In this article
- The hierarchy that worked without anyone teaching it
- The flattening of fluency
- The criteria being built by feel
- Where it hurts
- The asymmetry that worries me
Definitions · Referencias · También te interesa · En otros sitios
For years, glancing at the packaging was enough. A New York Times headline wasn't confused with a blog, a paper in Nature wasn't confused with a Twitter thread, a medicines-agency statement wasn't confused with an uncle's forward. Each rung brought its own protocol, its known filter, its name to hold to account. Generative AI hasn't torn down that hierarchy. It has flattened it in the one thing I could see without effort: the surface. Any model output now has the same smoothness as an editorial worked over for a week, and the marker I used to decide what's worth something has stopped being where I had it. Nobody yet knows where to put it back.
The hierarchy that worked without anyone teaching it
Almost nobody got a class on how to rank sources. And yet most people operated with three fairly clear rungs, learned by rubbing up against them. At the top, the reference press: professional newsroom, an editorial leadership that answers, corrections, civil liability for defamation. Validation happened before printing. The reporter verified, the editor reviewed, legal filtered, and the brand stood as proof that someone had looked twice.
Below that, the academy. Here the judgment didn't rest on one person: two or three anonymous reviewers opined before publishing, and afterwards the community reacted with citations, replies and fights that lasted years. On the third rung, the official sources —state bulletins, regulators' statements, reports from international bodies—, where whoever signs answers with their career and sometimes with their assets.
Worth saying before someone accuses me of nostalgia: none of the three was infallible, and the three failed loudly. The reference press produced the Jayson Blair case, the New York Times reporter who fabricated or plagiarized dozens of pieces until his resignation in 2003, and before that it had run catastrophic coverage like the one on weapons of mass destruction in Iraq. The academy birthed the fraud of Jan Hendrik Schön at Bell Labs, whose investigative committee confirmed in September 2002 the falsification or fabrication of data in sixteen cases, and the study by Andrew Wakefield that in 1998 invented a link between the MMR vaccine and autism. The official sources signed biased reports, from the Pentagon Papers to the memos that legitimized torture.
What the hierarchy guaranteed was not accuracy. It was something less showy. When something went wrong, there was a name behind it, a procedure for rectifying and a possible sanction. A structure of responsibility, not a promise of truth. It's that structure —not the truth, which was never assured— that's been left with no recognizable packaging.
The flattening of fluency
What generative AI touches is not the facts contained in the texts. It touches a surface property almost nobody talked about because almost nobody knew they were using it: the visual heuristic of quality, that mental rule that infers a text's reliability from its look. Before the mass model I separated professional press from amateur blog by formal signals —the page layout, the typography, the register, the careful spelling, the byline, the mention of sources. None of those signals guaranteed quality. They correlated with it, which is not the same, and that correlation was enough for me.
The language model produces text with all those signals in place and the finish flawless. Spelling without a single stumble, register fitted to the genre, clean transitions, vocabulary that doesn't grate. The text of an untrained user now meets the visual heuristic exactly as well as that of a veteran journalist. The amateur blog of 2010 gave itself away with its typos and its clumsy syntax. The one of 2026, fed by a model, no longer gives itself away by anything.
So the visual heuristic has stopped filtering, and whoever depended on it has been left out in the open. I'm not talking about the information professional, nor the academic, nor the analyst who has other mechanisms. I'm talking about almost everyone, who distinguished sources by the look of the text and by nothing else. The loss doesn't affect an expert minority. It's wide, and it's silent, because few knew they leaned on that marker until it stopped holding them up.
The criteria being built by feel
If fluency no longer discriminates, something will have to, and here the discussion is wide open, with nothing settled. The most worked-out idea is to trace origin: being able to check who produced a content, with what primary sources and through what editing chain. The C2PA initiative (Coalition for Content Provenance and Authenticity) works along that line, founded in February 2021 by Adobe, Arm, BBC, Intel, Microsoft and Truepic, which seeks to mark cryptographically the origin and the successive modifications of images and videos. The European AI Regulation (EU 2024/1689) advances along the legal route and obliges the marking of machine-generated content, with the transparency obligation of its article 50 fully enforceable from 2 August 2026. Marking origin, mind you, only works if almost everyone marks, and that broad adoption doesn't yet exist.
There's a second route, more artisanal: cross-checking independent sources against each other. It's the old triangulation of the investigative journalist, who collates a claim with several sources that have different incentives, now carried over to the ordinary reader. If three voices that gain different things coincide, the probability of error drops. The problem, besides being humbling, is one of time: triangulating takes effort and almost nobody invests it to decide whether a headline is true. And there's a fine trap on top. Models trained on similar corpora can manufacture an illusory consistency: three answers that seem to confirm one another in fact drink from the same well, so what the reader takes for triangulation is one single voice repeated three times with three faces.
There remains the third route, the most uncomfortable: certifying that behind a text there's a person and not a model, with cryptographic verification that separates what a human wrote from what a machine generated. Sam Altman's Worldcoin project proposes iris scanning for this, while Adobe and Microsoft work on verifiable credentials. They're experiments, and they drag along reasonable resistances over privacy and concentration of power that shouldn't be dismissed as paranoia.
