Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Newsroom AI Becomes the Expertise Catch-22

Mari Reisjå and Anders Sundnes Løvlie followed generative-AI plans and practice in four Norwegian news organizations through the 2025 parliamentary election.

An editorial expertise receipt records which human knowledge checks AI-assisted work, whether that knowledge is being preserved, and whether one model is quietly reviewing another model's influence.

The Paper

The paper is Mari Reisjå and Anders Sundnes Løvlie's The GenAI Catch-22: Use of Generative Artificial Intelligence in Norwegian Newsrooms During the 2025 Parliamentary Election, arXiv:2608.10773v1 [cs.CY], cross-listed in cs.AI and submitted August 11, 2026. Its 28-page PDF lists both authors at the IT University of Copenhagen.

The authors conducted 13 semi-structured interviews with nine informants from NRK, TV2, VG, and Dagbladet between January and November 2025. Four managers responsible for AI adoption were interviewed before and after the election; five journalists working on party-selection quizzes were interviewed about practice. This is a longitudinal qualitative case study of four organizations, not a prevalence survey of journalism.

The Public Ambitions Receded

Early plans were conspicuous and audience-facing. Managers discussed election chatbots, personalized services, and live fact-checking. By the campaign, all four organizations had shelved plans for GenAI-powered party-selection quizzes. The paper reports accuracy problems, weak user interest in open-ended chatbot interaction, difficulty identifying which debate claims were worth checking, and review burdens that erased the imagined scale advantage.

That retreat matters because professional judgment did stop some deployments. It also risks becoming the reassuring headline. A canceled chatbot is easy to see; a model used to navigate party programs, condense text, suggest questions, or simplify language is ordinary enough to disappear into the workflow.

The Internal Use Remained

The four newsrooms returned to familiar quiz formats while journalists used internal and commercial GenAI tools behind them. Reported tasks included navigating party programs, translating or condensing the journalists' own text, generating summaries and ideas, identifying possible blind spots, weighting parties on a predefined scale, and adapting language for younger readers.

The paper does not treat all assistance as failure. Some participants reported saving time. One experienced NRK quiz maker, however, largely abandoned the tool because correcting its suggestions took longer than doing the work directly. The relevant unit is therefore not output speed. It is the full editorial loop: assignment, source reading, generation, checking, correction, senior review, publication, and later repair.

The Expertise Catch-22

Managers repeatedly located safety in journalists' source criticism and professional competence. Yet the study describes work being assigned on the assumption that AI could compensate for limited prior experience. A TV2 student intern reported four or five quiz errors connected to incomplete checking or definitions. In another instance, an economics journalist caught a politically charged phrase in an AI-assisted question that the less experienced quiz maker had not recognized.

This is the paper's catch-22. Human expertise is needed to notice plausible errors, loaded language, omitted context, and misplaced emphasis. If AI use reduces the time, staffing, apprenticeship, or repeated practice through which that expertise develops, the safeguard can weaken while formal human approval remains. A person can remain formally in the loop after the institution has thinned out the knowledge required to make the loop meaningful.

Review Needs Provenance

The study found limited internal visibility into colleagues' AI use. One participant could not tell whether colleagues checking AI-assisted work had themselves used GenAI. That creates a review-provenance problem: two apparent checks may share a model, prompt pattern, source omission, or framing bias. Agreement is weaker evidence when its dependencies are hidden.

Public disclosure also receded during the study period. The paper reports that labeling policies became less strict or less specific, and none of the four party-selection quizzes disclosed the reported GenAI assistance. Whether every minor tool use needs a public label is a separate policy question. Internally, however, consequential election work needs enough provenance to reconstruct what the tool did and which review was independent of it.

The Governance Reading

An editorial expertise receipt should name the assignment, source corpus, AI tool, task delegated, generated artifact, journalist's relevant domain experience, factual checks, framing checks, senior reviewer, reviewer tool use, corrected errors, unresolved disagreements, disclosure decision, publication owner, and correction path. It should distinguish source-grounded retrieval from generated interpretation and independent review from same-system comparison.

The labor controls follow from that record: reserve paid review time, preserve experienced staff, keep apprentices doing enough first-hand source work to acquire judgment, test whether claimed time savings survive correction costs, and never use an AI availability claim as evidence that a less experienced worker can safely replace domain expertise. Human oversight is a maintained organizational capacity, not a checkbox attached to publication.

What the Study Does Not Establish

The sample is small, purposive, national, and based largely on interviews. The first author disclosed being on study leave from her position at NRK and having previously worked at VG, while the second author provided an external perspective. The study did not perform a content analysis that could identify causal effects on coverage, did not test audience trust, and explicitly says it cannot determine whether GenAI exerted undue or detrimental influence on the election coverage.

Its contribution is narrower and useful: it documents a gap between plans and practice, reported errors and review dependencies, reduced labeling, and a plausible long-term risk to the expertise on which oversight depends. Those findings justify better records and further observation, not a verdict on every newsroom or every use of AI.

Source Discipline

The factual source is the arXiv abstract and complete version 1 PDF. The paper was checked for bibliographic metadata, affiliations, interview design, reported newsroom practices, examples, disclosure changes, positionality, author-declared AI use, and limits. This essay paraphrases rather than reproduces interview quotations and separates the authors' observations from its proposed governance controls.

Sources

Return to Blog