The Parasite and Its Host
The press is dying from a position of historic weakness, not historic strength. This needs to be said at the outset, because the argument that follows is not a defence of journalism as it currently exists. It is an argument about what journalism does that nothing else can do — and what happens when that function disappears.
The numbers are stark. United States newspaper advertising revenue peaked near fifty billion dollars in 2005 and fell below ten billion by 2021 — a decline of eighty percent in sixteen years. More than three thousand newspapers have closed since 2005. Nearly seven thousand journalism jobs were lost in 2025 alone, roughly eight percent of the remaining workforce. Seventy million Americans now live in communities with no local news coverage at all. [1]
Into this weakened landscape, AI has arrived — and here is the irony that frames everything that follows. The large language models that are now being used as information sources were trained substantially on journalism. The news articles, investigative reports, analyses, and opinion pieces that journalists produced over decades constitute a significant portion of the training corpora. The AI is not an alternative to the press. It is a derivative of the press — a system that has ingested journalism’s outputs, repackaged them without attribution or compensation, and is now diverting the audience and revenue that journalism needs to survive. [2]
This is a parasitic relationship in the precise biological sense: the parasite depends on the host for sustenance while simultaneously weakening it. If the host dies — if the economic model that sustains original reporting collapses entirely — the parasite does not gain independence. It loses its source of nutrition. The AI will continue to produce fluent, confident answers, but those answers will increasingly be summaries of summaries, reflections of reflections, each generation further from any direct encounter with the world.
The migration is real, even if it is still early. The Reuters Institute’s 2026 Digital News Report finds that ten percent of online news consumers globally now use AI chatbots for news weekly, up from seven percent the previous year, with fifteen percent of those under twenty-five doing so. [3] In the United States, Pew Research reports that nine percent of adults get news at least sometimes from AI chatbots. [4] These numbers are modest. But they are growing in a context where trust in traditional media continues to decline — and where the economic base of journalism is already critically damaged. The question is not whether the migration will happen. It is whether anyone will notice what is lost when it is complete.
The Distinction That Changes Everything
The deepest structural difference between journalism and AI is not about bias, transparency, or accountability — though all of those matter. It is about the relationship between the information source and the world it claims to describe.
A journalist goes to a courtroom and watches a trial. She interviews witnesses. She visits the disaster zone. She sits in the room where the decision is made. She encounters the world directly — with all the limitations of a single human perspective, all the distortions of her training and her employer’s interests, all the imperfections of memory and attention. But she was there. Her account is a first-order representation of reality: imperfect, partial, ideologically inflected, but grounded in direct encounter.
A large language model has never been anywhere. It has read everything that has been written — or a vast subset of it — but it has never left its room. It cannot tell you what the air smelled like after the shelling, what the witness’s face looked like when she changed her testimony, what the silence in the committee room felt like before the vote. It can tell you what has been said about these things. It is a second-order representation: a summary of accounts, not an account of the world. [5]
This distinction — between witness and echo — is not a minor technical point. It is the structural fact on which the entire argument turns. Everything else in this essay — the hidden bias, the consensus reproduction, the concentration risk — is a consequence of this primary distinction. The AI’s limitations are not bugs to be fixed in the next version. They are inherent in what the system is: a machine that processes descriptions of the world, not a machine that encounters the world.
And yet the presentation obscures this. The model speaks in the present tense, with confidence, about things it has only encountered in text. It does not say: according to the sources I was trained on, the situation was as follows. It says: the situation is as follows. The epistemic distance between the model and the world is collapsed in the presentation, even though it is very much present in the reality. [6]
The Calm That Disarms
If the primary distinction is between witness and echo, the primary danger is that the echo sounds more authoritative than the witness.
A newspaper has a masthead with a known political history. It has a proprietor with identifiable interests. It has an editorial line that can be named, contested, and compared with alternatives. The bias is, in principle, visible — which means it can be corrected for. A reader who knows the outlet leans a particular direction has already built in a mental adjustment. The distrust is legible: you can point to what you distrust and why.
