Democratic institutions are not failing because the public sphere collapsed. They are failing because the infrastructure beneath it was tuned to maximise attention rather than to carry public judgment. Audrey Tang calls that stance MaxOS, the maximisation operating system, and argues that naming it is the first condition of being able to change it.
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:
The failure is not in the conversation but in the substrate. Where systems are tuned to maximise reaction, deliberative life is made expensive and reactive life is made free.
MaxOS has three habits that matter for democracy: it collapses judgment into score, it treats attention as extractable fuel, and it inserts the person into the loop of the optimiser.
Taiwan shows the alternative in practice. A deliberative poll, a formal response path and the Fraud Crime Hazard Prevention Act cut detected high-risk scam advertisements from a weekly peak of 77,484 to about 837.
About the author: Audrey Tang is Taiwan’s cyber ambassador and a 2025 Right Livelihood Laureate. Read the biography →
Democracy cannot repair its conversations until it names the maximisation operating system that shapes what can be heard, known and changed by the public.
Democratic institutions are not failing because the public sphere has collapsed. They are failing because the operating system beneath that sphere was never built for democratic life.
I mean operating system in a plain sense: The defaults that decide what counts as a signal, what gets ranked and which actions are cheap enough to become habits. For more than a decade, those defaults have rewarded attention captured, reaction accelerated and people made measurable. We open our phone to check on the world, and the session quietly reverses the relationship until the world is checking on us.
Call it MaxOS, the maximisation operating system.
MaxOS names a design stance that migrated from industrial logistics into public life, crossing companies, models and ideologies along the way. The name does not attack optimisation as such. Optimisation can be honest work inside a bounded task with a named owner and a stop condition. MaxOS is the extension of that stance into domains where no metric can legitimately stand for the good, no affected public authorised the objective and no stop condition bounds jurisdiction. Once a society accepts that stance as neutral infrastructure, democratic talk continues on the surface while the conditions of knowing and deciding are rewritten underneath.
Wrong unit of analysis
We often diagnose this as a crisis of conversation. Listening has become harder. Trust has fallen. Polarisation has risen. Each observation can be true in a room and still misname the actual failure.
Conversation runs on infrastructure. What a public can hear, compare, contest and remember depends on the systems that filter, rank, summarise, recommend and generate the material through which a polity perceives itself. When those systems are tuned to maximise reaction, the failure mode is not that citizens forgot how to deliberate. It is that deliberative life has been made expensive, while reactive life has been made free.
That is why “rebuild the public sphere” is too coarse a slogan. Public life has always run through many rooms, many languages, many communities of practice. What changed is the substrate those rooms sit on, and a substrate tuned for capture punishes the slow virtues democratic life depends on: patience, second thoughts and room to change one’s mind in public without penalty.
The object of democratic theory has to move with that fact. Beyond asking how citizens exchange reasons, it must now ask how the infrastructures that form the conditions of knowing, deciding and acting are built, owned, governed and, when necessary, interrupted.
MaxOS habits that matter
MaxOS has three habits that matter for democracy.
First, it collapses judgment into score. A ranking that is always on teaches people to perform, and it makes what can be counted look representative of what cannot. The feed does not ask whether a claim is answerable. It asks whether a claim travels.
Second, it treats attention as extractable fuel. Herbert Simon saw the shape of this long before generative models arrived: A wealth of information creates a poverty of attention. The feed economy industrialised that poverty by selecting reaction as its cheapest feedback signal, so the loop closes fastest exactly where reflection is thinnest.
Third, it inserts the person into the loop of the optimiser. “Human in the loop” sounds like a safeguard. Often it is a staffing plan for the machine. The human labels, confirms, reacts, steers a little and supplies the one resource the system cannot synthesise: lived attention. The phrase only earns its safety grammar when the objective can be inspected, the recommendation refused and the system interrupted.
None of this requires malice, and none of it is ownerless. The defaults that reward capture are held in place by revenue models, product roadmaps and institutional incentives, not simply by inheritance. Care, mutual comprehension and democratic correction are goods in their own right. Efficiency answers to them, not the reverse. Generative systems entered this pattern and made it fluent. A model that can draft, summarise, simulate consensus or produce a face on demand raises the throughput of capture faster than institutions raise the bandwidth of answerability.
