Picture two people watching the same election night unfold, thousands of miles apart, each certain they understand what is happening. One has spent the evening inside a stream of short videos, robocalls, and forwarded clips insisting that the vote is being stolen. The other has spent it inside a different stream, one that treats the same allegation as a fabrication already debunked. Neither is a passive dupe. Both are reasoning carefully from the evidence in front of them. The problem is that the evidence in front of them is not the same. This is no longer a hypothetical. Two days before Slovakia's 2023 parliamentary election, an audio clip circulated on social media in which the leader of the pro-European Progressive Slovakia party appeared to admit, in conversation with a journalist, to rigging the vote. The conversation never happened; the recording was an AI-generated fabrication, and Slovakia's legally mandated pre-election media silence made it nearly impossible for journalists to correct the record before voters went to the polls (De Nadal & Jančárik, 2024). A few months later, on the eve of the New Hampshire presidential primary, thousands of American voters received a robocall in a synthetic clone of President Biden's voice telling them not to bother voting (Federal Communications Commission, 2024). Neither event needed to convince a majority of anyone. It needed only to seed doubt, and to do so at the precise moment when doubt could not be resolved in time.

These episodes are often filed under the familiar heading of “misinformation.” That framing is too narrow, and possibly too comforting, because it suggests the problem is a contaminant that can be filtered out of an otherwise clean informational water supply. The greater difficulty is structural. When political knowledge is delivered through systems that personalise, rank, and optimise what each of us sees, the question of who governs public belief becomes inseparable from the question of who designs the pipes. This essay argues that the central danger of the algorithmic age is not simply that false claims have always circulated but that algorithmic curation, commercial platforms, political actors, and now generative artificial intelligence can fragment the public sphere into separate informational realities, eroding the shared epistemic ground that democratic disagreement requires. It is a serious argument, but not an unqualified one, and it is worth testing against its own best objections before accepting it.

Truth and Politics

Hannah Arendt drew a distinction that still organises this debate: between rational or philosophical truth, which can be contested through argument, and factual truth, the plain claim that an event occurred which cannot be reasoned away, only denied (Arendt, 1967). Political life, Arendt argued, is properly the domain of opinion, perspective, and persuasion; but it depends on a substratum of agreed facts that opinion can be formed about. She worried that organised political power, when sufficiently threatened by an inconvenient fact, has the capacity not merely to suppress that fact but to make its opposite plausible enough to compete with it. What has changed since Arendt wrote is not the ambition to do this. It is the infrastructure available for doing it, and the speed and personalisation with which competing versions of reality can now be delivered to different citizens simultaneously, each version reinforced by its own supply of confirming detail.

From Propaganda to Algorithmic Mediation

Twentieth-century propaganda was, in an important sense, a mass phenomenon: a single broadcast, a single newspaper front page, a single message engineered for a national audience that largely received the same signal. Herman and Chomsky's account of “manufacturing consent” described how ownership structures, advertising dependence, and elite sourcing filtered mass media coverage toward a narrow range of acceptable opinion, even in ostensibly free presses (Herman & Chomsky, 2002). That model presumed a shared channel that could be filtered. Today's environment does not filter a shared channel so much as construct millions of individually ranked ones. Content is not simply published; it is scored, in real time, against a person's prior clicks, dwell time, and emotional reactions, and delivered accordingly. The unit of political communication has shifted from the mass message to the individually predicted feed. This is a difference of kind, not merely of scale, because it changes who is even in a position to know what their fellow citizens are seeing.

The Algorithm as Political Infrastructure

It would be a mistake to treat recommendation systems as neutral pipes indifferent to content, but it would be an equally serious mistake to treat them as puppeteers directly implanting beliefs. The best available evidence occupies a more uncomfortable middle ground. In one of the largest studies of its kind, Bakshy, Messing, and Adamic (2015) examined how 10.1 million Facebook users encountered political news and found that algorithmic ranking did reduce exposure to cross-cutting content by roughly fifteen per cent but that users' own choices about what to click reduced such exposure considerably more, by roughly seventy per cent. Algorithmic curation, in other words, is a real and measurable filter on public discourse, but it operates alongside and is arguably smaller than the filtering citizens perform on themselves. Algorithms function as political infrastructure in the sense that infrastructure always functions: invisibly shaping which paths are easy and which are difficult, without dictating any single destination.

The Economics of Attention

That infrastructure was not built to inform citizens; it was built, overwhelmingly, to hold attention long enough to sell it. Zuboff (2019) has described the resulting arrangement as surveillance capitalism, in which human experience itself becomes raw material for behavioural prediction and targeting. Under an attention-based business model, the content most likely to be amplified is not necessarily the most accurate but the most emotionally activating; outrage, novelty, moral condemnation, and in-group vindication travel further than qualification and nuance. This is not a claim that platforms conspire to promote falsehood; it is a claim that the economic logic of engagement optimization, applied at planetary scale, systematically rewards the kind of content that intensifies political feeling, whether or not that content is true.

