Six Provocations about AI in News

This post (I don’t want to call it an article because it really isn’t structured enough for that) lists six provocations that I put to a panel of industry leaders at the NAMS 2026 conference – with a big thank you to Naja Nielsen (SVT), Stine Thorsgaard Kjær (TV2 Østjylland), Aida Kokanovic (brAIght) for their participation!

The below is a summary of these propositions and the notes I had on them – they should not be read strictly as “this is definitely the case” but as a mixture of my own interpretation of the evidence on this and some conjecture. I did use some Claude to polish these (so using Pangram on this won’t be the “gotcha” you think it is…)

Observation 1: The low hanging fruits of AI in journalism have been picked. The way forward is not obvious.

My argument here: The easy story was “AI will make newsrooms more efficient”; the hard story is what editorial, economic, and organisational advantage remains once everyone can transcribe, translate, summarise, and draft – so once the obvious applications of AI in news have sort of been done (which I think is a point we have passed).

Some evidence

• In a broad UK survey, 56% of journalists used AI professionally at least weekly, and the most common uses were classic low-friction tasks: transcription/captioning, translation, grammar/copy-editing, research, brainstorming, headlines, and drafting. (epub.ub.uni-muenchen.de)

• News leaders also see the most immediate value in back-end automation. In the 2026 Reuters Institute trends survey, 97% of respondents said back-end automation would be important, with high expectations for newsgathering and product/coding too. (reutersinstitute.politics.ox.ac.uk)

• Fundamentally, I have not seen anything to date that convinced me that my argument from a few years ago – the the use of AI in news itself (not in the environment around it) is predominantly about re-tooling, with the ends of news still the same, and mostly the means changing. See below for what I mean by “ends” and read the paper this is based on here.

  

• We’ve heard time and time again here is that the deeper challenge is that AI changes the environment around news, not only the tools inside newsrooms. The “way forward” I would argue is less about adopting AI as such and more about deciding what journalism remains defensible when generic information that fulfills similar functions becomes cheap.

Likely counter-arguments:

“Many newsrooms have barely started, so the low-hanging fruit has not been picked.” Fair. The point is not that every newsroom has implemented the easy use cases well. It is that the easy use cases are now known: transcription, translation, summarisation, copy-editing, data extraction, alerts, and ideation, AI chatbots. The unresolved question is not “can AI help?” but “what advantage remains once these tools and systems are common?”

“There are already impressive examples, so the path is clear.”
The strongest examples to me are narrow, workflow-embedded, and require expertise, data access, editorial judgement, and organisational capacity. They are closer to building newsroom infrastructure rather than radically transforming what a news organisation could be.

Observation 2: The use of AI in news pushes more responsibility on individuals for tasks that used to be professionalised. Capabilities increase but at the risk of thinner expertise.

My argument: AI makes individual journalists more capable, but it can also turn them into one-person production stacks: researcher, translator, picture editor, data analyst, fact-checker, lawyer, and prompt engineer — without giving them the institutional support those roles used to require, and without them necessarily having the expertise to do some of these things.

Some evidence

• This all inspired by my colleague Carl-Benedikt Frey and his recent piece on AI and busywork in the New York Times.

• AI use is often happening at the level of individual journalists. In the recent UK survey by my colleagues, 56% used AI weekly, but only around a third said their organisation provided AI training.

• AI is used increasingly for editorially significant tasks: story research, fact-checking, verification, and data analysis. But not every journalist is trained to work e.g. on a data analysis task.

• People can over-rely on automated recommendations and output, especially when verification is difficult, time pressure is high, or explanations create misplaced confidence (see Romeo & Conti, 2025)

Likely counter-arguments:

“AI democratises expertise and empowers small newsrooms.” Yes, and that is the strongest version of the pro-AI case in news in my view. But alas access to a capability is not the same as expertise. A reporter can now translate, visualise, summarise, code, and generate media and for many of these uses this will be fine but not for all — and without training and institutional guardrails, you may simply be shifting risk downwards.

“Guidelines solve this.” No. Guidelines are necessary, but even the evidence from the earlier versions suggests they are uneven and incomplete. Policies also don’t solve issues of lack of training or lack of expertise

“Journalists have always adopted new tools.” True, but generative AI is not just a faster typewriter or spellchecker. It produces plausible claims and analysis. That means the journalist is not merely using a tool but an epistemic system.

