‘AI Has Just Changed Everything’ – for the Seventeenth Time: Hype as a Journalistic Genre

The curtain rises. A person in trainers walks on stage before a screen the size of a small country. The new system answers a question, translates a document and generates a video. Everything works on the fir st attempt. The audience applauds. Reporters take notes. A press release is already waiting in their inboxes: ‘a new era’, ‘a breakthrough’, ‘a tool that will transform work’. Before the day is over, headlines ask whether an entire profession has just ended.

The following morning, an accounts clerk tries the same trick on a real spreadsheet. The system loses a table. A teacher discovers that the licence does not cover pupil data. A newsroom learns that ‘automation’ requires hours of correction. Such scenes happen after the music stops, without a carefully selected input.

Hype is not simply a matter of overexcited adjectives. It organises time. The present becomes a threshold, the future becomes inevitable, and reflection begins to look like lateness. Hype has acquired the recurring parts of a journalistic genre: hero, conflict, turning point and a missing chapter about maintenance.

Hype has a format, not merely a temperature

A familiar AI story begins by turning a product into a subject: it understands, creates, discovers, learns. Then comes a demonstration selected by the maker, followed by a quotation from somebody who benefits if the technology succeeds. An analogy to electricity, the printing press or the Industrial Revolution supplies historical weight. The conclusion informs organisations and workers that they must adapt immediately.

The simplest alternative is often missing. How does a person, a spreadsheet, a search engine or an older system perform the same task? Without that baseline, ‘90% accurate’ is a number wearing a smart suit but carrying no address. We do not know the test data, refusal rate, correction cost or whether errors fall evenly across groups.

UNESCO’s handbook for journalism educators identifies recurring pitfalls: hyperbole, unjustified claims about future progress, false comparisons between narrow systems and broad human abilities, recycled PR language, buried limitations, invisible labour and performance figures without adequate conditions or uncertainty (Jaakkola, 2023, pp. 110–113). These are not cosmetic faults. Each one changes the model of reality that a reader takes away.

Framing: the spotlight chooses the revolution

Robert Entman describes framing as selecting aspects of perceived reality and making them salient in ways that promote a problem definition, causal interpretation, moral evaluation or proposed remedy (Entman, 1993, pp. 51– 58). Framing need not involve a lie. A spotlight may illuminate a genuine object. It still leaves the rest of the stage in shadow.

When launch coverage focuses on generation speed, efficiency defines the problem. The central question becomes ‘how much time will it save?’. Other definitions fall outside the beam: who verifies, who answers for error, whose work supplied the data, whom the system serves poorly, and what happens to the person affected by its decision. A ‘race’ frame makes scrutiny look

obstructive. If competitors are already running, testing and regulation resemble somebody tying their shoelaces after the gun.

Verbs distribute responsibility too. ‘AI decided’ removes the organisation that selected the model, threshold and data. ‘The model learnt’ can obscure the people who labelled examples, moderated disturbing content and corrected results. ‘The system replaced editors’ erases the managers who chose to reduce staff. Grammar settles accountability before an ethics committee has entered the room.

Technology does not descend from the cloud

The social construction of technology offers an alternative to stories in which an invention has one obvious nature and pushes history by itself. Trevor Pinch and Wiebe Bijker showed how technologies acquire meaning and form through negotiations among relevant social groups that define problems and successful solutions differently (Pinch & Bijker, 1984, pp. 399–441).

The ‘same’ AI tool is therefore different things to a vendor, investor, editor, reporter, trade union, source and audience member. A vendor sees a scalable service. A reporter sees help with transcription. A manager sees a cost reduction. An informant sees a new participant in a confidential process. A reader sees an invisible co-author. There is no socially neutral innovation waiting before use; there are design choices, contracts, metrics and institutions.

Journalists should ask not only ‘what can the system do?’ but ‘for whom is this a problem, and who gets to declare it solved?’. A tool may summarise routine documents well while failing people whose language or circumstances are poorly represented. It may save an author five minutes and add twenty minutes to a checker’s work. ‘Efficiency’ without a named beneficiary is a bill without a currency.

Expectations work before the product does

The sociology of expectations shows that visions of the future do more than describe what may happen. They coordinate present action, attract

investment, recruit people and provide legitimacy. Expectations can align networks and justify choices before a technology reaches the promised maturity (Borup et al., 2006, pp. 285–298).

That is the performative power of hype. ‘AI will revolutionise education’ can prompt a university to buy licences, create a post and change a programme. Once those investments exist, the technology becomes more present not simply because the forecast was correct, but because the forecast helped construct the conditions of its own fulfilment. News coverage participates by supplying visibility and the language through which a board justifies its decision.

Expectations are not inherently deceptive. Research and innovation require imagined futures. The problem begins when a vision has no condition of failure and the cost of disappointment falls on somebody other than the promiser. If the system does not work today, we hear that ‘this is only the beginning’. If it works on one task, scale is described as inevitable. If it harms, the organisation is said to be ‘not ready’. Hype behaves like a forecast with diplomatic immunity.

Steve Woolgar proposes analytical scepticism: reflective distance from seductive claims about novelty and impact. His instruction is wonderfully short: ‘We need that question mark’ (as cited in Jaakkola, 2023, p. 78). The question mark is not an innovation brake. It is a seat belt for a sentence travelling too fast.

Who gets to speak in AI news

Hype grows easily when access to the stage is unequal. UNESCO’s handbook cites research into UK coverage in which almost 60% of articles concerned industry products or initiatives. Roughly one third of unique sources were affiliated with industry — almost twice the number from academia and six times the number from government. Nearly 12% of all articles referred to Elon Musk (Jaakkola, 2023, p. 84).

