The GenAI Catch-22: Use of Generative Artificial Intelligence in Norwegian Newsrooms During the 2025 Parliamentary Election

The GenAI Catch-22: Use of Generative

Artificial Intelligence in Norwegian Newsrooms

During the 2025 Parliamentary Election

Mari Reisjå

IT University of Copenhagen

marre@itu.dk

Anders Sundnes Løvlie

IT University of Copenhagen

asun@itu.dk

Abstract

The increasing use of Generative Artificial Intelligence (GenAI) in journalism raises

concerns about possible detrimental effects both on journalism and its democratic

function. We explore these risks through a case study of GenAI in Norwegian Newsrooms

during the 2025 parliamentary election campaign. Based on interviews with managers and

journalists over a ten-month period, we analyse how ambitious visions fared in the face of

technological and practical challenges. We highlight the risk of an internal threat stemming

from the journalists' own use of AI, contrasting the dominant focus on external

disinformation threats. We show how newsroom managers shared sociotechnical

imaginaries resulting in unrealistically optimistic beliefs about the capabilities of the

technology and the pace of development, leading to plans for audience-facing GenAI

services collapsing and giving way to more mundane uses of GenAI tools internally in the

newsrooms. Furthermore, we identify a vulnerability in the newsroom's resilience against

GenAI influence: a GenAI Catch-22. In order to monitor the GenAI tools and prevent errors

and undue influence, newsrooms rely on human expertise. But by using GenAI extensively,

the newsrooms risk a deterioration of human expertise, preventing them from monitoring

the GenAI systems adequately.

Keywords: Generative artificial intelligence, automated journalism, data journalism, robot

journalism, elections, newsroom studies

1. Introduction

Generative Artificial Intelligence (GenAI) has rapidly become a part of everyday life in

many newsrooms (Simon 2024; de-Lima-Santos et al. 2025; Dodds, Ngai Yeung, et al.

2026; Dodds, Zamith, et al. 2026). While some view this integration optimistically as a

technology-driven advance (Opdahl et al. 2023; Kotliar 2026), others have raised

concerns about the risk of detrimental effects on democracy (Nemitz 2018; Kreps and

Kriner 2023) as well as journalism (Mahony and Chen 2025; Møller et al. 2025).

Previous research has provided valuable insight into how news actors assess GenAI and

which areas of use are most widespread (Dodds, Ngai Yeung, et al. 2026; de-Lima-Santos

et al. 2025; Fridman et al. 2025; Amponsah and Atianashie 2024). Nevertheless, there is

still limited empirical knowledge about how these technologies are applied in journalistic

practice over time, and especially in situations where the democratic function of

journalism is at stake. Much public discourse and hype focus on grand visions of rapid

technological change which can lead both to optimistic views of the benefits the

technology may bring about, as well as concern for its detrimental side effects. Meanwhile,

there is a risk that such grand visions make us blind to small and gradual changes which

are introduced without adequate caution. For instance, many scholars have already noted

that use of GenAI may lead to increased requirements for quality assurance, which may

raise further concerns for editorial autonomy and ethical integrity (Amponsah and

Atianashie 2024; Simon 2024; Shi and Sun 2024; Dodds, Ngai Yeung, et al. 2026).

This study explores the implications of journalistic reliance on GenAI using the

Norwegian parliamentary election in 2025 as a crucial case (Gerring 2007). A national

election campaign is a situation where the democratic role of the news media is crucial,

and where routines for editorial control and quality assurance are put to the test. This

makes election coverage ideal as a crucial case: If unintended technological influence or

errors occur during an election campaign, it is highly likely that the same will also happen

during normal times when journalists and editors are less on guard.

While trust in the news media have been on decline in many parts of the world, Norwegian

news media still enjoy a high level of trust from the population (The Norwegian Media

Authority 2025). The major national news providers function as shared arenas for

information and debate which gather broad audience sections across party-political

divides. Following international concerns about the impact of AI on elections (Jungherr

et al. 2026; Stockwell et al. 2024), in 2025 a government-appointed expert group

highlighted professional journalistic news media as a key element of society's resilience

against AI-enhanced disinformation (Schia et al. 2025). At the same time, all the largest

news organisations in Norway were experimenting with how to apply GenAI in their work

processes in general, and the election coverage in particular. Thus, while news media were

expected to act as a democratic counterweight to AI-related risks, they were

simultaneously becoming extensive users of the very same technologies.

