News avoidance is a growing issue that leads to less informed citizens and endangers democratic processes. This also poses problems in news recommender environments, as 'unpleasant' news content could be avoided through personalized algorithms. To 'lighten the user's mood', this paper investigates whether satirical re-framing of news article summaries, generated by Large Language Models (LLMs), can mitigate news avoidance by making news content more engaging. Through two online experiments ( = 89 ; = 151 ), we tested various prompting techniques, assessing the impact on user perception, humor, understanding, and news consumption choices. Results indicate that satirically re-framed summaries were perceived to be engaging and informative. Less frequent news consumers showed a stronger preference for satirical content, suggesting that satire could be a tool for reconnecting with disengaged audiences. These findings show the promise of AI-generated personalized satire as an innovative approach to reducing news avoidance.
News avoidance has become an increasingly pressing issue in recent years. It can be defined as the
action of intentional disconnection or rejection from news content [
This leads to the question: Why do people avoid news? Prior research suggests a combination of
psychological and content-based drivers [
A news category that tends to be perceived as more positive is satire. In general, satire intends to
entertain and critique, and these dimensions should interact [
This paper examines to what extent satirical re-framing of news content can persuade readers to increase their news consumption. Satire has a long history in journalism and entertainment as a tool to actively engage audiences by reporting events in a lighter, more humorous way. Examples include The Proceedings of the 13th International Workshop on News Recommendation and Analytics (INRA 2025), co-located with the 19th
CEUR
ceur-ws.org
Daily Show and Last Week Tonight, that have demonstrated how blending serious events with humor
can attract viewers [
We will examine the potential of using large language models for the re-framing task. We use such LLMs to summarize existing news articles in a more satirical tone, applying it to a news recommender system interface. The combination of satire and news recommender is an underexplored research area, which leads to the following research questions: • RQ1a: How does consuming news in a satirical format afect the user’s sense of feeling informed compared to traditional news formats? • RQ1b: What prompting style performs best on humor when re-framing news articles to be more satirical? • RQ2a: To what extent can satirical news article summaries created by AI help reduce the occurrence of news avoidance? • RQ2b: How does news consumption frequency influence the efectiveness of AI-generated satirical summaries in reducing the avoidance of news content?
Recent advances in AI have made our strategy of using satirical re-framing feasible at scale. Powerful large language models (LLMs) such as GPT-4 enable the automation of human-like text generation, as also shown by Jeng et al. [11]. However, employing AI for humor introduces specific challenges. Previous research has highlighted audience biases toward AI-generated content, indicating that users may approach such material with skepticism or low expectations, especially due to the subjective nature of humor [12]. However, when users are unaware of the source, AI-created jokes can be rated equally as human-created [12, 13]. To better understand this dynamic, we explore in detail how diferent prompt styles impact humor perception and user engagement (RQ1b). Specifically, we develop and compare three distinct strategies for creating satirical news summaries based on insights from previous studies.
News Avoidance is typically defined as the practice of limiting ones exposure to news, either in a
more general manner or within certain topics [
Both factors play a role in the overall decline in news consumption. Researchers describe this as the
“News Avoidance Paradox”: As access to news has never been easier, significant proportions of the
population are actively disengaging from it [
Satire tends to be higher in entertainment [
The role of satire in a news recommender system is very limited, as it is not mentioned in literature surveys and reviews [17, 18]. What stands out in recommender research is that by adapting the selection and ranking of news stories to each user’s interest and behavior, this improves the user experience by increasing relevance [17, 19, 18]. Important challenges are news avoidance and selective exposure, for algorithms may simply reinforce user preferences without considering what is important to inform a user. The failure to consider normative aspects of news recommendations may lead to less informed citizens [20, 21]. For example, only considering personalization may lead to the omission of harder news categories, such as economic news or war coverage, and may marginalize certain voices in the public debate [22].
Our research takes a more novel angle on personalization: rather than recommending only content that users like, we recommended content we presume they tend to avoid, but in a format they might like. This approach still respects preferences at some level, as we use personalization to identify avoided topics and then tailor how we present these topics. Instead of completely disregarding a user preference, we re-frame the content fit more to the users liking. The idea builds on the intuition that news avoiders are not indiferent to important issues but rather turned away by how its presented. By re-presenting the same factual content in a more engaging frame, we seek to explore ways to reconnect with segments of the lost audiences. Consequently, we had the following behavioral hypothesis: • H1: Users are more likely to choose satirical versions of news articles in non-preferred categories, compared to those in preferred categories.
LLMs have been used to a limited extent to re-frame news articles in the context of a simple recommender system. Jeng et al. [11] show that using ChatGPT to write news articles in a diferent emotional tone can change user preferences. In contrast, the application of LLMs for generating news satire is new. Computational humor has historically posed significant challenges due to humor’s dependency on creativity, context, and nuanced language [23]. With recent advances in natural language processing, LLMs such as GPT-4 now possess the ability to generate more sophisticated humor, closely approximating human-written jokes when prompted efectively [ 13, 12].
