In this article
Today we get into a specific mechanism that circulates in disguise, and it's worth stripping it bare. The thesis: the press release from OpenAI — and from any AI lab with a serious communications department — is not information. It's an announcement written by people paid to sell you something, approved by lawyers, distributed by a PR agency and signed by the very company that is the subject of the announcement. When that release reaches the Spanish reader as news, there's a sleight of hand worth watching. Why does the distinction matter? Because it makes you believe you're getting independent information when you're getting advertising. My opinion is biased, because I've worked with PR departments and I know how the sausage is made. Form your own.
For two years I've been clipping OpenAI releases and comparing them with the coverage they get in the Spanish and international press. The conclusion is that the vast majority of the pieces the reader consumes as news about OpenAI are, technically, the more or less expanded translation of the original release. That isn't journalism in the classic sense. But it's signed as if it were, and read as if it were. Let's look at the mechanism slowly.
The technical difference between a release and a news story
A press release is written by the company itself, generally through its corporate communications department. The legal team approves it, it's coordinated with the product launch, and a PR agency distributes it to a list of journalists selected at the company's discretion. Its declared aim is for it to be published with the frame the company wants. Its undeclared aim is for the frame to carry over unquestioned.
A release has a predictable structure that repeats month to month. A pre-prepared headline. A quote from Sam Altman, Dario Amodei or the CEO of the moment. A benchmark figure, generally accompanied by a favourable comparison with a previous model. A statement about the sector's progress and the company's commitment to safety. A final paragraph inviting you to try the product. There's no surprise. There's no tension. There's nothing the company hasn't decided to publish.
A journalistic piece should do exactly the opposite. Start from the release's information, yes, but cross-check it with outside sources. Ask independent researchers for reaction. Verify the figures with third-party benchmarks. Mention what the release doesn't include. Recall what the company announced six months ago and didn't deliver. Identify what's being announced that competes with what within the company's own catalogue.
The distance between the two things is enormous. And yet the typical Spanish press piece on an OpenAI launch stays much closer to the release than to the ideal piece. Not out of bad faith, but because of the operational constraints I described in another instalment of this batch.
The concrete cycle, step by step
It's worth reconstructing the cycle from end to end, because the abstraction doesn't help and the steps are always the same.
At 16:00 Pacific time — midnight in Madrid — OpenAI publishes the announcement blogpost at openai.com/blog. The blogpost has a header with the model or product, a short video, a series of benchmark comparisons against the previous model, a couple of polished demonstrations, quotes from Altman and from some researcher, and links to the technical documentation.
Simultaneously, the international agencies — Reuters, Associated Press, Bloomberg, Agence France-Presse, Efe — post a wire of about 400 to 600 words reproducing the release's central points. That wire is prepared in large part from material pre-distributed by OpenAI to the agencies under embargo, which means the wire exists before the moment of publication but is only released at the exact agreed hour. The wires are distributed to subscribing newsrooms worldwide.
In Spain, between midnight and 8 in the morning, the newsrooms with a night shift publish a translated version of the agency wire. Those without a shift publish between 8 and 10. By midday, the five big dailies and the main portals have all published a very similar piece. The structure is almost identical. The headline is a variation on the release's central phrase. The figures are those of the benchmark OpenAI has published. The quotes are those of the release. The piece runs between 350 and 700 words.
What appears in none of the 30 to 50 Spanish pieces that come out that day is practically nothing. There's no reaction from independent researchers. There's no verification with third-party benchmarks. There's no mention of what OpenAI said six months ago about the same product and whether it delivered. There's no reminder of the pending litigation with publishers over unlicensed content use. There's no contrast with the competing models from Anthropic or Google that will probably publish something similar the following week. There is, instead, what the release wanted there to be: the chosen headline, the chosen figure, the chosen quote, the chosen frame.
The embargo and exclusive trick
There's a finer mechanism worth naming, because it explains why some pieces seem to be an exception to the previous pattern and, looked at closely, aren't.