A diagnostic from Taiwan
In early 2024, deepfake investment advertisements masked with famous faces spread across platforms in a society boasting an internet ranked the most free in Asia, according to Freedom House. This was an incompatible setting for a takedown-first reflex. Removing speech was never the plan; the target would be paid reach, not expression. The harm had to become governable without anyone being appointed editor-in-chief of reality.
The civic infrastructure that later made that path available was already standing. On March 23, 2024, while I was minister of digital affairs, we ran a deliberative poll titled “Utilizing AI to Enhance Information Integrity.” It was not convened in response to the deepfake advertisements; it existed so that broad listening would be ready before any single shock arrived. The government sent 200,000 random text invitations; 1,760 people responded, or 0.88 percent of those sent. Recruitment mixed the SMS pool with practitioner outreach: 447 participants deliberated in 44 facilitated groups, 12 of practitioners and 32 drawn more broadly by gender, age and residence. This was designed to approximate the population rather than mirror it, with the published analysis weights a sample of 436. The groups questioned experts and reconsidered proposals item by item.
Support for mandatory algorithm disclosure fell 27.5 points to 55.6 percent after participants considered the proposal and its trade-offs; support for detecting AI-generated content rose 2.6 points, while support for watermarking fell 2.5 points and for media-literacy segments fell 9.3 points. The largest swing ran against the most intuitive regulatory instrument, the opposite of what a convener seeking ratification would engineer. Large swings on technical questions can still reflect framing power; what made the process defensible was a response path anyone could contest rather than governmental pre-commitment to the result.
The assembly secured a formal response path while a government bill I had helped prepare was already moving, rather than a pre-commitment to adopt its every word. That sequencing is not a defect to hide: the meeting widened listening and tested a targeted, procedurally bounded design against a live legislative track.
Mandate-setting in layers
The Fraud Crime Hazard Prevention Act was promulgated July 31, 2024. It required stronger identity for advertisers; when a platform knowingly left scam advertising up, it could share liability with the advertisers. The act did not set the removal deadline itself; it authorised the ministry to prescribe one, and a subsidiary announcement fixed 24 hours from Nov. 30, 2024. That is mandate-setting in layers, each layer with its own author.
The act also carried fines of up to NT$100 million (US$3.1 million) and authorised measures such as traffic management, suspension of domain-name resolution and access restriction, on commencement dates the Executive Yuan would set separately: the most interruptive powers were the ones held back, and powers of that kind require transparent criteria, proportionality and a contest path.
By the ministry’s own account, detected high-risk investment-scam advertisements fell from a single-week peak of 77,484 in 2024 to a weekly average of about 837 by September 2025, a decline that began before the new powers took effect and belongs to the full set of instruments working together. The assembly complemented representation, widened the listening and left a trail anyone could examine.
Seen through MaxOS, the episode reads simply. Where the substrate rewards speed and virality, fraud becomes a native application. Where a society has invested in broad listening, contestable process and a response path, the same shock becomes governable. Deliberation of this kind is infrastructure, and it must exist before the week it is needed. The method travelled with civic habits built over years; without those, the lesson is to build the response path before copying any assembly.
Mediation to epistemic agency
Taiwan’s case extends beyond fraud. It shows institutions keeping a capacity that the maximising substrate erodes everywhere else: the capacity to notice, weigh and decide in public.
For years the democratic concern about technology was mostly about mediation: Who carries the message, who moderates the square, who owns the press? That concern remains, now joined by a deeper one. AI systems perform epistemic functions. They search, rank, summarise, advise, moderate and act. They shape what a ministry notices, what a classroom treats as a finished answer, what a voter sees before a claim can be checked and what someone reaching for company at 2 a.m. is offered. The stake is whether publics and institutions keep the capacity to know and decide for themselves, or cede it by default.
