How Public Knowledge Is Made Now

I

Asked what they expect of news produced with the help of artificial intelligence, audiences across forty-eight markets reply in a consistent pattern. They expect such news to be cheaper to make, by a net margin of twenty-nine percentage points, and more up-to-date, by sixteen. They expect it to be less transparent and less accurate by eight points each, and less trustworthy by eighteen. The Reuters Institute Digital News Report 2025, on which these figures rest, surveyed nearly a hundred thousand respondents across the survey set. A supplementary Reuters Institute report published in October 2025 in six countries found the equivalent figures, on the four dimensions the supplementary report tracks, had hardened rather than softened: thirty-nine, twenty-two, minus eight, minus nineteen.

These are not the views of a public that does not understand what is happening. They are the views of a public that perceives a trade-off and has not yet been given the public language to discuss it. Fifty-eight per cent of respondents worldwide report concern about telling real news from fake news online. In Australia, one in five say they are comfortable with AI-produced news that has had little human oversight, up four points on the year before; fifty-four per cent say they are uncomfortable with such news. The figures track in the same direction across the survey set. The audience holds the trade-off in mind, and the trade-off is sharpening.

Walker Briefing is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

What the audience perceives describes a real change in the conditions under which the news the audience reads is produced. The next part of the article describes that change.

II

The change runs through several institutions of public-knowledge production at once. Four are examined here: the architecture of search, the institutional capacity of newspaper journalism in Australia, the new economy of paid-newsletter writing, and the production of academic and parliamentary research.

First comes the search. In a study published by the Pew Research Center in July 2025, the click-through rate on conventional search results fell from fifteen per cent without an AI Overview present to eight per cent with one. Users clicked into sources cited inside the Overview itself one per cent of the time. Browsing sessions ended on the search page sixteen per cent of the time without an Overview, and twenty-six per cent with one. The data was drawn from sixty-eight thousand unique searches by nine hundred US adults in March 2025. Independent corroboration has come from multiple methodologies. Ahrefs, having analysed approximately three hundred thousand keywords, reported in February 2026 that the position-one click-through rate for AI-Overview-triggering queries had fallen to 1.6 per cent in December 2025, a fifty-eight per cent reduction once general trends were controlled for. Authoritas, in submission to the UK Competition and Markets Authority in April 2025, measured a 47.5 per cent click-through decline on desktop and 37.7 per cent on mobile when AI Overviews were present.

Google’s Head of Search, Liz Reid, has argued repeatedly between August 2025 and April 2026 that AI Overviews reduce low-value bounce clicks while leaving high-value engagement clicks intact. Google measures retention to its own surface; publishers need a metric of traffic to theirs; the company has not published the query-level data that would resolve the dispute.

The second is institutional newspaper journalism in Australia. IBISWorld data, cited in the Media, Entertainment and Arts Alliance’s 2024 submission to the federal government’s News MAP consultation, tracks Australian newspaper publishing employment from 23,472 in 2010-11 to 12,531 in 2023, a forty-seven per cent decline over twelve years. The trajectory has continued. IBISWorld’s 2025 industry analysis records sector revenue contracting at a compound annual rate of 5.2 per cent across 2020 to 2025, with a 6.8 per cent fall in 2025-26 alone, and a similar contraction in the number of businesses in the industry. The named workforce reductions through 2024 are consistent with the trajectory. Nine Entertainment announced around two hundred redundancies in 2024, including between seventy and ninety in publishing, accompanied by a national strike of around five hundred journalists. Seven West Media, News Corp, and Australian Community Media announced parallel cuts. More than a hundred regional and community newspapers ceased printing across Australia between 2020 and 2025, on top of the 106 local and regional newspaper titles that closed between 2008 and 2018.

Next we have the paid-newsletter economy. Substack announced in March 2025 that it had passed five million paid subscriptions. Secondary platform-statistics reporting later put total active subscriptions at about thirty-five million as of September 2025. The platform also reports more than seventeen thousand writers earning subscription revenue through it, while its top ten authors collectively earn more than forty million US dollars annually and more than fifty individuals reportedly earn a million dollars or more from paid subscriptions alone. These platform-reported figures are not independently audited in the way that newspaper circulation figures historically were. A Series C round closed in July 2025 valued the company at $1.1 billion. The platform’s economic model tilts towards individual analytical voice rather than the institutional capacity for sustained reporting that newspapers historically funded: court reporting, council reporting, regional reporting, investigative work. Newsroom-style operations on the platform remain rare and small-scale relative to the institutional capacity newspapers historically provided, and the underlying economics weight returns toward individual subscriber relationships rather than sustained collective reporting. The new capacity sits beside the contracting institutional capacity rather than replacing it. The paid-newsletter economy carries a generalisability constraint of its own. The Reuters Institute Digital News Report finds that the share of audiences paying for any online news, across twenty richer countries, is stable at eighteen per cent and is not growing.

Finally we see academic and parliamentary research production. By late 2023, approximately thirty per cent of researchers reported having used AI tools to help write manuscripts. The figure, drawn together by Gray at University College London from multiple surveys including Nature’s, captures self-reported AI assistance of all kinds, from grammar to substantive drafting. A separate measurement, by Liang and colleagues using a population-level word-frequency method to estimate the share of LLM-modified content in published abstracts, returns figures that vary sharply by field: up to about twenty-two per cent in computer science papers, and up to about nine per cent in mathematics papers and in papers in the Nature portfolio. The two methods measure different signals, one tracking visible markers of AI use, the other tracking shifts in word frequency at population scale, and the picture they produce together is one of substantial and field-variable adoption rather than of any single rate. Major academic publishers, including Wiley, Elsevier, and Nature, have policies as of 2024 and 2025 requiring disclosure of substantive AI use; universities are still defining theirs. The parliamentary side carries thinner data. There is no published equivalent of the academic surveys for parliamentary research staff in the United Kingdom, Australia, Canada, or comparable Westminster jurisdictions. The thinness of the public record is itself a finding. The institutions whose research production most directly shapes legislative deliberation have produced the least public information about how that production is changing.

