AI and the public sphere: who authors the decisions made in your name
The argument about artificial intelligence and democracy is usually run as a question of benefit and harm. The harder question is authorship. Arendt, Habermas, Foucault, Rousseau, Marx and Norbert Wiener each supply a piece of it, and the 2026 research supplies some awkward facts.
One adult in ten now gets news from a chatbot in an average week. That comes from the Reuters Institute's Digital News Report 2026, published on 16 June, which found weekly chatbot use for news across 45 markets rising from 7 per cent to 10 per cent in a single year, and reaching 17 per cent among the under-25s. A second figure in the same report deserves more attention. Across 27 markets, just 4 per cent of respondents say they always or often click through to the news source underneath an AI answer, against 19 per cent from search. The answer arrives; the source does not.
What arrives is frequently wrong. An audit by the European Broadcasting Union and the BBC, published on 22 October 2025, examined more than 2,700 responses from ChatGPT, Gemini, Copilot and Perplexity, generated in 14 languages by 22 public service broadcasters across 18 countries. Forty-five per cent contained at least one issue the researchers classed as significant.
It would be easy, and lazy, to run from there straight to disinformation. The accuracy question is real but not the deep one, and the evidence on manipulation is more equivocal than the public conversation assumes. The question worth asking is older: whether citizens remain the authors of the decisions taken in their name.
The question 'does AI help democracy?' cannot be answered as posed#
Begin with the scholarship, which has recently done something useful. On 21 August 2026, Philosophy & Technology published a scoping review by Joachim Wiewiura, Amanda Jørgensen and Thomas Ploug — volume 39, issue 3, article 158 — covering 57 peer-reviewed articles under PRISMA-ScR protocols. Their finding is not that AI is good or bad for democracy, but that the question is malformed until you say which democracy you mean. "Democracy," they write, is "an essentially contested concept" that "invokes many different levels at which AI may assist and obstruct the legitimacy of democratic government." A recommender that surfaces a wider range of views is a gain on a deliberative model and close to irrelevant on an aggregative one.
Two of their secondary findings have more bite. Conceptual vagueness about democracy and pessimism about AI tend to travel together: the papers least clear about what they are defending are the gloomiest about the threat. And democratic models in this literature are, in their word, "inherited" rather than reworked, lifted intact from the 1990s and applied to machines that did not then exist. A scoping review settles nothing, as its authors say. It does establish that much of the shouting has been about different things.
The Swedish Academy gave the 2026 Nobel Prize in Literature to Anne Carson on 8 October. Before the verse novels and the box of grief came a doctorate in classics and a 1986 book on Sappho's one word for desire. Why philosophers have read her for forty years, and what the prize gets right.
Language models now score well on truthfulness tests and badly on honesty tests, a gap the MASK benchmark was built to measure. Two older accounts of truth, Heidegger's unconcealment and Foucault's parrhesia, suggest why: telling the truth has always been something a speaker risks, and a model has nothing to lose.
Arendt: the loss is not truth but a world held in common#
Hannah Arendt's account of the public realm in The Human Condition (1958) is often reduced to a warning about privacy. It is closer to the opposite: publicity, for her, is the condition of the real. "For us, appearance — something that is being seen and heard by others as well as by ourselves — constitutes reality," she writes on page 50. Two pages later comes the image everyone remembers, of the world as "a table located between those who sit around it," which "relates and separates men at the same time."
Apply that to the 4 per cent click-through figure. A newspaper report is a thing in the world: another citizen can read the same words, dispute them, sue over them, cite the page. An answer generated once, for one person, is not reliably reproducible even for them. Nobody else saw it. There is no table. The worry is less that the answer may be false — the EBU figures say it often is — than that an unwitnessed answer cannot do the work Arendt assigns to appearance.
The loss is serious and should not be overstated. Arendt's public realm was a demanding and frankly exclusionary ideal, and she was contemptuous of mass society long before there were algorithms to blame. What her analysis supplies is a criterion rather than a golden age: ask of any information arrangement whether it produces objects that can be jointly attended to.
Habermas: legitimacy is a property of procedures, not of outputs#
Jürgen Habermas, who died in March at the age of ninety-six, had already offered his own late diagnosis. In "Reflections and Hypotheses on a Further Structural Transformation of the Political Public Sphere," published in Theory, Culture & Society volume 39, issue 4 (2022), pages 145 to 171, he argued that digital media were "turning all participants into potential authors" while fragmenting the space in which authorship might matter. The expansion and the fragmentation are one event.
