Welcome to the Brave New World of Prompted Reality
A growing share of the internet is assembled by chatbots and checked by nobody. Sites that look like news outlets are produced this way, and so are databases quoted in industry reports. Some of what they contain is accurate, some is outdated, distorted or invented, but all of it reads fluently. Once online, it is copied and cited, and the next generation of AI systems absorbs it.
The result is a new world founded on a mixture of truths, half-truths and lies. Each claim is calibrated to sound plausible, and repetition gets it accepted. Call this world prompted reality: a version of events that exists because someone prompted a model and published the answer.
Its building blocks are what one can describe recursive truths. A recursive truth is a claim that becomes a fact by being copied from one machine-made source to the next, until the copies look like confirmation. Inside prompted reality, the false ones among them become a working form of fact. Reviewers, regulators and reporters who want to operate in it will have to rely on its directories and numbers, whatever their origin.
How Prompted Reality Is Built
Understanding how prompted reality grows starts with the sites themselves, because the barrier to building one has all but disappeared and a working website can now be assembled in an afternoon by someone who has never written a line of code. Lovable, one of several services that turn chat prompts into working websites, said in June 2026 that its users were creating about a million new projects a week. Many are throwaway prototypes. Others go live as directories, trackers and reference pages, filled with whatever the chatbot found.
Databases about the news industry itself are compiled with the same tools, and their own documentation shows how unreliable those tools can be. One widely promoted database of the sector studied as part of the research for this article, built by a single author who uses AI research assistants, reports in its methodology note that in one audited batch of chatbot search results, about three in four findings carried fabricated or mismatched sources. Those were caught and discarded; a compiler who does not check would have published them. The same note acknowledges that the database has roughly doubled since its last full recheck of source links, so its newer entries rest on checks made when they were added.
At the industrial end, NewsGuard had counted 3,749 AI content farms in 16 languages by June 2026, up from 49 in May 2023. These sites dress as news outlets and publish mostly machine-written articles, with little human oversight and no disclosure. Some are written for machines alone. The Moscow-based Pravda network published an estimated 3.6 million articles in 2024, and a NewsGuard audit found that ten leading chatbots repeated its false narratives a third of the time.
The harm these sites do lies mostly in the way they blend accurate and inaccurate material, so that the reliable entries lend their credibility to the unreliable ones and readers have little reason to doubt either. Picture a directory of local news outlets compiled by a chatbot. It lists hundreds of genuine publications, which is why it looks trustworthy. It also keeps newspapers that closed years ago and names owners who have since sold up. Its audience figures are estimates presented as measurements, and community broadcasters with little online presence are missing. Anyone who checks a few familiar entries finds them correct and comes away reassured about a picture of the country that is essentially wrong.
Repetition turns such entries into recursive truths. A figure copied into a report, then into a second database, then into a chatbot answer looks independently confirmed. Every version, though, descends from one unchecked entry. Asking a second model to confirm the first only adds another copy.
The copies then train the machines. In a 2024 Nature paper, researchers showed that models trained repeatedly on generated content lose the rare cases at the edges of the original data, a process they called model collapse. In one experiment, text about medieval architecture had turned into a list of jackrabbits by the ninth generation. Each cycle makes prompted reality smoother and more confident.
A Research System With Nobody Inside
Science is the system built precisely to test claims against the world, and catching recursive truths before they spread is, in principle, part of its job. Academic publishing, however, is being drawn into the same loop, largely because the volume of work now exceeds what human reviewers can properly read. The International Communication Association received more than 12,000 individual submissions for its 2027 conference in Glasgow, more than double the 5,670 it reported for 2024. Two half-hour reviews for each would take 12,000 hours of reading.
As part of the research for this article, MediaPowerMonitor spoke to a dozen people working in academia. Their accounts converged on one claim: there is no longer a human being in the loop. An article is drafted by AI and submitted unchecked. A reviewer short of time passes it to a chatbot, which writes the report. The author then feeds the report into another model to produce the revisions. Human names appear on every form, while machines do the reading.
The figures bear them out. When Pangram Labs scanned the 75,800 peer reviews submitted to ICLR 2026, a major machine-learning conference, it classified about 21% as fully AI-generated. Some authors have adapted to this audience. In 2025, Nikkei Asia found hidden instructions in 17 preprints, set in white or microscopic type, telling AI reviewers to praise the work.
Some recursive truths are eventually exposed, usually when a person named in them or a careful colleague notices that a quotation or a reference matches nothing in the real world. The most striking recent case came from the top of European journalism. In March 2026, Mediahuis suspended Peter Vandermeersch, a former editor-in-chief of NRC, after NRC itself found fabricated quotations in 15 of his 53 posts. He said they came from AI summaries he had trusted to be accurate.
Such cases surface only when someone compares a claim with evidence outside the generated text. Most generated material never meets that test, and what escapes it becomes the working reality. That reality already shapes decisions. In October 2026, a South African labour court halted an arbitration because the arbitrator had relied on a Constitutional Court judgment that does not exist. A faulty media directory can decide which communities qualify for journalism support, and a wrong ownership map can point regulators away from the companies in control.
Operating in prompted reality increasingly means treating its falsehoods as facts. A researcher answers the review a machine wrote. A funder allocates grants from the directory it was handed. A journalist cites the figure every chatbot repeats. Accepting a recursive truth is faster than checking it, and few institutions pay for the checking.