We explore the tension between the news media’s role as a democratic safeguard against

AI risks and their own use of GenAI through a longitudinal qualitative study following

journalists and editors in four key Norwegian news organizations: the two main public

broadcasters, NRK and TV2, as well as the two largest tabloid newspapers, VG and

Dagbladet. We trace how ambitious initial GenAI visions collapsed and gave way to more

mundane internal uses. Furthermore, we identify a structural vulnerability in newsroom

oversight, a GenAI Catch-22: While newsrooms depend on human expertise to safeguard

against GenAI-induced errors and bias, the use of GenAI leads to a risk of internal deskilling and competence deterioration, ultimately undermining the human oversight

required to monitor GenAI systems effectively.

2. Related work

2.1. Concerns about GenAI as a threat to democracy

GenAI technologies have received much attention since the launch of ChatGPT in 2022,

often described as a hype (Kotliar 2026). The AI boom has also caused much concern,

and researchers warn that AI can rob us of democratic control (Nemitz 2018; Kreps and

Kriner 2023). Such threats may be divided into two main categories: External and

internal threats. Threats stemming from external forces using AI have been the subject

of much attention, often focused on the role of AI in disinformation and manipulation of

elections (Bontridder and Poullet 2021; Stockwell et al. 2024; von Sikorski and Hameleers

2025). The Norwegian expert group mapping risks ahead of the 2025 election similarly

focused on external threats, and highlighted the role of journalistic institutions in

countering such threats (Schia et al. 2025, 67), in particular through fact-checking and

verification, which has had a rising importance within journalism, and which AIgenerated content may further increase the need for (Steensen et al. 2024; Kavtaradze

and Kalsnes 2024).

Internal threats, on the other hand, refers to the risk that participants in a democracy –

including journalists – can be influenced through their own use of GenAI systems. Such

threats are difficult to measure and define, as they rest on the sometimes subtle power

that technology can have to influence our actions, thoughts and values (Jungherr and

Schroeder 2023; Nandini et al. 2024). There is much concern about how large language

models (LLMs) – which form the basis of most contemporary GenAI systems, such as

ChatGPT – can perpetuate stereotypes, misunderstandings and derogatory forms of

expression from the training material, which may disproportionately affect vulnerable

and marginalised groups (Bender et al. 2021; Dodge et al. 2021; Gallegos et al. 2024). The

trustworthiness of LLM outputs is also much debated, as LLMs are prone to give

convincing-looking but inaccurate outputs, sometimes referred to as hallucination

(McIntosh et al. 2024; Huang et al. 2025; Zhang et al. 2025). In spite of their unreliability,

experimental studies have shown that LLMs may be highly effective in influencing

people's opinions (Hackenburg et al. 2025; Salvi et al. 2025; Costello et al. 2026; Steyvers

et al. 2025). Studies have also demonstrated a troublesome tendency for people to

overtrust outputs from LLMs, even in critical domains such as medicine, finance and

military applications (Klingbeil et al. 2024; Holbrook et al. 2024; Shekar et al. 2025). This

raises the possibility that use of GenAI may influence news journalists in ways that may

occur hidden and gradually and be difficult to detect (Shi and Sun 2024).

2.2. GenAI in the newsrooms

GenAI is often presented as a way to increase efficiency at a time when media economies

are under pressure, and it is argued that GenAI systems may take over simple tasks and

free up journalists for investigative journalism (Moran and Shaikh 2022; Opdahl et al.

2023; Fridman et al. 2025). Empirical research indicates that news organisations mainly

use GenAI for support and routine tasks such as transcribing interviews and processing

large data sets (Amponsah and Atianashie 2024; Simon 2024). However, while

automation may save time, several studies have shown that journalists are increasingly

given new tasks related to monitoring, control, and quality assurance of GenAI systems

(Sonni et al. 2024; Simon 2024; Møller et al. 2025). Dodds and colleagues (2026) argue

that the integration of GenAI technologies in journalistic practices can be described as a

process of controlled change, in which journalists proactively use GenAI in a deliberate

way to preserve their professional authority, with GenAI functioning as a supplement

rather than a substitute for human expertise.