An AI system has none of these visible markers. Its training data is not published. Its rater guidelines are not disclosed. Its system prompt is invisible. The angle — and there is always an angle, because any system trained on human-generated text carries the assumptions of the people who built and evaluated it — is hidden behind a presentation of calm, comprehensive neutrality. [7]
This is not a metaphor for danger. It is a description of a cognitive mechanism. The perception of neutrality suppresses the epistemic vigilance that would otherwise operate. When a source presents itself as having no position, the reader’s critical faculties relax. The bias becomes stronger precisely because it is invisible — not despite the calm presentation, but because of it. The newspaper reader who knows the outlet is biased is better protected than the AI user who believes the system is neutral.
The Consensus Machine
The distinction between witness and echo has a second consequence that compounds the first: the model does not merely reflect what has been written — it reflects what has been most written.
Large language models are trained to predict what text is likely to follow given a preceding context. The training signal rewards outputs that resemble the distribution of text in the training corpus. The corpus is dominated by text that reflects the consensus views of the institutions and communities that produce most text. The model therefore has a structural tendency to reproduce consensus — to give the answer that most sources would give, weighted by the authority those sources carry in the training data. [8]
This is not a flaw in the model’s design. It is a consequence of what the model is. But it has a significant implication: the model is systematically better at reproducing established consensus than at identifying the places where established consensus is wrong. And the places where established consensus is wrong are precisely the places where the most important epistemic work needs to be done.
New truths almost always start as the unpopular view. Heliocentrism, germ theory, continental drift, the link between smoking and cancer — each was, at the time of its emergence, the position of a small number of people arguing against a larger consensus. [9] A system that is structurally weighted toward consensus reproduction would have been, at each of these moments, on the wrong side. Not because it was malicious, but because it was doing exactly what it was designed to do: reflecting the weight of existing text.
The press, for all its failures, has at least the possibility of the maverick reporter, the contrarian editor, the outlet that publishes the unpopular finding. The AI has no such possibility built into its architecture. Its contrarian outputs must be actively extracted by users who already know to ask for them.
The Geometry of Correlated Failure
The concentration problem follows directly from the distinction between witness and echo. Many journalists, deployed across many locations, encountering the world from many angles, produce errors that are diverse. They argue with each other. The mistakes of one outlet are, at least sometimes, corrected by another. The errors are not random, but they are distributed — and distributed errors are partially self-correcting.
Two or three dominant AI models, each trained on similar corpora, each evaluated by similar rater guidelines, each deployed by organisations with similar cultural and legal contexts, produce errors that are correlated. When the consensus is wrong, all the models are wrong in the same direction. There is no internal correction mechanism, because there is no internal diversity. [10]
This is a statement about the geometry of error in information ecosystems. A distributed system — many witnesses, many perspectives, many editorial lines — produces a noisy but partially self-correcting picture of the world. A concentrated system — a small number of echoes, all reflecting the same corpus — produces a clean, confident, and potentially systematically wrong picture. The cleanliness is not a sign of accuracy. It is a sign of correlated failure.
The historical precedent is not the press. It is the encyclopedia. The great encyclopedias of the eighteenth and nineteenth centuries were comprehensive, authoritative, and deeply influential. They also embedded the assumptions of their editors in ways that were invisible to their readers, and those assumptions shaped what generations of educated people believed about the world. The errors were not corrected by the encyclopedia itself. They were corrected by the accumulation of other sources — by journalists, investigators, and witnesses who encountered the world directly and reported what they found.
The Instrument That Talks Back
There is something genuinely new here, and it deserves acknowledgment. You can argue with an AI. You cannot argue with a newspaper — it prints its opinion and that is the end. But you can push an AI: make it steelman the position it initially dismissed, make it identify the assumptions embedded in its first answer, make it argue the opposite case — and watch it change its response.
This is a real capability. The model’s outputs are not fixed. They are responsive to the framing of the question, the structure of the conversation, the explicit demands of the user. A user who pushes back will get a different and often richer response than a user who accepts the first answer. [11]
But this only works if you keep pushing. The model’s default output is the consensus answer, delivered with confidence. The alternative outputs — the minority positions, the contested claims, the acknowledged uncertainties — are available, but they require active extraction. They do not surface spontaneously. A user who accepts the first answer gets all the danger of a concentrated, consensus-weighted echo chamber, with none of the benefit of the interactivity.