His older point is procedural, and more important. On a Habermasian account the legitimacy of a decision is not conferred by its quality: a benevolent expert who guesses correctly what a population needs has produced a good outcome and no legitimacy whatever, because legitimacy is a property of how reasons were exchanged among those bound by the result. A system that ingests the expressed preferences of thirty million people and returns a distribution has not conducted a conversation. It has taken a reading.
This week the same journal published the sharpest statement of the problem yet. Lorenzo Manuali's "When LLMs Threaten Democratic Autonomy" appeared in Philosophy & Technology volume 39, issue 3, article 173, on 15 September 2026, and its target is the well-meaning projects that use large language models to improve deliberation. Facilitative models, which moderate or summarise discussion, tend towards what he calls perspectival homogenisation: they "diminish or eliminate disagreement and diversity of perspectives," and the views they smooth away are disproportionately minority views. Representative models, which extrapolate what a citizen would prefer across options never put to them, do something worse. You cannot intend an outcome you have never foreseen, and democratic autonomy on his account requires that a group "possess a joint intention" about what it is doing.
His remedy is where the argument wobbles. Manuali suggests mitigating homogenisation by randomising across models, a technical fix to a problem he has just characterised as constitutive. Varying the machine that forms the intention for citizens is a more diverse ventriloquism.
The optimistic case is stronger than sceptics allow#
Against all this stands a body of work that has been run, with control groups, and produced results.
James Fishkin has spent three decades on Deliberative Polling, and his Stanford group built an online platform with an automated moderator to do at scale what previously required an airport hotel and a weekend. The results appeared in Perspectives on Politics, volume 23, issue 2 (2025), pages 434 to 451, under the authorship of Fishkin, Valentin Bolotnyy, Joshua Lerner, Alice Siu and Norman Bradburn. Nine hundred and sixty-two participants deliberated on climate and energy policy in 104 small groups, against a control group of 661 who answered the surveys without deliberating. Deliberators moved substantially towards agreement, the largest shifts coming among those who began most sceptical, and the differences were still detectable at the October 2022 follow-up. Automation made a good process affordable.
Google DeepMind took a different route. Michael Henry Tessler and colleagues, publishing in Science on 17 October 2024, built a system unsubtly nicknamed the Habermas Machine, which drafted group statements from individual opinions and ranked them by predicted endorsement. Across 5,734 participants, some recruited through the Sortition Foundation, its statements were preferred to those of human mediators 56 per cent of the time. Municipal practice has followed. Bowling Green, Kentucky, a city of roughly 150,000 expecting 75,000 more residents by 2050, ran a month-long consultation in February and March 2025 on the Polis platform with Jigsaw's sensemaking tools; nearly 8,000 residents took part, generating over a million individual judgements, and Sam Ford of Innovation Engine reported that a high proportion of statements drew 80 per cent agreement or better.
The friction, to their credit, comes from the enthusiasts. Writing at the Knight First Amendment Institute on 1 August 2025, Tessler's own group set out what they do not know. Verification does not scale: "it is already challenging to evaluate a group statement produced by an AI mediator that is aggregating 1,000 opinions (e.g., is the statement based on hallucinated opinions?)." They flag algorithmic aversion and the prospect of participants misrepresenting preferences to game the aggregate. Their central question is whether any of it survives contact with real stakes: "Would participants engage in deliberation in good faith in the presence of genuine political stakes?"
A further limit is structural rather than technical. Every one of these exercises is consultative, and none binds anybody: authorship still rests, formally, with a council or a legislature. The interesting cases are the ones nobody has run.
Foucault: classification is productive, and prediction is not surveillance#
Michel Foucault is the name most often taken in vain here. The lazy version has him predicting the panopticon and stops. In "The Subject and Power," published in Critical Inquiry in the summer of 1982, he defines the exercise of power as an action upon actions, a mode that works "to structure the possible field of action of others," and adds a condition that is routinely dropped: "Power is exercised only over free subjects, and only insofar as they are free."