OpenAI — like Apple, Google, Microsoft or any other company with professional PR machinery — grants early access to a limited number of selected journalists. It generally selects three to five international bylines: The New York Times, The Verge, Wired, Bloomberg, occasionally the Financial Times or The Information. The selected journalists receive access to the product a few days before the public announcement, under embargo: they can only publish at the agreed hour, and they can't leak to anyone in the interval.
In exchange for early access, the journalist produces a longer, more informed and apparently more critical piece than the release. But the access is renewable. The company decides who it will give early access to for the next launch. That introduces a silent incentive. The journalist who criticises too hard loses access. The journalist who keeps to the agreed line retains it. The criticism permitted is the nuanced criticism, the elegant doubt, the "more could have been expected" that doesn't discomfort the PR department. The deep criticism — "this model is worse calibrated than advertised," "the X problem the company has spent years promising to solve is still open," "the benchmark figures are measured under a non-comparable inference regime" — practically never appears in pieces with early access.
This isn't conspiracy theory. It's standard editorial economics. The well-managed embargo is an instrument for governing coverage, perfectly legal and perfectly effective. Its technical name, in the literature of source journalism, is soft editorial capture. It requires bribing no one. It requires only that the access be worth more than the criticism.
The DeepSeek case as a contrast
There's a recent case that illuminates the mechanism by contrast, and it's worth keeping at hand. DeepSeek published its R1 model on 20 January 2025 with no Western PR machinery, no embargo to selected journalists, no CEO statement tour. They uploaded the paper to arXiv, published the weights on HuggingFace and that was that.
For the first 48 hours, Western media coverage of DeepSeek was almost non-existent. The journalists had no agency wire, no executive quote, no choreographed demo. What they had was the technical paper, the public weights, and what was being said in specialist circles on Twitter and in Chinese forums.
When, on Monday 27 January, Nvidia dropped 589 billion dollars in a single session because of the DeepSeek effect, the coverage exploded. But the journalistic piece produced in those days was notably different from the typical OpenAI piece. There was more cross-checking with technical sources, more verification, more nuance, more original reporting. The reason is operational: since there was no official release to stick to, the journalists had to work.
That difference between DeepSeek 2025 coverage and typical OpenAI coverage is the difference between journalism and the relay of a release. Same profession, same masthead, same journalist; but, without the PR crutch, the result is structurally different.
The operational test
There's a simple routine to tell, faced with a specific piece, whether it's journalism or an expanded release. Four questions the reader can ask in thirty seconds.
First question: does the piece mention what the company hasn't solved in this launch? If a new model is announced and the piece doesn't mention the problems the model doesn't solve — long context, hallucinations, energy consumption, bias in non-English languages, known failures on specific tasks — the piece is incomplete. Every technical product has limitations and an honest journalist names them.
Second question: does the piece compare with direct competitors? If OpenAI launches a model and the piece doesn't mention what it does better or worse than Claude, Gemini, Llama or DeepSeek, the piece is incomplete. The sector is understood only in comparison.
Third question: does the piece recall earlier promises? If the company announced six months ago that it would solve X and now announces it has launched Y without mentioning X, the piece is incomplete. Temporal traceability is part of the analysis.
Fourth question: does the piece quote anyone who isn't an employee or former employee of the company? Academic researchers, independent analysts, sector critics. If the only voices in the piece are those of the release, it isn't journalism. It's transcription.
If the piece fails all four questions, it isn't journalism. It's an expanded release with the journalist's byline on top. If it passes one or two, it's weak journalism. If it passes three or four, it's decent journalism. The vast majority of Spanish pieces on OpenAI launches fail three or four.
What the reader can do
There are three habits that change the consumption of this stream.
The first is to always read the original release before the coverage. Five minutes at openai.com/blog save forty reading twenty pieces that say the same thing. And, having read the original, the reader has the judgement to see what each piece adds.
The second is to seek out at least one piece that passes the four-question test. Stratechery usually passes it. Platformer usually passes it. Wired sometimes passes it. MIT Technology Review sometimes passes it. Spanish general-interest pieces almost never pass it. That active search shifts consumption from passive to deliberate.
The third is to wait seventy-two hours. Almost everything that matters about a launch is understood better three days later, when the first technical criticisms, the first external benchmarks, and the first competitor reaction have come out. The urgency of the first-day headline almost never makes up for the loss of context.