Observation 3: Having a human-in-the-loop at all times is a paradigm that will not hold.

“Human in the loop” sounds so, so reassuring, doesn’t it? If just the humans stay in control at all times things will be fine, right. Well, no. For one, it often risks becoming a slogan and second, HITL at all times defeats the benefits of one key AI use: scaling. HITL can also be mind-numbing for the people having to do it and ineffective. The real question going forward is where humans need to be in the broader system: before publication, after publication, in audits, in escalation, in design, or in accountability.

Some evidence

• Personalised summaries, AI chat interfaces, synthetic audio, and automated briefings make the “every output checked by a human” model increasingly unrealistic. The more personalised and dynamic the product (or the internal tool), the harder it is to apply traditional pre-publication review to every instance.

• The public is already sceptical that human checking happens consistently. Only 33% of respondents thought journalists always or often check AI outputs before publication in our survey of various countries, and belief in oversight was strongly linked to prior trust in news, so it does not even help that much in making the public trust AI use in news

  

  

• Evidence from decision-making research warns that human oversight can be performative. Human-in-the-loop systems can increase trust in automation without improving accuracy, and sometimes reduce it. (see this experiment in PLOS)

Likely counter-arguments

“We will always have a human review anything we publish.” The reply here is to ask what counts as “publish”. A static article can be reviewed. But what about a personalised summary, a chatbot answer, an automatically generated audio briefing, a translation, or an AI search result based partly on your journalism?

“Without human-in-the-loop, trust will collapse.” Maybe, maybe not. To be honest I don’t think most people care as much about this in reality. And empty claims about oversight may be worse. A more credible approach might be to say: humans must remain accountable at the end of the day, but not every low-risk output can or should be individually pre-approved. High-risk areas get stronger review, while lower-risk areas need get more general monitoring, audits, logs, red-teaming, and so forth.

Observation 4: AI is likely to emerge as a (the?) central way of accessing information and news. There is no guarantee that there is a human premium or one for “trustworthy news” beyond a minority.

Basic point: AI will become the front door to information for most people. In that world, journalism may still be valuable, but not necessarily visible, credited, or paid for beyond a small minority.

Some evidence

• Information-seeking is becoming a core use case. Weekly use of generative AI for “getting information” rose from 11% to 24% between 2024 and 2025

• Explicit news use remains smaller, but it is growing. DNR data from 2025 found that 7% overall use AI for news weekls (and 15% of under-25s), with these numbers going up. The wider trend is already away from direct relationships with news brands.

• There is evidence of that people prefer “human” journalism — but it is conditional. trusted news brands and public-service brands remain important when people want to check whether something is true, but people may need them less often if AI systems provide convenient answers first. And convenience could easily trump loyalty for most people.
  

Audience experiments do not give a simple “humans always win” story. A meta-analysis found no clear credibility advantage for human-written over automated news, though human-written news had advantages on perceived quality and especially readability; other studies find mixed or context-dependent effects. (ideas.repec.org)

Likely counter-arguments

“People still trust news more than ChatGPT.” Yes, and that is important but also trust is not the same as behaviour. People often satisfice: they accept something that is quick, plausible, and good enough. AI search answers may win in most routine moments even when users say, abstractly, that journalism is more trustworthy.

“Only a small minority use AI for news.” That is true for explicit chatbot news use today. But the larger shift is AI-mediated information access through search, mobile assistants, browsers, summaries, and social platforms. I don’t see this abating anytime soon, just look at e.g. Google’s latest IO.

“In a synthetic information environment, trusted news brands will become more valuable.” They may — especially for high-stakes verification and for a politically interested minority. The uncomfortable question is whether that premium is broad enough to sustain mass-market news, or whether it concentrates value among a few elite brands. We will see, to me this is an open question.

Observation 5: AI in journalism, too, will be about replacing people, not just augmenting their work.

I don’t think it’s defensible to either argue “AI will replace journalists” or “AI will only augment them”. I think that AI will replace some tasks, some roles, some entry-level pathways, and some news organisations’ willingness to hire.

Some evidence

• News leaders already report some labour effects. In the 2026 Reuters Institute trends survey, 16% said AI efficiencies had slightly reduced staff numbers, while 67% said no jobs had yet been saved and 9% said AI had added roles or costs. That is not mass replacement but it is also not pure augmentation either, sorry.