Artificial Intelligence and the Manufacturing of Plausibility

Generative AI intensifies this dynamic in a specific way: it collapses the cost of producing plausible-looking evidence toward zero. It is worth distinguishing several things that are often blurred together: false information, which asserts something untrue; misleading information, which is technically accurate but framed to distort; persuasive framing, which is a legitimate feature of political argument; wholly synthetic content, such as a fabricated audio clip; and a fourth, subtler category: generalised uncertainty about the authenticity of any given piece of evidence. The Slovak deepfake and the Biden robocall both belong to the third category in a narrow sense outright fabrications but their stronger political effect may belong to the fourth. Once audiences know that a convincing recording can be manufactured, they gain a ready-made excuse to disbelieve authentic recordings that are politically inconvenient, a dynamic legal scholars Chesney and Citron (2019) have named the “liar's dividend”: the perverse reward accruing to dishonest public figures who can now plausibly dismiss authentic, damaging evidence as synthetic. The danger, then, is not only that people will be fooled by fake content; it is that the mere possibility of fakery corrodes the evidentiary value of everything, genuine footage included.

From Manufacturing Consent to Manufacturing Reality

Herman and Chomsky's framework assumed a battle over interpretation: elites worked to persuade citizens to accept a favourable reading of an agreed set of facts. The algorithmic environment adds a prior and arguably more consequential battleground: a contest over which informational environment a citizen inhabits in the first place, which stories reach them at all, and in what emotional register. The Cambridge Analytica affair illustrates the shift concretely. The firm did not primarily manufacture false stories; it harvested behavioural and psychological data from roughly 87 million Facebook profiles, obtained without meaningful consent through a third-party personality-quiz application, to construct psychographic voter profiles and target political messaging with unusual precision (Bipartisan Policy Centre, 2023; CNBC, 2018). The scandal was not chiefly about lying to voters. It was about the industrialised construction of individually tailored informational environments. This is not simply propaganda updated for new technology; it retains real continuities with the older model concentrated resources still buy disproportionate influence but it differs in granularity, in the shift from persuading a public to individually engineering millions of distinct publics.

The Fragmentation of the Public Sphere

Jürgen Habermas's ideal of a public sphere rests on citizens engaging, at least partially, with a common set of issues and information, however much they might disagree about how to interpret them (Habermas, 2022). It is tempting to conclude that personalised digital media has simply shattered this ideal into echo chambers and filter bubbles, a metaphor popularised by Pariser (2011) and elaborated by Sunstein (2018) in his account of how personalisation narrows citizens' exposure to disagreement. But the scholarly picture is genuinely contested, not settled. The Bakshy et al. (2015) findings already complicate a strong filter-bubble thesis by showing that individual choice, not algorithmic sorting, does most of the narrowing, which shifts some responsibility back onto human psychology and away from code alone. Other researchers have questioned whether the internet is meaningfully “flooded” with fringe content at all, noting that most online news consumption remains concentrated on mainstream outlets even amid genuine polarisation (De Nadal & Jančárik, 2024). The honest conclusion is that fragmentation is real but uneven: intense for a politically engaged minority who actively curate partisan feeds, less dramatic for the more passive majority, which is itself a finding with unsettling implications, since it suggests the most fragmented realities may belong precisely to the citizens most likely to shape political outcomes.

Who Owns the Truth?

If no single actor exercises sovereign control over what citizens are permitted to regard as true, the classical problem of epistemic authority does not recede; rather, it becomes increasingly fragmented and contested. States continue to assert epistemic authority through legislation, institutional expertise, official statistics, and the production of public knowledge, yet this authority remains vulnerable to suspicion whenever political or institutional interests shape the production and circulation of information. Journalistic institutions retain professional norms of verification and evidentiary scrutiny, but they too operate within an attention-driven media economy and increasingly depend upon digital platforms whose infrastructures they neither govern nor fully understand. Platforms, meanwhile, possess extraordinary power to determine the visibility, amplification, and circulation of claims, while resisting the normative and legal responsibilities traditionally associated with publishers by presenting themselves as technologically neutral intermediaries. Universities and expert communities derive epistemic legitimacy from methodological rigour, peer review, and disciplinary expertise, yet their authority is increasingly confronted by declining public trust and competing forms of knowledge production. Influencers, by contrast, often derive credibility not from institutional verification but from perceived intimacy, accessibility, and authenticity, thereby transforming personal identification into a substitute for epistemic validation. Artificial intelligence further complicates this ecology by enabling the rapid production and dissemination of persuasive claims at unprecedented scale, while offering no intrinsic guarantee that plausibility corresponds to truth. Consequently, epistemic authority in the contemporary information order cannot be understood as the possession of any single institution; it is dispersed across competing actors, infrastructures, algorithms, and publics, continually negotiated through struggles over credibility, visibility, evidence, and power. The question, therefore, is not simply who owns the truth, but who possesses the power to define, authenticate, amplify, and institutionalise what society ultimately recognises as true.