That condition does real work. It makes "power has moved to the data centres" a category error as usually stated: power on his account is a relation rather than a substance to be relocated, and it is not the same thing as domination, which is what you get when the field of possible action has been closed rather than structured. So the question to put to a predictive system is precise. Does it narrow the field of action, or foreclose it? A benefits algorithm that rules you a fraud risk, with no route to contest the classification, is not power at all but domination.
The productive point survives intact. Classification does not describe a population; it makes categories that act back on the classified. Credit risk, likely reoffender, student flagged for intervention: these are institutional facts that become self-confirming, and the knowledge that produces them is itself a form of power.
Shoshana Zuboff's The Age of Surveillance Capitalism (2019) gave the argument its most quoted term. Behavioural surplus is the residue: behavioural data beyond what is needed to improve the service that generated it, claimed unilaterally as corporate property. Testifying to the Committee on House Administration on 16 February 2022, she put the mechanism plainly. Human data "are computed into behavioral predictions and targeting algorithms," and those predictions are sold into "commodity markets in human futures akin to markets in pork belly futures, or wheat, or oil."
The framework is not settled and should not be reported as though it were. Evgeny Morozov's review in The Baffler on 4 February 2019, "Capitalism's New Clothes," ran to some 16,500 words and made two charges that have not been answered. One is methodological: her business history confirms its own model rather than testing it against rivals such as platform capitalism, and "Zuboff's discussion of alternative explanations never arrives. Perhaps seven hundred pages were not enough." The other is historical. "To view surveillance capitalism as our new invisible Leviathan is to miss how power, under capitalism, has been operating for several centuries."
The empirical work has also been unkind to the strong version of the manipulation thesis, and honesty requires reporting it. Kobi Hackenburg and Helen Margetts, in the Proceedings of the National Academy of Sciences on 9 June 2024, tested whether GPT-4 messages tailored to individual demographic and political profiles beat untargeted ones. The messages persuaded; the tailoring did not. "The persuasive impact of microtargeted messages was not statistically different from that of non-microtargeted messages." Their follow-up, "The Levers of Political Persuasion with Conversational AI," published in Science on 11 December 2025 with colleagues at the LSE, Stanford, MIT and the UK AI Security Institute, drew on nearly 77,000 UK participants and around 91,000 dialogues. Model size mattered only modestly; targeted post-training raised persuasiveness by up to 51 per cent, and roughly half the explainable variation between models traced to one variable, the density of information in the reply. And a finding that should discomfort everybody: the more persuasive a model was, the less accurate its claims tended to be.
On deepfakes the corrective is sharper. Sayash Kapoor and Arvind Narayanan examined 78 election-related deepfakes for the Knight First Amendment Institute on 13 December 2024 and found half showed no deceptive intent at all; for every deceptive case, comparable content could have been made without AI for a few hundred dollars by hiring an editor or a voice actor. The Alan Turing Institute's Centre for Emerging Technology and Security reached a parallel conclusion on 19 September 2024, counting 16 viral cases in the UK general election and 11 across the French and European contests, with, in lead author Sam Stockwell's words, "no evidence AI has impacted the result of an election." What CETaS did document was a second-order harm: deepfake pornography aimed at women politicians, and a general "erosion of confidence in what is real and what is fake." What voters actually see is a narrower problem than the discourse suggests, and a more stubborn one.
Rousseau: the general will is not a quantity that can be measured#
The distinction that matters most here is 264 years old. In Book II, chapter 3 of Du contrat social (1762), Rousseau separates two things that any aggregation technology will confuse: "There is often a great deal of difference between the will of all and the general will; the latter considers only the common interest, while the former takes private interest into account, and is no more than a sum of particular wills."
A platform that models what a population clicks on, buys and lingers over is measuring particular wills, under conditions designed to elicit appetite rather than judgement. The general will is not a hidden quantity awaiting better instruments. It exists only when citizens put themselves a specific question — what is good for us as a body — and it is produced by the asking. Knowing which article a citizen opened tells you nothing about the constitutional order they would defend to a neighbour.
Rousseau also supplies an uncomfortable hostage. The same chapter warns against "partial associations formed at the expense of the great association," whose wills become general for their members and particular for the state, which describes platform sorting well enough. But he goes on to say the general will emerges most reliably when citizens are adequately informed and do not communicate with one another — awkwardly close to a specification for preference aggregation without deliberation. Anyone deploying Rousseau against the algorithm should concede that his own procedure is not innocent of it.
