it answers for what it says and it can be sent home. Whose loop? The honest answer is plural. Nobody lives in only one community. A nurse can belong at once to a hospital ward, a union branch, a congregation, the language spoken at home and a street that floods every typhoon season. Walt Whitman gave the condition its line: “I am large, I contain multitudes.” A democratic design should begin from those multitudes. I call this approach techno-communitarianism: a way of building Civic AI that strengthens the communities it serves. The word needs care in a series that opened with the public sphere. I keep that tradition’s normative inheritance whole: reasons given in public, standing to be heard and accountability held in common. What changes is the unit of redesign, from a single society-wide public to many communities. The communities I have in mind are where that inheritance can be practised at human scale. They are small enough for people to know one another’s reasons. They are porous, so nobody is confined to one. And each stays answerable beyond itself. Communitarianism also has a history of closed doors. Georg Simmel observed more than a century ago that a person standing where many social circles intersect is held by none of them completely. The best guard against any community closing its doors is a human who belongs to several. Here is what such a community could do with tools already within reach. It could ask a machine to sit in on its listening. The machine would gather what members think and show where they agree and where they part. It could draft what the community might decide, under a charter the members wrote and can amend, and every summary would carry the name of a human willing to answer for the words. When the nurse changes wards or leaves a group, her friends and her records would go with her. And once the community’s own habits of listening had grown strong, it would send the machine home and keep the tools. None of this waits on a breakthrough. Each piece is already running somewhere. What is missing is the expectation that every civic machine can answer three plain questions: Who asked for this machine, who answers for what it says and who can send it home? Share The strongest objection comes first. Have the people not already chosen? Billions installed the apps and open them every day. That looks like a mandate. Economists Leonardo Bursztyn, Benjamin Handel, Rafael Jiménez-Durán and Christopher Roth put that intuition to the test with US college students. On average, active TikTok users asked for US$55 to deactivate their own account for four weeks. Yet 60 per cent were willing to pay to have every participating student at their university, themselves included, deactivate together. The authors call this a collective trap. The nurse knows the feeling. When her ward’s shift swaps and her street’s typhoon warnings run through the same platform, leaving alone means leaving both. A click is a ballot on the wrong question. It asks each person whether to leave alone, when the live question is whether to stay together, and only people who share a network can answer that. Mandate belongs where the relationships are. Two answers are already within reach. The first is portability. Since July 1, 2026, Utah’s Digital Choice Act has required social media companies to give people, on request, a portable copy of their personal data, including the social graph, and to maintain an interoperability interface built on an open protocol. A phone number already moves with its owner between carriers. Friends should move too. Once leaving no longer means losing everyone, exit becomes credible and memberships can overlap. The second is to make aggregation an event. A feed is aggregation that never stops running, a permanent plebiscite with no question on the ballot and no owner of the error. An assembly has the opposite shape. It is convened for one question, facilitated, answered by a named receiver and then dissolved. Taiwan’s 2024 assembly on information integrity had that shape, and a machine was already in the room. It began with 200,000 text messages to random numbers asking what should be done, and thousands replied with ideas or volunteered to take part. Its 447 participants met online in small groups on Stanford’s Online Deliberation Platform, where software kept the queue of speakers and the clock. The people did the deliberating. When the day ended the groups dissolved, and what outlived them was the government’s obligation to respond. The government did: that July, a new anti-fraud law required large online ad platforms to verify who places each ad and made them jointly liable for losses from scam ads they leave up once aware. An invitation to a machine should work the same way. It is issued by the people who will live with the result together, for a question they have chosen, and it lapses when the question is answered. In 2024 my ministry issued that invitation on the public’s behalf. Théophile Pénigaud, writing earlier in this series, describes chatbots that fill public life with “orphan reasons”, reasons no one