The political question
This is personal opinion, but I hold it with three years of archives. The degradation of tech journalism through dependence on the press release isn't a problem of a specific country. It's a business-model problem. Outlets that depend structurally on sector advertising can't afford to irritate the advertisers. Outlets that depend on early access can't afford to irritate the source. The outlets that can afford it — because their business is the reader's direct subscription, not advertising — are the ones offering honest coverage.
That correlation is informative. It suggests the structural solution to the problem doesn't run through the journalists' individual goodwill, but through changing what the outlets depend on economically. The reader's direct subscription is, today, the only structure proven capable of sustaining journalism critical of the tech sector. Any strategy that runs through continuing to read for free and hoping the coverage improves is an illusion.
What I'd ask — with no realistic hope, but with clarity — is that the readers with the capacity to pay subscribe to the outlets that apply the four-question test. That the non-tech companies needing to understand the sector pay subscriptions for their teams to Stratechery, MIT Technology Review or 404 Media as part of the training package. And that the general-interest press, in the meantime, at least link the original release alongside its coverage, so the reader can compare and judge.
A concrete figure closes it. In December 2024, the research team at Maldita.es — whose coverage of corporate releases has been documented for years at maldita.es — published an analysis of the coverage cycle of OpenAI's o1 launch in September 2024. Within 24 hours of the original release, the six big Spanish digital dailies had published a total of 17 pieces. Of those 17, 16 reproduced the release's benchmark figures without external verification, 15 quoted no independent researcher, 14 didn't compare with Claude or Gemini, and none recalled what OpenAI had promised six months earlier and not delivered. The figure is from Maldita Tecnología and, in itself, measures the exact dimension of the problem.
Definitions
Press release: a document written by a company or institution and distributed to media with the intention that it be published. Predictable structure, controlled frame, selected quotes. It's the starting material of the news, not the news itself.
Embargo: a condition agreed between source and journalist whereby the journalist receives information in advance in exchange for not publishing it until a specific hour. It's a legitimate coordination technique; its misuse is as a mechanism of selection and control.
Soft editorial capture: a situation in which an outlet or journalist modulates its coverage to keep preferential access to a source. It requires no explicit agreement; it's produced by economic and professional incentives sustained over time.
Agency wire: a brief piece of information produced by a news agency (Reuters, AP, AFP, Efe) and distributed to its subscriber outlets. It serves as a base for each outlet to produce its own piece, though in practice many outlets only translate it.
References
Maldita.es / Maldita Tecnología, analysis of corporate-release coverage on AI (2023–2025). Cited figures on the coverage of the o1 launch.
Reuters Institute for the Study of Journalism, Digital News Report 2025 (Oxford University, June 2025). Data on coverage structure and advertising dependence.
Edward S. Herman & Noam Chomsky, Manufacturing Consent: The Political Economy of the Mass Media (Pantheon, 1988). The founding text on the filters that act on media coverage.
Cory Doctorow, The Internet Con (Verso, 2023). A frame for understanding media concentration and the incentive structure of current tech journalism.
Neil Postman, Amusing Ourselves to Death (Viking, 1985). A precedent on how the format conditions the substance of the information transmitted.
Stratechery, recurring analysis of the release cycle of the big AI labs. Accessible examples at stratechery.com.
Going deeper
Ryan Holiday, Trust Me, I'm Lying: Confessions of a Media Manipulator (Portfolio, 2012, revised edition 2017). An insider manual on how a modern PR campaign is designed.
Jay Rosen, PressThink (pressthink.org). An academic blog on journalistic practice, useful for understanding the debate on embargo, access and capture.
Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025). In-depth reporting on OpenAI, an example of a piece that would pass the test's four questions.
The Markup (themarkup.org). A US non-profit outlet specialised in algorithmic investigations; an alternative funding model for critical coverage of the sector.
You might also like
- Why Spanish journalism about AI is very bad
- Stratechery, Platformer and 404 Media, AI journalism done seriously
- Ten reasons never to read another listicle about AI in your life
- Who OpenAI really is
Comments0
No comments yet.
Leave a comment