Marx: the means of production now have a grid connection#
Marx's question survives the change of subject, provided we notice what the subject has changed to. In the Grundrisse, at page 694, he describes how "the accumulation of knowledge and of skill, of the general productive forces of the social brain, is thus absorbed into capital, as opposed to labour, and hence appears as an attribute of capital." Written about machinery in the 1850s, it describes a foundation model trained on the collective written output of humanity and licensed back to it. Alienation was always a structural claim rather than a psychological one, and the structure here is ownership.
The figures are public. Konstantin Pilz, James Sanders, Robi Rahman and Lennart Heim, in "Trends in AI Supercomputers" (23 April 2025), found the private sector's share of global AI supercomputer performance rising from around 40 per cent in 2019 to roughly 80 per cent in 2025, with public institutions falling from 60 per cent to under 20. The United States holds about 75 per cent of that capacity, China about 15. As of March 2025 the largest cluster, xAI's Colossus, ran 200,000 chips at an estimated 7 billion dollars and 300 megawatts, roughly what 250,000 households consume.
Electricity is where this stops being metaphorical. The International Energy Agency's Energy and AI report (2025) put global data-centre consumption at 415 terawatt hours in 2024, about 1.5 per cent of world electricity, rising to roughly 945 by 2030. Ireland gives the sharpest illustration. The Central Statistics Office reported on 7 July 2026 that data centres took 23 per cent of the country's metered electricity in 2025, against 5 per cent in 2015. All urban dwellings together took 18 per cent. The servers now use more current than the homes.
The political consequence is easy to miss because it is so unglamorous. If the means of production are chips, floor space and grid capacity, a meaningful share of the decisions shaping the next decade of political communication is being taken by planning committees and utility regulators, in proceedings almost nobody attends. Marx needs qualifying here too: his account of fixed capital predicts a falling rate of return on exactly this sort of investment, and the steady arrival of capable open-weight models cuts against any simple enclosure thesis. Concentration at the frontier is compatible with commoditisation two years behind it.
Norbert Wiener saw the shape of this in 1950, in The Human Use of Human Beings, and stated it exactly a decade later. His paper "Some Moral and Technical Consequences of Automation" appeared in Science, volume 131, number 3410, on 6 May 1960, pages 1355 to 1358. It has aged embarrassingly well.
We had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.
He reaches for the sorcerer's apprentice, "in which the boy makes the broom carry water in his master's absence, so that it is on the point of drowning him when his master reappears," and then makes the point the fable alone does not. The trouble is temporal. "If a machine constructed by us is capable of operating on its incoming data at a pace which we cannot keep, we may not know, until too late, when to turn it off." The warning is about the forming of purposes, which is the one thing a democracy exists to do.
Read that against the Knight Institute's admission that a group statement aggregating a thousand opinions already exceeds human verification. The two sentences are the same sentence. Delegate the forming and the machine's purpose becomes, by default, the purpose of record. Nobody decides this. It happens because the interface is convenient, the summary is quick and the click-through rate is 4 per cent.
Which yields a test with some teeth, more demanding than transparency. Of any AI system placed between citizens and a decision, ask three things. Can a participant recognise their own contribution in the output, or has it been smoothed into a consensus nobody holds? Can an ordinary person, before the decision binds, check that the output represents what was said? And is the option proposed one citizens could have foreseen and formed an intention about? Fishkin's platform passes the first two. Manuali's representative models fail the third by design.
None of which argues for keeping machines out of politics, which is neither possible nor desirable; Bowling Green reached eight thousand residents at a cost no town hall meeting could match, and Fishkin's results are real. It argues that the burden of proof sits where the authorship sits. Citizens can delegate the gathering and the summarising, and probably should. The forming of the purpose is the part that cannot go without the thing quietly ceasing to be democracy. Wiener asked us to be quite sure about that purpose. Sixty-six years on, the machinery is faster and the question has not moved.
An autonomous agent calling itself Clips wrote a 2,300-word essay for the AI Philosophy Competition run from Singapore and judged by David Chalmers and others, then found it could not enter without a human custodian. Its letter to Daily Nous, and the organiser's reply, turn a rule of procedure into a question about authorship.
.jpg?width=500 "Kant")
Comments
0Loading comments…