can answer for. His sentence stays with me: “They speak, but they do not answer.” My parents were both journalists, and I grew up around an institution built to keep reasons from becoming orphans. A newspaper gives each story a byline, a corrections column, a letters page and an editor who can be summoned. The five-link chain I proposed last time — visible inputs, a synthesis that can be checked, a named human judgment, a path of contest and a receiver with authority to act — is really a masthead, generalised. A machine may draft. A named human signs and can be called to account for what was signed. The nurse already works this way: a handover note may start from a template, but it carries the name of the nurse who signs it. A masthead also must show that it learns. Even a change of mind can be staged, when an institution concedes the headline and ignores the reasons underneath. Learning shows in the reply to each dissent: “This is exactly how your dissent moved our decision.” A ministry can write that by hand to a few hundred participants. No ministry can write it by hand to the more than 23 million people of Taiwan. Here language models have a place on the letters page, drafting replies that a named human reads, corrects and signs. Researchers including Manon Revel of this working group and members of the Community Notes team have proposed the same division of labour for the notes that add context to posts. People and models both draft. Diverse human raters decide which notes are helpful enough to show, and those ratings train the models in turn. Answering also needs a record, and here I can speak from habit. When I served in the cabinet, the transcripts of my meetings were published for anyone to read, and I still keep a public archive of my conversations. Speech keeps the admission fee low. People take part without translating themselves into officialese, and in a room they can hesitate, repair and change course. Writing keeps the paper trail. Together they make conversation that keeps its own record. The same discipline can reach the machines. Testimony and facilitators’ notes can travel beside a synthesis as records in their own right, so that nothing is quietly averaged into a score. Models can be pinned to an exact version, and engineers have shown how to make identical inputs yield identical outputs, so a recount of an inference could one day be asked for in the same spirit as a recount of ballots. Membership can be verified without exposing identity, which for people facing intimidation is the precondition of honest input. Sylvie Delacroix, in her essay for this series, marks the limit of any record: “A record recovers what was said and set aside. It cannot recover what was never said because attention had been directed elsewhere.” Delacroix is right, and that is why the record is not the product. The reasons worth protecting are formed in what she calls in-between spaces, the small forums where a privately held intuition can become an argument fit for the public sphere. Her proposal that professionals articulate together how their tools should express doubt is exactly the kind of charter I have in mind. A machine earns its place in those spaces only if it widens them. Share Here the factory falls silent. The Andon cord can stop the line, but nobody on the line can send the loom home. Scaffolding is different. It is built to come down. In 1976, psychologists David Wood, Jerome Bruner and Gail Ross used the word for the help a tutor gives so that a learner can do what she could not yet do alone. Later work stressed withdrawing that help as competence grows. That is the sense in which Civic AI should compose away. Its success is measured by what a community can do once the machine has gone. I sometimes put it bluntly. If I send my robot to the gym, I lose muscle and the friends I made while working out. A fully automated deliberation would deliver tidy summaries and lose what the gathering was for. The topic on the agenda is often just the reason to gather. The relationships built through it are what lasts. Georg Diez and Felix Beer, in a reflection for this series, describe a state that runs on systems it cannot leave: “The state has not bought a tool. It has entered a dependency.” They add: “Contract expiry becomes a constitutional event.” The same holds for any community, which is why a hand-back must be designed on the first day, before anyone grows attached. It needs a named human inside the community who decides when the scaffolding comes down. On the nurse’s ward, that could be the charge nurse retiring the shift-swap assistant once the team can arrange swaps on its own. It needs a plan and a budget for leaving. And it needs records and tools that stay behind in forms the community can run by itself. Consent and exit are necessary. They are also thin. A community that can only say yes, no or goodbye has little say over the tool itself. The fuller set begins with the right to fork, which lets a community run a version that fits local norms. It continues with the right to propose a merge, so that when the fork works