Ways Out
Most of the researchers and academics interviewed for this article agreed that prompted reality is already taking shape. Asked what could be done about it, they offered answers that ranged from rebuilding the network itself to quietly protecting small pockets of verified knowledge.
The most radical among them argued that the internet in its present form cannot be repaired. For them, the only solution is to shut it down and build a new one. The proposal sounds extreme, yet it follows from the diagnosis. On today’s web, every open page can be scraped and fed into the next model. A correction published online is therefore copied, mixed and recycled like everything else, and it enters prompted reality as one more claim among many. Fixing errors one at a time cannot keep pace with systems that generate new ones faster than anyone can check them. Any rebuilt network would need different rules from the start. Knowing who published a piece of information, when, and on what evidence would be part of its architecture, and anything lacking that record would count as unverified by default. Such a project would take years and would depend on governments and on technology companies that profit from the network as it stands.
A second group proposed new knowledge spaces closed to AI systems and to the researchers who build them. These would be libraries, repositories or networks whose contents could not be scraped, summarised or fed into training data. Material deposited there would stay exactly as its authors wrote it, and anyone consulting it would read the original text, free of a model’s paraphrase. Fragments of the idea already exist, from news sites that block AI crawlers to scholarly archives behind institutional logins. The price is reach. Knowledge kept away from the machines is also kept away from the many people who now meet the world through chatbots and AI search, and those people would depend on prompted reality all the more. Closed spaces also rely on constant enforcement, because a single leak or scrape can carry their contents back onto the open web.
A third group spoke of private spaces where the truth could be preserved. The second proposal aims to keep knowledge out of AI’s reach; this one aims to keep a reliable record intact for the future. Such spaces would work like archives or seed banks. They would hold verified documents, datasets and testimony in their original form, together with a record of how each item was checked. Their value would grow with time. As recursive truths multiply, an untouched original becomes the only way to establish what was actually observed, and every later version can be measured against it. The difficulty lies in deciding who keeps these archives, who may open them and who pays for them, since a private record of the truth helps only those allowed to consult it.
MPMonitor, the site publishing this article, is pursuing a fourth approach, one that keeps verified information on the open web and makes its origins visible to anyone who looks. Its Source Passports, a project of the Media and Journalism Research Center (MJRC), form a public verification system of provenance. Each passport is a public record card for an organisation that shapes what people read, watch and hear, such as a news outlet, a political party, a regulator, a platform or a funder. It answers three questions in one place: who owns the organisation, who pays for it and where it publishes.
Every fact in a passport names its source and carries a status label. A fact is marked verified only after a researcher has checked it against that source. Others are shown as self-declared, pending, disputed or not disclosed, so readers can see exactly what has been confirmed and what has not. Software helps locate documents, while researchers decide what counts as fact. Each change is dated in a public history, organisations and readers can dispute any entry, and the whole register is released as open data. On 4 November 2026 it held 4,348 passports. In a world of recursive truths, the design offers readers a fixed point to check claims against.
However, this approach is slow by design, and it is expensive. To keep its output at this standard, MJRC has had to double its human editing capacity. It does not know how long it can sustain that if demand for real, checked truth does not grow.
None of these solutions will arrive quickly, and some may never arrive at all. Until they do, people may have to accept that they have to live within this brave new world of prompted reality and adjust to it. Asked what that adjustment would look like, the researchers and academics interviewed for this article described a few possible futures, which differ sharply in what they would mean for society.
In the first future they described, checked information becomes a premium product. Organisations that can pay for verified data, through subscriptions, licences or their own researchers, would work from an accurate picture of the world. Everyone else would rely on whatever chatbots and free databases return. Accuracy would become a privilege, and the gap between the informed and the merely answered would follow familiar lines of wealth and institutional power. Public debate would split between people working from different versions of the same facts. This future would keep the truth alive while making it unequal.
Other interviewees pictured a future, in which provenance travels with every claim, much as origin and ingredients travel with food. Readers would grow used to asking where a figure came from and whether anyone checked it, and an unlabelled claim would be treated as a probable recursive truth. Professions that depend on facts would build the habit into their routines, as some courts have already begun to do by requiring lawyers to disclose their use of AI in filings. This future would ask the most of everyone in time, money and attention. In return, prompted reality would stay visible and open to challenge, and a shared public record would survive.
In another future, as interviewees described it, institutions simply learn to run on prompted reality, accepting its figures because nothing better is available in the time they have. Grants would be awarded, markets sized and policies drafted on numbers that many suspect and few check, with errors discovered only when they cause visible harm. Public arguments would increasingly be settled by which version the most systems repeat. This is the easiest future to drift into, because it requires no decision at all. It would also weaken accountability, since nobody could say with confidence what had actually happened, and it would hand the power to define facts to whoever can flood the system with content.
Finally, the fourth future foreseen by our interviewees is one in which the past becomes a scarce resource. Archives, old newspapers and datasets created before generative AI spread would gain value as the last material known to be free of machine contamination, much as steel made before the first nuclear tests is still sought for some radiation-sensitive instruments. Historians, researchers and AI developers would compete for access to these records. The pre-AI record would survive as a reference point, but whoever controls access to it would hold considerable power over what can be verified, and knowledge produced after the shift would remain hard to trust.
The four futures can unfold side by side, and early signs of each are already visible. For society, the second would be the best outcome and the third the worst. The first and the fourth fall in between: each preserves the truth for some while leaving most people inside prompted reality. Which one prevails will depend partly on whether people and institutions are willing to pay for the difference between prompted reality and the world it imitates.