The four threads describe different sites of change. They share an architecture. In each, the conditions under which public-facing knowledge is produced have shifted within the last decade, and in some cases within the last two years. The shift has happened ahead of the public conversation about it.

III

The conditions under which a public-facing analytical piece is produced in 2026 are not the conditions under which a comparable piece was produced in 2016. The shift is the same shift the previous section described, encountered now from inside the production of a single article rather than across institutions in aggregate.

The first part of the shift is in research. A writer working on a public-affairs question in 2016 began with a list of sources returned by a search. A writer working on the same question in 2026 begins with whatever the search returns. What the search returns now, on a meaningful and growing share of queries, is an aggregated summary written by an artificial-intelligence model with the sources listed below it. Some writers will follow the sources to their primary form before drawing on them. Others will not. Either way, the encounter with the question has been pre-shaped by an architecture the writer did not design and most readers do not see.

The next shift is in the reference frame the writer can draw on. An analytical piece of public-affairs writing rests, often without acknowledging it, on a substrate of reported fact produced by other people: court reporters, council reporters, regional reporters, investigative teams, the slow accumulated work of newsroom institutions. The Australian data above is the relevant index of how that substrate has thinned. A writer in 2026 working on an Australian regional question has fewer reported court appearances available, fewer council meetings covered, fewer regional angles to draw on than a writer working on the same question in 2016. The shape of the analytical piece is unchanged. The substrate beneath it has changed.

Then we have the relationship between the writer and the reader. The paid-newsletter economy has put many analytical writers in direct economic relation with their readers, without the institutional intermediary that the newspaper or magazine once provided. The relation is more intimate. The disclosure obligations that apply to it are different in kind from the disclosure obligations of institutional journalism, and the public norms that would specify them are not yet formed. The reader who reads an analytical piece in 2026 reads the writer, not the institution.

These three changes are not separable. They constitute a single shift in what producing a public-facing analytical piece in 2026 actually involves. The shift is real, it is structural, and it is mostly invisible to readers because the output looks like the output always looked. An article in 2026 reads like an article in 2016. The conditions of its production do not.

IV

Architectural transitions in public knowledge have happened before. Two are useful for locating the present moment.

The maturation of wire-service journalism, driven first by the telegraph and then complemented by the telephone, restructured newsroom workflow and political communication between the 1850s and the 1920s. The wire bureau was a new institution, the real-time deadline was a new pressure, the foreign correspondent’s relationship to the editor was a new geography of work. Public conversation about what these changes meant for the production of news lagged the practical adoption by roughly a generation. This is a structural lesson. When the architecture of news production changes, it changes before the public language for discussing the change has been built.

The shift from paper to electronic clinical records, between roughly 2000 and 2020 depending on jurisdiction, changed what doctors did in consultations, what was recorded and what was not, who could see the record, and how the record could be searched. The change carried ethical weight. The relation of trust between patient and clinician was reshaped by an architecture that neither patient nor clinician had designed. Public conversation lagged the practice change by roughly a decade and remains incomplete in 2026. The lesson is that architectural changes to relations of trust raise ethical questions that the people inside the relation cannot fully see while the change is happening, and that the public conversation about those questions has to be made rather than awaited.

Together the two comparators locate the present moment. The architecture of public-knowledge production is changing in the way the architecture of news production changed across the late nineteenth and early twentieth centuries, and is reshaping a relation of trust in the way clinical records reshaped the relation between patient and clinician. The lag between the practice change and the public conversation does not close itself.

V

The architecture of public knowledge production is upstream of what the reading public can read, what writers can produce, and what kinds of public reasoning a society has available to it. Downstream debates, about misinformation, about polarisation, about institutional trust, about media bias, sit on top of the architectural conditions. Changing the downstream debates without changing the architectural conditions is rearranging furniture in a room whose walls are moving.

The architectural conditions are upstream because they shape the possibilities and constraints within which all of the downstream conversations are conducted. A reader who finds news through an AI-mediated summary, a writer who works on a substrate of thinning institutional reporting, a researcher who drafts within an emerging set of disclosure norms, a citizen who reads the writer rather than the institution: each is participating in a public conversation whose conditions of production were not chosen and have not been collectively examined.

The change is being made by entities whose decisions are mostly local and whose collective effect is structural. Platform companies whose product changes alter the economics of every publisher dependent on search and social distribution. Employment markets that have hollowed parts of the institutional reporting capacity. Individual writers whose paid-newsletter ventures are now part of the new architecture. Academic institutions still defining their disclosure norms. AI tool providers whose adoption curves are running ahead of the policy conversation that would govern them. None of these decisions, taken individually, would warrant the description “structural change”. Taken together, that is what they are.

The lag between practical adoption and public conversation, which the historical comparators locate as a structural condition rather than as an oversight, is the present condition. The architecture is changing without a public conversation that matches the change. The conditions under which public-facing analytical writing is produced are upstream of every downstream argument about what the public knows, trusts, or believes.

The conversation that the present moment asks for is not technical, and it is not procedural. It is structural. The conditions under which a society’s public-facing analytical writing is produced are not given facts. They are produced by decisions, taken in different places by different parties for different reasons, whose collective effect is the architecture within which public reason is conducted. The conversation has to be made.

This article is Part Five of The Available Remedies, a series on the Walker Briefing examining structural responses to the conditions diagnosed in the preceding Zhōng Yōng series. The full archive is available at bfwalker.substack.com.

Walker Briefing is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.