Several studies discuss concerns about whether demands for efficiency may lead to

journalists becoming dependent on GenAI technology to carry out their tasks, and

whether the combination of time pressure and GenAI use may increase the risk of errors

and omissions (Simon 2024; Dodds, Ngai Yeung, et al. 2026). Several studies argue that

there is an urgent need for clear guidelines for the use of GenAI in news production

(Amponsah and Atianashie 2024; Mahony and Chen 2025). De-Lima-Santos and

colleagues (2025) analysed the AI guidelines of newsrooms in 17 countries, and found

that almost all guidelines include a strong requirement that humans must always

supervise the technology. Dodds and colleagues (2026) note that constant changes to the

AI guidelines create uncertainty among journalists about what is allowed and what is not.

Several studies highlight the need for transparency around the use of AI, in order to

ensure that the process can be controlled and the risk of errors or hidden influence from

the systems is minimized (Shi and Sun 2024; Mahony and Chen 2025; de-Lima-Santos

et al. 2025).

3. Method

This is a qualitative case study of editorial use of generative AI, focusing on a crucial case

of the type Gerring (2007) terms most likely. Coverage of a national election campaign

is such a case because we may assume that during this time, journalists and editors are

extra observant of routines and control of GenAI technology given that elections are a

critical situation for society and the press.

3.1. Sampling and collection of empirical data

The news organisations represented in the study were the four largest in the country in

terms of digital readership: The two main public broadcasters, NRK and TV2, as well as

the two largest newspapers, VG and Dagbladet. The study was designed as a longitudinal

investigation, based on interviews over a period of ten months from January to November

2025, in order to capture both the planning period ahead of the election campaign, the

coverage running up to the election on 8 September 2025, as well as capturing evaluations

and reflections in the aftermath. This has made it possible to compare the news

organisations' initial plans and visions with how the technology ended up being used in

practice.

The data consists of 13 semi-structured interviews with nine informants. Four of the

informants were managers responsible for AI adoption (hereafter: AI managers) who

were each interviewed twice: First in the planning phase in January 2025, and then again

after the election was over in October 2025. Furthermore, we had intermittent email

correspondence with the managers during the ten months period in order to follow the

developments in the newsrooms. The five remaining participants were journalists who

were responsible for an important part of the election coverage: the party selection

quizzes offered by each news website. These quizzes invite voters to answer a number of

questions about key political issues and receive a score indicating which parties that most

agrees with their own views. These services are used by a large proportion of the

population ahead of each national election. 2025 was the first year in which GenAI tools

was used in the development of the quizzes, and the five journalists were interviewed

about their practical experiences with using GenAI for this purpose. In the case of NRK

two journalists were interviewed, as one of them was in charge of the quiz and the other

was tasked with making a party guide accompanying the quiz (hereafter referred to as

NRK quiz maker and NRK party guide maker). Interviewing both managers and

journalists made it possible to examine differences between strategic ambitions and

practical implementation of GenAI.

The study was planned and carried out in line with the ethical standards of our university,

which does not require formal approval from the Institutional Review Board for this type

of study. Each interviewee signed a consent form. In the initial interviews with managers

we made agreements to keep their plans confidential until after the election, in order to

allow them to speak freely. We agreed with the interviewees that we would withhold their

names but attribute quotes with the name of their organisation and their professional role.

3.2. Analytical approach

The interviews were recorded and later transcribed for thematic analysis using a

codebook structured around three main themes with up to nine sub-codes each. The first

theme was imaginaries and justifications for GenAI use, which focused on uncovering

prevalent sociotechnical imaginaries (Jasanoff 2015) as revealed by commonalities in

statements appearing over time and across the various informants. The second theme was

practical use, control and ethics, referring to patterns that could provide insight into how

generative AI affected the journalistic process and its results, linked to theories about

power dynamics, automation and expertise (Bråten 1983; Bainbridge 1983; Klein and

Hoffman 1993). The third theme was organisation, power and vulnerabilities, which

looked at how technological infrastructures affected the news organisations' room for

manoeuvre in election coverage.

