A headline that says the opposite of its body text

France-Soir’s headline reads: « OpenAI inquiète Wall Street avec 20 milliards de dollars de revenus en moins qu’annoncé » (OpenAI worries Wall Street with 20 billion dollars less in revenue than announced) [1]. A few paragraphs later, the same article concedes: « OpenAI n’a pas réellement perdu 20 milliards de dollars » (OpenAI did not actually lose 20 billion dollars) [1]. The reader in a hurry remembers the headline. This is a classic newspeak mechanism: the word « moins » (less) suggests a real shortfall, while the text describes a difference between two estimates.

The Guardian does the same: “OpenAI projected to bring in $20bn less in revenue than expected” [2]. The paper also writes that OpenAI “is making about $20bn less in projected revenue” [2]. The phrasing mixes the present (“is making”) with projection, which gives the figure the weight of an accomplished fact.

A gap that is first of all accounting

Yet the methodological detail is in all three articles. According to The Guardian, the gap “arose from attempts by OpenAI investors to provide a more direct comparison” [2] with Anthropic, which includes sales through AWS and Google Cloud, unlike OpenAI [2]. Courrier international reports that « les deux entreprises calculent ce chiffre de façon différente » (the two companies calculate this figure differently) and that the 70 billion figure « était extrapolé selon des paramètres identiques » (was extrapolated using identical parameters) [3]. The same article speaks of a « correction » (correction) [3].

One point remains unclear, and we do not settle it. Courrier international writes that the 70 billion figure was the one « dont la presse faisait état » (that the press reported), « un montant que l’entreprise n’avait pas démenti » (an amount the company had not denied) [3]. The Guardian says it appeared in information provided to investors and was “widely reported” [2]. Who produced this 70 billion, and on what basis? No article says so clearly. OpenAI’s failure to deny it, noted only by Courrier international [3], deserves to be front and center.

The Nasdaq as proof by outcome

All three texts run the same sequence: revelation, then fall. France-Soir reports a Nasdaq down 1.25% [1]; The Guardian and Courrier international say 1.4% [2][3]. This divergence is not explained, and we cannot settle it from these articles alone. For Nvidia, the figures roughly converge (« près de 3 % » (nearly 3%) [1], 2.9% [2][3]).

France-Soir writes that « plusieurs valeurs emblématiques de l’intelligence artificielle ont ainsi été sanctionnées » (several emblematic AI stocks were thus punished) [1]. The verb « sanctionner » (to punish) turns the market into a moral judge. Courrier international says that the « résultat de ce scoop […] ne s’est pas fait attendre » (result of this scoop […] was not long in coming) [3]. The causal link is presented as self-evident. Yet no article shows that this revelation alone explains the decline, and The Guardian points to other contextual factors at the same moment, such as safety concerns around AI [2]. We are describing a possible reading here, not a demonstration.

The vocabulary of the bubble, or anxiety as a product

In Huxley, conditioning works through emotion more than through coercion. Here, the coverage oscillates between « inquiétudes » (worries) [1][3], « vacillé » (wavered) [1] and “buffeted” [2]. Courrier international’s footer links (« Faut-il craindre une explosion de la bulle IA ? » (Should we fear the AI bubble bursting?)) extend this register [3], even if this is editorial packaging and not part of the article body.

At the same time, a figure of 50 billion is still presented as « spectaculaire » (spectacular) growth [1] or « rapide » (rapid) growth [3]. France-Soir notes that OpenAI was at 20 billion annualized at the start of the year, against 6 billion in 2024, according to Reuters [1]. In other words, the “bad news” is a figure that more than doubled. The articles do not examine this paradox: it is expectations that are disappointed, not activity that is collapsing.

What is repeated without being examined

The notion of “annualized revenue” is explained only by France-Soir, which calls it « projection théorique, particulièrement sensible aux hypothèses retenues » (a theoretical projection, particularly sensitive to the assumptions used) [1]. The Guardian describes it as “closely watched by markets” [2], without defining it. Courrier international, citing the Financial Times, makes it « l’indicateur décisif » (the decisive indicator) of AI demand [3]. A non-standardized metric thus becomes the yardstick of an entire industry, precisely because everyone repeats it.

France-Soir also cites an estimate of 10,300 billion dollars in investments by 2032, and calculations from Columbia University suggesting 3,550 billion in annual revenue would be needed [1]. These figures are reported via Reuters, and we cannot verify their method. They are used to conclude that « les marchés valorisent aujourd’hui des entreprises en anticipant des recettes futures considérables » (markets today value companies by anticipating considerable future revenues) [1], which is a description, not a revelation.

What is hardly said

What a reader would need instead