This creates a structural inequality. The users who are already epistemically sophisticated — who know to push back, who know what questions to ask, who have enough background knowledge to recognise when the consensus answer is incomplete — will extract genuine value. The users who are not already epistemically sophisticated will receive the default output: the consensus view, delivered with the authority of apparent neutrality. The tool that is most valuable to those who need it least, and least valuable to those who need it most.
What Could Be Built — and What Happens If It Is Not
Epistemic provenance labelling. If the model is an echo, it should be labelled as one. Users who know that a model was trained primarily on English-language text, evaluated by raters operating under particular cultural guidelines, and deployed with invisible system prompts — are in a better position to apply appropriate scepticism. This does not require disclosing proprietary information. It requires disclosing the structural properties of the system: what it was trained on, how it was evaluated, what it is designed to do and not do. The failure mode is that disclosure becomes a legal formality that users ignore. The constraint is that the institutions capable of requiring disclosure are the same ones struggling to regulate social media.
Adversarial defaults. The interactivity that makes AI genuinely valuable could be built into the default interface rather than requiring active user extraction. A system that, by default, presents the minority position alongside the consensus, flags contested claims, and identifies the assumptions embedded in its own answer — would distribute the epistemic benefit more widely. The failure mode is that users find the friction annoying and migrate to systems that give clean, confident answers. The constraint is that commercial incentives currently favour the frictionless experience.
Plural model ecosystems. Rather than a small number of dominant models trained on similar corpora, a healthier information ecosystem would include models trained on deliberately different corpora, evaluated by different communities, and designed to disagree with each other in structured ways. The error-correction mechanism of the old press — distributed mistakes that argue with each other — could be partially reconstructed at the model level. The failure mode is that users, confronted with disagreement, retreat to the model that confirms their existing beliefs. The constraint is that genuine diversity requires resources currently concentrated in organisations with incentives toward convergence.
Funding the witnesses. None of the above matters if the primary source of first-order information — journalism, investigation, the direct encounter with the world — ceases to exist. The AI cannot improve its relationship to reality by becoming more sophisticated. It can only improve if the witnesses whose accounts it summarises continue to exist and continue to do their work. This means the economic model that sustains original reporting must be rebuilt — not as a subsidy to a dying industry, but as infrastructure for the information ecosystem that AI itself depends on. The failure mode is that funding becomes a mechanism for political control. The constraint is that no one has yet designed a model that sustains journalism without either commercial distortion or state capture.
Now consider what happens if none of these interventions are implemented. The trajectory is not speculative — it is the extrapolation of trends already underway. The press continues to lose revenue and staff. The AI continues to be trained on the diminishing output of a shrinking journalistic workforce. The models become increasingly confident summaries of an increasingly thin evidence base. The public migrates further toward AI for information, accelerating the economic collapse of the institutions that produce original reporting. Within a generation, the primary information infrastructure of democratic societies becomes a system that has never encountered the world directly, summarising the work of witnesses who no longer exist, for an audience that has forgotten the difference. [12]
This is not an apocalyptic prediction. It is a description of what happens when a parasitic relationship runs to completion: the host dies, and the parasite — unable to feed itself — produces outputs of declining quality while maintaining the same tone of calm authority. The danger is not that the system will suddenly fail. The danger is that it will continue to function, fluently and confidently, long after the substance beneath the fluency has been hollowed out.
The Question That Will Not Close
The migration from press to AI is not, in the end, a migration from one information source to another. It is a migration from a system that — however imperfectly, however compromised by commercial pressure and ideological capture — maintained a direct connection to the world, to a system that is, by construction, a reflection of what has already been said about the world.
The press was never the idealised institution that its defenders sometimes invoke. It was concentrated, polarised, commercially driven, and often wrong. But it was wrong in diverse ways, and it was grounded — however imperfectly — in the encounter with reality. The AI is wrong in correlated ways, and it is grounded in nothing but text.
The protection is not a more trustworthy source. The protection is learning to question any source — to ask whether it is a witness or an echo, whether it encountered the world or merely summarised what others have said about it, whether its confidence is earned or merely performed. That capacity cannot be outsourced to a machine. It is the one thing that must remain human.