The interviews from each news organisation were analyzed pairwise, comparing the

statements made by the manager and journalist(s) in that organisation and summarising

findings for each organisation. Subsequently all the quotes were grouped thematically and

compared across organisations and roles, examining common tendencies and differences.

3.3. Positionality

While this study was conducted the first author (FA) was on study leave from her position

as journalist in one of the organisations studied, NRK. The manager interviewed was not

her direct manager and had not been involved in any of her previous projects. Agreements

were made between the FA and her managers to ensure that the project did not conflict

with the employer's principles of loyalty. The FA had also previously worked for another

of the organisations in the study, VG. Such proximity to informants is not unusual in

Scandinavian journalism research, as the professional communities are relatively small.

These connections helped the FA gain access to key informants in the newsrooms,

establish trust with the informants, and understand nuances in their descriptions of their

work practices. At the same time, it was important to make agreements that allowed the

informants to trust that sensitive information about future plans would not be shared with

competitors. The proximity to the field also entails a risk of bias. The FA has reflected on

her position throughout the research process and strived for a transparent anal…

Based on the paper, here are specific improvements AI systems can make:

1. Implement fact-check worthiness detection: AI systems should be trained to distinguish between claims that are journalistically relevant and important to verify versus those that are not. The paper shows NRK struggled with this—the system couldn't identify which of 300 claims per debate were worth fact-checking. An improved AI would have a built-in relevance filter that prioritizes claims based on political significance, novelty, and potential impact on voters.

2. Reduce politically charged language generation: The TV2 case showed GenAI generated a quiz question using working capital, a phrase with right-wing political connotations, without flagging it. Improved AI systems should have political neutrality detection—identifying terms that carry partisan weight and either avoiding them or explicitly flagging them for human review.

3. Provide source references for all outputs: The TV2 journalist preferred NotebookLM because it gave clear source references to party programs. AI systems should be improved to always cite specific sources (e.g., exact passages from party manifestos) rather than generating summaries without traceable origins, enabling journalists to verify claims quickly.

4. Detect and flag its own limitations: The NRK quiz maker abandoned GenAI because it produced serious errors that took more time to correct than doing the work manually. AI systems should be improved to self-assess confidence levels and explicitly warn users when outputs may be unreliable, especially in high-stakes domains like elections.

5. Prevent AI-on-AI quality assurance: The paper notes a risk where AI-generated content might be checked by another AI without humans knowing. Improved AI systems should embed metadata or watermarks indicating AI involvement in content creation, so downstream verification processes can detect and route such content to human reviewers.

6. Support expertise preservation: The GenAI Catch-22 describes how AI use degrades human skills needed to monitor AI. Improved AI systems should include training modes that help users maintain and develop domain expertise—e.g., by explaining reasoning, offering educational feedback, or prompting users to verify facts manually before accepting outputs—rather than fully automating tasks and causing deskilling.

7. Improve accuracy in multi-party political contexts: VG noted difficulties with Norwegian elections having 22 parties versus US elections. AI systems should be improved to handle multi-party systems with nuanced policy positions, including better training on smaller parties' programs and avoiding oversimplification.

8. Enhance transparency features: The paper shows newsrooms reduced AI labeling because AI was involved in so many processes. Improved AI systems should offer granular, automatic disclosure mechanisms—e.g., generating clear statements of AI involvement that can be embedded in articles without disrupting workflow, helping maintain audience trust.

9. Detect blind spots in coverage: The Dagbladet journalist used GenAI to ask about topics they might have missed. Improved AI systems should proactively suggest under-covered policy areas or underrepresented party positions during election coverage, helping journalists maintain balanced reporting.

10. Prevent subtle bias introduction: The TV2 journalist noted AI introduced a politically slanted question she lacked expertise to catch. Improved AI systems should undergo bias audits specifically for political neutrality, and flag any output that could be interpreted as favoring or disfavoring any party or ideology.