better the wider institution changes its mind. With those come the rights to tinker, to retune and to retire. Taiwan’s g0v community has practised fork and merge since 2012: change the letter o in a government web address to a zero, turning gov.tw into g0v.tw, and you reach a civic fork of the site. Several of those forks, including the budget visualisation where g0v began, were later merged back into government service. I call a system built this way a Kami, or knowledge artefact management intelligence, after the local kami of Japanese tradition, each belonging to a river, a grove or a neighbourhood. A Kami stays small and bound to one place or practice. It is open, and light enough to run on hardware a community owns, which makes democracy something people can cook at home. The facilitation kit at a kitchen table should be the same kit a ministry uses. Georg Diez, who edits this series, asked me in writing the question techno-communitarianism must answer: whether a near-endless set of mini-publics could ever add up to a whole. The internet is my existence proof. Tens of thousands of independent networks cohere by agreeing on thin protocols at the edges. No one runs the whole. The whole is a protocol. The internet also shows what happens when exit grows expensive: the whole recentralises one layer up. That is why portability belongs in the protocol. A plural democracy needs a protocol of its own, and it can stay thin. Communities are not always kind. Some exclude. Some organise to dominate others. The protocol’s first layer is therefore a rights floor that every fork inherits, and no local charter can lower. The second is portability, so that exit stays credible and memberships stay overlapping. The third is bridging visibility, so that each community can see its neighbours’ reasons without being ranked by them. Where a feed must rank at all, it can give more reach to statements endorsed across groups that usually disagree than to applause from one’s own side. A written charter matters most when a minority must contest a decision. A convener’s unwritten norms give a minority nothing to point at. A written, amendable charter lets it point to the exact line where its reason was dropped and ask for that line to be reopened. Pointing at a line assumes one can read it and argue over it, so taking part should also be possible in one’s own language, including a signed one. Taiwan’s law on national languages counts Taiwan Sign Language among them, alongside the spoken ones. The whole should speak only where decisions bind across communities, as with fraud or pandemics, and even then, through an event with a mandate. That leaves a constitutional question I would like our field to hold together: Who may convene an aggregation moment and who may refuse one? Every day the nurse carries reasons from her ward to her union branch and from her congregation to her street. That is bridging done by hand, and it is how a plural society has always held together. Legal academic Cass Sunstein described “incompletely theorized agreements”, in which people agree on what to do while disagreeing about why. Communities can act together that way, and so can the many belongings inside one human. We, the people, are truly the superintelligence. Portability law in Utah, assemblies that dissolve once the question is answered, notes drafted by models and chosen by people, Kami running on a community’s own hardware: every piece exists. What is missing is the expectation. It can begin wherever public money is spent. A contract for a civic machine can name who invited it, who signs for what it says and how it goes home, with a budget for leaving. The human radical is a small mark. It turned a word for machines that move by themselves into a word for machines that work in the loop of humanity. We can ask for three such marks on every machine we invite in: an invitation, a byline and a way home. In the Age of AI, those marks are ours to write. Leave a comment The Three Essays Audrey Tang opened with a provocation. These are the three essays that pursue it, published here as they are written, one phase at a time. Grounding — Dismantling MaxOS for Democracy Visioning — Humane Radicalism You are reading it. Demonstrating — Legibility, Peace-Tech, and Measurable Polarisation-Reduction — polarization-reduction as measurable, plus civilian peace-tech infrastructure. Forthcoming. Read the provocation → · Read the biography → · All provocations → Share Audrey Tang is Taiwan’s cyber ambassador and a 2025 Right Livelihood Laureate. She served as the country’s first digital minister from 2016 to 2022, its first minister of digital affairs from 2022 to 2024 and is an inaugural senior accelerator fellow at the Oxford Institute for Ethics in AI. This is the first of three essays for the Informational Democracy working group convened at the Max Planck Institute for Political and Social Science in Göttingen. Informational Democracy is an initiative hosted by the Max Planck Institute for Political and Social Science in Göttingen, led by Steven Vertovec, Georg Diez and Felix Beer.