The three proposals are worth half each and none on its own replaces the hierarchy that has collapsed. What would turn them into something resembling a system is coordination among them, and that coordination has no owner. No government has taken it on, nor industry, nor the academy. Each pushes its own piece and glances sideways at the next.
Where it hurts
The harm is not shared evenly. It concentrates, and it concentrates first in health policy. The covid pandemic was an open-air laboratory: the reader received at once information from the WHO, from national agencies, from the general press, from the specialized press, from individual medical blogs and from professional-looking accounts that answered to no one. When those voices contradicted each other, they had nothing to order them by. What in the earlier framework would have been resolved by deference to the sources with a protocol turned into a tribal fight over masks, vaccines and meters of distance.
In political opinion, the line separating informed analysis from partisan comment and from manufactured propaganda blurred as soon as the three things acquired the same smoothness. The Reuters Institute's Digital News Report 2024, based on a survey of more than 95,000 people in 47 markets, documented that the share of adults who avoid the news sometimes or often rose to 39%, three points above the previous year and the highest level since the report exists. People declare they're fleeing because of the overwhelming volume and the wear as much as because of the difficulty of telling what's worth it; the exhaustion before a flood that can no longer be ordered is part of the same picture.
Then there's informal legal and financial advice, where forums, networks and newsletters produce texts with the look of a professional opinion. On one same plane coexist the practicing lawyer, the law student, the well-read amateur and a model's output, and the decision the reader is going to make —sign, invest, sue— rests on a source they couldn't place even if they tried. With science communication the same happens: the trained communicator, the science journalist, the active researcher, the YouTube enthusiast and the automatic generator share a channel with identical presentation, and the viewer has no way to tell them apart.
The asymmetry that worries me
What comes now is my opinion, and I mark it as such even though it leans on what the literature on disinformation has been observing since 2016. A society without an operating hierarchy of sources hasn't just become more confused. It has become unevenly vulnerable, because the cost of attacking it and the cost of defending it have moved in opposite directions.
The actor with resources —the authoritarian state with its troll factory, the platform interested in positioning, the influence agency with a paying client— can today generate authorized-looking content in massive quantity, sow it across many channels and synchronize it among voices that appear independent. Producing that appearance has become dirt cheap with generative models. Evaluating it, by contrast, has grown costlier, because the heuristics the reader filtered with no longer filter. The attacker only needs to saturate the space until telling things apart stops being worth it. They don't even need to lie.
And the institutional response arrives late to everything. Industry has partial incentives and promotes its own content credentials; governments legislate with delay, with the European marking clause coming into force almost four years after the phenomenon turned mass; the academy investigates without the capacity to execute; the platforms adopt minimal and disparate measures. Meanwhile, the same Digital News Report 2024 records that only four in ten respondents, 40%, say they trust most of the news most of the time, a figure stable relative to 2023 and four points below the peak reached during the pandemic. What collapses is not exactly trust, which was already low. It's differential trust, the kind that presupposed some sources are worth more than others, and it sinks in favor of a flat distrust that doesn't tell the serious paper from the pamphlet. Whoever suspects everything equally hasn't become a skeptic. They've given up before starting, and whoever suspects everything ends up believing anything, because they no longer have anything to choose with.
Definitions
Source hierarchy. An ordering, implicit or explicit, that assigns different initial credibility to different emitters according to their validation protocol.
Visual heuristic of quality. A mental rule that infers a content's reliability from formal signals: layout, typography, register, spelling correctness.
C2PA. Coalition for Content Provenance and Authenticity, an industry initiative founded in February 2021 to mark cryptographically the origin and editing chain of multimedia content.
Triangulation. A journalistic and forensic technique that collates a claim with several independent sources that have different incentives from one another.
Referencias
Reuters Institute — Digital News Report 2024 (University of Oxford, June 2024). Source of the data on news avoidance (39%, three points more than the previous year) and trust in most of the news (40%, stable relative to 2023 and four points below the pandemic peak). https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2024/dnr-executive-summary
C2PA — Content Provenance and Authenticity Specifications. Initiative founded in February 2021 by Adobe, Arm, BBC, Intel, Microsoft and Truepic. https://c2pa.org/
Regulation (EU) 2024/1689 — European AI Regulation, article 50 on transparency and marking of AI-generated content; obligation fully enforceable from 2 August 2026. https://artificialintelligenceact.eu/article/50/
American Physical Society — September 2002: Schön Scandal Report is Released. On the Bell Labs committee report that confirmed the data falsification of Jan Hendrik Schön. https://www.aps.org/apsnews/2022/08/september-2002-schon-scandal-report
Postman, N. — Amusing Ourselves to Death (Viking, 1985). On the erosion of public discourse.
Frankfurt, H. — On Bullshit (Princeton UP, 2005). On speech indifferent to truth.
Pariser, E. — The Filter Bubble (Penguin, 2011). On the algorithmic filtering of information.
McIntyre, L. — Post-Truth (MIT Press, 2018). General framework of the phenomenon.
Bender, E. et al. — On the Dangers of Stochastic Parrots. FAccT 2021. On the risks of large language models.
Allcott, H. & Gentzkow, M. — Social Media and Fake News in the 2016 Election. Journal of Economic Perspectives 31 (2017). Starting point of the cited literature on disinformation.
Comments0
No comments yet.
Leave a comment