This is the shift from information mediation to epistemic agency. Mediation moves the material. Agency chooses the question, decides when it is settled and acts on the answer. It is also why the old marketplace of ideas metaphor fails: A marketplace assumes buyers who can still inspect the goods, and the substrate trains the inspector to move at the tempo of the inventory.
A failing consultation theatre
Many governments and platforms now know they must listen, so they run consultations, panels, feedback chatbots and sentiment dashboards. Voice becomes abundant while consequence stays scarce.
Another channel for expression rarely helps. Nancy Fraser’s distinction between publics that only form opinion and publics positioned to influence authoritative decision names the fracture; the missing piece is a pipeline from collective sense-making to institutional action that humans can inspect: inputs that can be seen, synthesis that can be checked, a judgment that can be named, a path of contest and a receiver with authority to act. Without that loop, participation teaches cynicism. People learn that speaking soothes the institution without binding it.
The gap is easy to miss because simulated responsiveness is cheap: a personalised acknowledgment, a room summarised in seconds, the appearance of having taken something in. Legitimacy still comes from elsewhere, from disagreement that stays visible, a person who stands behind the decision and an affected public that can interrupt the system when it drifts.
A single, unified rational public sphere is therefore the wrong recovery fantasy. Premature consensus is MaxOS at its most polite. It smooths away the plural exactly where the plural is the democratic signal, pressing many publics into the feed’s single template. What deserves protection is the plural itself: many publics, each keeping its own record of what remains unsettled. Fraser’s counterpublics make the case: Where a society is stratified, a plurality of contesting publics serves participatory parity better than one comprehensive sphere.
Design stance we mistook for fate
If the diagnosis stopped at “platforms are too powerful”, the prescription would narrow to antitrust or better recommender systems. Both can matter, and both leave a deeper confusion untouched.
The confusion is between two kinds of design. One is proper to optimisation under constraint: Choose a metric, reduce friction, scale the winner. The other is proper to shared life: Name who is owed an answer, preserve room for disagreement, keep correction cheaper than denial. MaxOS extends the first stance into the second domain and calls the result progress. Inevitability is the story an installed system tells about itself. Installed systems can be relocated, bounded, slowed and, when they fail the people who live with them, retired.
The same extension explains how convenience keeps turning into jurisdiction without a public hearing. A tool acquires authority by becoming difficult to refuse, then treats its accumulated reach as a mandate. Scale expands reach, and reach keeps being mistaken for standing. Seth Lazar’s account of the algorithmic city asks how intermediary power can claim procedural legitimacy. The claim here is adjacent and narrower: An optimisation stance installs that power, and an interruptible response loop is what can still reclaim it. Democratic infrastructure reverses that sequence: the mandate is named first, use stays contestable and the people affected can pause the system without waiting for permission. Interruptibility is not a flaw in democratic technology. It is one of its constitutional virtues.
Naming the foundational substrate
A first essay should open a field rather than settle it. Naming MaxOS does not choose the site of repair: professional communities can articulate their own norms, institutions can bind listening to delivery, publics can govern attention and computational power as common concerns. Nor does it settle what synthesis may do. Some work needs outputs an institution can run on, and some work must refuse to erase uncertainty, authorship or minority dissent in order to look decisive. The live question sits where those meet: Who sets a system’s mandate, who inspects its synthesis, who may reopen what it has closed?
The next essays turn from diagnosis to design: AI in the loop of humanity, entering communities as scaffolding for mutual listening that is meant to compose away, and the practical case that polarisation-reduction can become as measurable as carbon became for climate policy.
The diagnosis comes first: democratic speech cannot recover while the systems beneath it reward capture, speed and legibility. Once the operating system has a name, it stops passing for fate. The defaults we inherited become defaults a public can choose.
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. You are reading it.
Visioning — AI in the Loop of Humanity, and the Many Bounded Communities — techno-communitarianism and Civic AI as scaffolding that “composes away.” Forthcoming.
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 →
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.
Feels like the strongest point is that giving people more ways to speak is not the same as giving them influence. A democratic system must keep a record of how public input reached a decision, who was responsible, and how mistakes can be challenged and corrected. That record must survive when the technology changes. AI should strengthen this chain, not quietly replace it.