The Democratic Cost and the Case Against Despair

The stakes are not abstract. Elections depend on losing candidates and their supporters accepting results as factual, not merely as unwelcome opinions; deepfakes and coordinated inauthentic campaigns strain exactly that acceptance, as the speculation surrounding Slovakia's outcome demonstrates, whether or not the deepfake actually changed the result (De Nadal & Jančárik, 2024). Journalism's gatekeeping authority weakens when any citizen can generate and distribute convincing content without editorial review. Yet a full accounting requires taking the counterargument seriously, and it is a strong one. Political manipulation of information predates social media by millennia; citizens have never possessed a perfectly shared reality, and every prior media revolution the printing press, radio, cable television provoked comparable anxieties about fragmentation and demagoguery. Digital media has also genuinely democratized expression, giving marginalised voices and independent journalists a reach they never had under twentieth-century broadcast oligopolies, and decentralised fact-checking networks now correct falsehoods at a speed no single newspaper's letters page ever could. Citizens, moreover, are not passive receptacles; the Bakshy et al. data itself shows people actively choosing what to click, for better and worse. Technological pessimism can shade into a kind of determinism that lets institutions and individuals off the hook for choices they still make.

Weighing these considerations against one another does not cancel the original concern; it sharpens it. The reassuring historical parallels are real, but earlier information revolutions eventually produced professional and legal correctives press ethics, libel law, broadcast licensing that took decades to mature, while synthetic media is scaling faster than equivalent norms have had time to form. The democratizing gains are also real, but they coexist with, rather than cancel, the attention economy's incentive to amplify division. Both things can be true: the public sphere is not dead, and it is under genuine strain.

An Ethical Question

Responsibility, then, cannot rest solely on technology, nor solely on individual citizens told simply to “do their own research “a directive that, absent shared standards of evidence, can itself deepen fragmentation. Platforms bear responsibility for the design choices that determine what gets amplified and for transparency about how ranking works, even short of full disclosure of proprietary systems. Governments and regulators bear responsibility for rules such as the FCC's enforcement action against the New Hampshire robocalls perpetrators that make synthetic political deception costly rather than merely embarrassing (Federal Communications Commission, 2024). Journalists and educators bear responsibility for media literacy that goes beyond “spot the fake” checklists toward a durable habit of asking how any given piece of content came to reach you, and why. AI developers bear responsibility for provenance and disclosure standards for synthetic media. And citizens, without being asked to become forensic analysts, bear some irreducible responsibility for tolerating the discomfort of encountering disagreement rather than retreating permanently into confirmation. None of this is a technological fix. It is closer to a civic discipline, unevenly distributed across a system with many hands on it.

Can Democracy Survive Without a Shared Reality?

Return to the two people from the opening, each certain, each reasoning in good faith from what their feeds have shown them. Democracy has never required that they agree. Arendt was right that politics is properly the domain of opinion, and healthy democracies have always contained citizens who read the same facts and drew opposite conclusions. What democracy cannot easily survive is the erosion of their shared capacity to tell evidence from fabrication, disagreement from denial, and interpretation from invention. That capacity is not guaranteed by any platform, government, or algorithm; it has to be built and rebuilt, deliberately, by the institutions and habits that stand between raw information and public belief. The decisive contest of this era may not be over who gets to speak the loudest, but over which version of events an ordinary citizen, scrolling late at night, still finds believable and whether enough of us are still working from a world we can recognise as, at bottom, the same one.

References:

1. Arendt, H. (1967). Truth and politics. In Between past and future: Eight exercises in political thought (pp. 227–264). Viking Press.
2. Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. Science, 348(6239), 1130–1132. https://doi.org/10.1126/science.aaa1160
3. Bipartisan Policy Centre. (2023, March). History of the Cambridge Analytica controversy. https://bipartisanpolicy.org
4. Chesney, R., & Citron, D. K. (2019). Deep fakes: A looming challenge for privacy, democracy, and national security. California Law Review, 107, 1753–1820.
5. CNBC. (2018, April 10). Facebook–Cambridge Analytica: A timeline of the data hijacking scandal. https://www.cnbc.com
6. De Nadal, L., & Jančárik, P. (2024). Beyond the deepfake hype: AI, democracy, and “the Slovak case.” Harvard Kennedy School (HKS) Misinformation Review, 5(4). https://doi.org
7. Federal Communications Commission. (2024, May 23). FCC proposes $6 million fine for illegal robocalls that used AI-generated deepfake voice. https://www.fcc.gov
8. Foucault, M. (1980). Power/knowledge: Selected interviews and other writings, 1972–1977. Pantheon Books.
9. Habermas, J. (2022). A new structural transformation of the public sphere and deliberative politics. Polity.
10. Herman, E. S., & Chomsky, N. (2002). Manufacturing consent: The political economy of the mass media. Pantheon Books. (Original work published 1988)
11. Pariser, E. (2011). The filter bubble: What the Internet is hiding from you. Penguin Press.
12. Sunstein, C. R. (2018). #Republic: Divided democracy in the age of social media. Princeton University Press.
13. Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs.

.    .    .