This is not an argument for excluding business leaders. They know a great deal about their products. They are also interested sources, like a minister promoting a reform or a football club presenting a new signing. Expertise does

not remove interest. A rigorous article needs test conditions and voices beyond the launch: users, workers, independent domain experts, regulators, affected communities and people who encounter failures after the photographers leave.

Jan Kreft presents AI in journalism as passing through phases of niche interest, euphoria, negotiation and normalisation. He also invokes Roy Amara’s observation that people tend to overestimate technology’s short-term effects and underestimate its longer-term consequences (Kreft, 2025, pp. 7– 10). Launch-day coverage is poorly equipped for that timescale. The largest change may arrive after years of quiet integration into employment, education, language and procedure.

Hype has casualties, not merely disappointed customers

Imagine a university buying a system advertised as ‘automated personalised support’. Funding for human tutors is reduced because the demonstration promises scale. After one term, the tool handles standard questions but misses unusual cases and fails to recognise when a student needs a person. Confident prompt writers manage. A student working in a second language receives a series of wrong directions and misses a deadline.

This example is hypothetical. It shows that framing errors have a distribution. Senior management can list the deployment in a strategy, the provider can place it in a presentation, and the media can leave the launch story in an archive. The student pays with a term. The same structure appears when a newsroom cuts copy-editing, a public body closes a simpler human route, or a hospital buys on the strength of a result achieved in different conditions.

Human dignity requires that people are not portrayed as drag on an adoption curve. Freedom requires a real alternative, particularly where a system affects rights, work or essential services. Accountability requires a named decision- maker and a route of appeal. Hype removes these elements quietly by recasting a political or organisational choice as ‘inevitable progress’.

The anti-hype audit: turn the promise into a test

An audit is not designed to prove that technology is bad. It converts a promise into a proposition that can be tested honestly. Create a dossier for every major claim in the story.

1. Measurable claim. Replace ‘revolutionises’, ‘understands’ and ‘works like an expert’ with a task, population, conditions and outcome. ‘The system reduces mean transcription time for files of type X from A to B at error rate C’ is less glamorous and more useful.

2. Baseline. What is the comparison: a person with stated experience, an older system, a simple procedure or no intervention? The benchmark must resemble real use. A pristine launch file is not a newsroom at 5.55 pm.

3. Evidence and conditions. Who ran the evaluation, on what sample, with which exclusions and measures? Is the result from a laboratory, pilot or ordinary operation? How often does the tool refuse? What is the worst plausible error? Has anybody independent replicated the result?

4. Limitations near the top. Do not bury the decisive caveat in the penultimate paragraph after a parade of enthusiastic quotations. If the system works only in a narrow setting, that is part of the product definition, not a ‘sceptic’s view’.

5. Source interest. Record financial, professional and reputational ties. Ask who was absent from the demonstration. An industry source can supply facts; promotional claims still need an independent test.

6. Horizon and failure condition. When should the prediction come true? What observation would show that it did not? A claim confirmed by every possible future is strategic narrative, not a journalistic finding. If the deadline moves, record the movement.

7. Invisible labour. Count the people preparing data, correcting output, handling exceptions, moderating content, training users and answering complaints. ‘Automated’ often means that labour moved to another room, supplier or country.

8. External cost and distribution. Who pays in energy, data, copyright, stress, lost work, exclusion or wrong decisions? Who gains, and who carries risk? An average improvement may conceal concentrated harm.

9. Exit, appeal and correction. Can a user choose a person, challenge a result, recover data and return to the previous route? Who repairs an error after deployment? A technology without an exit plan is less an innovation than furniture placed against the door.

Turning the dossier into readable journalism

The audit need not make an article sound like procurement paperwork. Begin with a person doing a concrete task, then show the promise, test and consequence. Put the vendor’s demonstration beside a simple real-world trial. Instead of asking ‘will AI replace journalists?’, ask which tasks change, who acquires control, and what happens to the time supposedly saved at the first stage.

A strong article keeps three clocks. The first measures present capability. The second measures promised development. The third measures the time institutions and people need to adapt. Hype is manufactured by mixing the clocks: a prototype result in the present is written in the language of full-scale deployment, while a future forecast adopts the grammar of fact.

Headlines can retain drama without settling the evidence in advance. A question mark is not clickbait when the article genuinely maps what is unknown. Sometimes the strongest conclusion will be that a tool performs one narrow task well but there is no evidence yet that it improves the whole process. That is less spectacular than ‘the end of a profession’ and more likely to prevent a bad decision.

The limits of scepticism

Anti-hype can become a genre of its own. ‘AI is only statistics’ may be as reductive as ‘AI understands the world’. Excessive scepticism can miss slow structural change or dismiss a useful tool because it does not resemble science fiction. Criticising industry sources does not excuse technical ignorance, and identifying a risk does not prove harm in every use.

Evidence will often be incomplete while systems change faster than research cycles. Journalists must then report uncertainty, separating observation from forecast and naming the forecaster. The audit does not manufacture certainty; it produces a more honest map of ignorance.

Nor can performance testing settle values. A system may be cheaper and more accurate yet still be inappropriate for a decision without meaningful human control. Choosing what to optimise — and what must never be traded away — is an ethical and political act, not merely a technical one.

The revolution can wait until paragraph two

Technology deserves curiosity. That is precisely why its vocabulary should not be outsourced to press releases. The best AI stories do not ask whether the future is coming. They show who is building it, under which conditions, with whose money and with whose right to object.

When the screen the size of a small country announces the next breakthrough, a reporter may applaud — and then open the dossier. Enter the baseline, source interest, horizon, cost and failure condition. Leave a question mark in the headline until the evidence earns a full stop. That does not kill the story. It stops the story selling the future on credit taken out in somebody else’s name.