• Public expectations are also strongly substitutional. Pew found that 59% of Americans expected AI to lead to fewer journalism jobs over the next two decades, while only 5% expected more journalism jobs.

• Labour-market evidence for now seems to suggests transformation rather than simple disappearance, but transformation can still mean fewer jobs in particular functions. The International Labour Organisation’s 2025 update says most jobs are more likely to be transformed than made redundant, while also noting that media and web occupations have become more exposed as AI has improved in voice, image, and video. (International Labour Organization)

• Generally, as Stanford’s Bharat Chandar notes the evidence on all of this is not great. But he and colleagues argue that “since the widespread adoption of generative AI, early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 16 percent relative decline in employment even after controlling for firm-level shocks. In contrast, employment for workers in less exposed fields and more experienced workers in the same occupations has remained stable or continued to grow. We also find that adjustments occur primarily through employment rather than compensation. Furthermore, employment declines are concentrated in occupations where AI is more likely to automate, rather than augment, human labor. Our results are robust to alternative explanations, such as excluding technology-related firms and excluding occupations amenable to remote work. These six facts provide early, large-scale evidence consistent with the hypothesis that the AI revolution is beginning to have a significant and disproportionate impact on entry-level workers in the American labor market.” The question this raises for me: why should it be different in the news industry? Unless you have a good argument why the news should somehow be exempt from these dynamics, to keep on claiming that AI will not lead to job losses in news seems in the best case scenario delusional thinking and in the worst case scenario a thinly veiled untruth.

Likely counter-arguments

“AI will free journalists to do deeper reporting.” Sometimes, yes, but that’s a (self-motivated) claim, not a fact. It is the best-case organisational choice: automate drudgery and reinvest the saved time in reporting. But market incentives often convert efficiency into more output, lower cost, or reduced hiring. Augmentation is a choice, not an automatic outcome.

“AI transforms tasks, not jobs.” That distinction is useful but can obscure the employment consequences. Jobs are bundles of tasks and ff enough tasks in copy-editing, translation, monitoring, social packaging, summaries, SEO, service journalism, and basic rewriting (the list continues) are automated, staffing models change even if “the journalist” is not fully replaced.

“AI cannot replace the judgement and relationships of good reporters.” Agreed (or at least, not yet). But many newsroom roles are not pure investigative reporting. AI does not need to replace the best reporter to reshape hiring, promotion, entry-level training, and the broader economics of routine news production.

Observation 6: People judge AI in news by different standards than AI in other sectors – and no amount of “responsible AI” will change that anytime soon.

Finally, I do believe that AI in news is not judged like AI in entertainment or in search and that this is part of why “responsible AI” approaches – well-intentioned and sensible they are – will not cut it, as people’s attitudes towards AI use in news are about more than the technology and it use itself. And when I say more it’s people’s interest and trust in news.

Some evidence

• I and colleagues found that people were more pessimistic about AI in news media than in many other sectors. News media, government, and political parties were among the few areas where pessimists outnumbered optimists.

• Rising general AI use has not yet normalised AI news. Instead, attitudes have hardened with only a minority of about 12% comfortable with news made entirely by AI, 21% with news made mostly by AI with some human oversight, 43% with mostly human news using some AI, and 62% with entirely human-made news.

• Acceptance is highly task-specific. People are more comfortable with AI for back-end task than with AI-created realistic images, artificial presenters, or artificial authors. This suggests the concern is not simply “AI yes/no”, but whether AI touches authorship, representation, deception, or editorial judgement.

• People expect AI to make news cheaper and more up to date, but also less transparent and less trustworthy. That is a difficult trade-off for news organisations: people might see why news organisations would want to use this technology, but still not necessarily admire it.

• Public scepticism is amplified by doubts about newsroom oversight. Only a third of respondents believed journalists always or often check AI outputs, and those who already trusted news were much more likely to believe oversight happens.

Likely counter-arguments

“Transparency and responsible AI policies will solve this.” They are necessary, but not sufficient. Disclosure can sometimes reduce trust rather than increase it, and audiences may interpret AI use as a signal of cost-cutting, lower care, or weaker accountability.

“People will get used to AI once it becomes useful.”
I generally believe there will be a domestication effect unfolding over time. The evidence already shows greater comfort with spelling, translation, summaries, and other utility functions. But news is still (and will likely remain) a high-normativity domain: people judge it through concerns about the society they live in (or want to live in), politics, and things like institutional trust – and I don’t expect this to magically disappear.