Previous research indicates that participants often struggle to distinguish between AI- and humangenerated humor, indicating that advanced models can produce convincing comedic content [13]. However, transparency influences a user’s perception, as they tend to judge AI-generated humorous content more critically. Building on these findings, our prompting strategies incorporate proven humor techniques from previous studies. One condition leverages lighter satire, known to maximize engagement while maintaining user trust [16]. Another employs more aggressive humor, aligning with previous observations that ChatGPT performs particularly well with bold comedic styles [13]. All prompting strategies were carefully structured, as structured prompts consistently yield the best humor output from LLMs [13]. Given the role trust plays in user reactions to AI-generated humor, we formulate the following hypothesis: • H2: A user’s general trust in generative AI positively afects the reported trust in satirically manipulated articles.
We present two studies in which we examine the efects of satirically re-framing news articles and how they compare to their original counterparts. This is the first study to use this strategy to reduce news avoidance, as well as the first study to do so in relation to news preferences and recommender systems. It is also one of the first to apply LLMs to re-frame a news article to present an alternative, new version of an existing one. We show how satirical versions of news articles could have positive efects on people who are more likely to avoid news consumption. In the first study, we tested which type of prompting technique is efective in this context, while the second examined the user’s evaluation of both original and satire-based news articles. 3. Study 1
3.1.1. Dataset In order to address RQ1a and RQ1b, we performed an online experiment designed to evaluate how satirical re-framing of news summaries using Large Language Models (LLMs) influences reader perceptions, enjoyment, and comprehension. Participants were asked to evaluate nine diferent news article summaries, each manipulated to be more satirical. We also assessed which prompting techniques are most efective to generate satirical news articles.
We created a dataset consisting of 90 news article summaries, which were evenly distributed across nine diferent categories (10 articles per category). For representativeness for U.S. audiences, categories were based on the most frequently read categories in the Washington Post (2012-2018) and the most published article categories by Hufington post (2012-2022) [ 24, 25, 26, 27]. News article summaries were selected to represent the U.S. media landscape, based on their popularity and political alignment, using rankings provided by the news aggregator Ground News [28]. Summaries of news articles, along with essential metadata, were retrieved in real time using the NewsCatcher API [29], which has been previously used for research purposes. Article summaries were manipulated into a satirical format by OpenAI’s ChatGPT-4 model. This was an advanced large language model selected due to its accessibility, cost-efectiveness, and prior success in generating humorous content in related studies [ 13]. Each summary was approximately 80 words in length, truncated consistently across original and manipulated versions to maintain comparability.
Three distinct satirical re-framing techniques were tested across the dataset. Condition A (gentle
satire) featured subtle humor and mild wordplay without significant exaggeration. Condition B ( bold
satire) employed exaggerated and playful humor inspired by traditional satirical outlets such as The
Onion (cf. [
A total of 103 participants were recruited via the crowdsourcing platform Prolific. They were residents of the United States, fluent in English, and maintained at least a 95% approval rating on the platform. The study took 12 minutes to complete, for 9 GBP per hour. We eventually retained = 89 participants, removed those who did not pass the attention check or provided incomplete data.
Participants were randomly assigned to diferent conditions. We used a mixed 2 (between) x 3 (within) design to examine the impact of satirical re-framing of news article summaries on the perceptions and comprehension of the participants. The between factor involved transparency of LLM use, where participants were told about the use of AI for summaries or not. The within factor concerned the prompting style, which varied in terms of the level and nature of humor employed. The temperature of the ChatGPT 4 model was set to a more balanced 0.6 to ensure engagement and creativity, but also in an attempt to give the diferent prompting techniques more control over the outputted content. Each participant evaluated a total of nine summaries, three from each prompting style. The presentation order of article categories and prompting styles was fully randomized for each participant to control for potential order efects.
We first inquired on participants’ demographics, news frequency use and possibly reasons for news avoidance. Then, participants in the transparent condition received a message that the summaries were manipulated to be more satirical by generative AI. All participants then evaluated nine news summaries, each corresponding to one distinct topical category. For each summary, participants responded to five questionnaire items designed to measure their perceptions and engagement with the content. Each of these items was measured using 7-point Likert scales. An attention check item appeared randomly during evaluations to confirm participants’ attention. After evaluating all summaries, participants in the transparent condition completed an additional measure assessing their beliefs about the human- or AI-generated nature of the summaries.
Participants evaluated each news article summary using five distinct measures, each designed to capture a specific aspect of their experience with the content. All measures utilized 7-point Likert scales. Perceived Enjoyment was assessed with the statement “I found this article summary enjoyable,” while Perceived Fun was measured by “I found this article summary funny.” Understanding was evaluated by asking participants to rate their agreement with the statement “I clearly understood the key message of this article summary.” The Intention to Share was measured using “I would be likely to share this article summary with a friend.” Lastly, Perceived Summary Quality was captured through the statement “This article summary is well-written.”
We analyzed the users’ experience with the AI-based re-framed news article summaries. In the full sample ( = 792 trials), participants’ perceived fun rating ( = 3.93 , = 1.90 ) did not difer from the neutral midpoint of 4, (791) = −1.05 , = .29 . In contrast, we found that perceived summary quality ratings were significantly above neutral ( = 4.90 , = 1.68 ), (791) = 15.01 , < .001 , the transparency condition did not have an impact (758.1) = −0.56 , = .57 . Perceived enjoyment ratings also exceeded the midpoint ( = 4.55 , = 1.76 ), (791) = 8.76 , < .001 , indicating that most participants found the satirical summaries engaging and entertaining despite not reporting very high humor scores. Lastly, intention to share, was significantly below the neutral midpoint of 4 ( = 3.70 , (791) = −3.98 , < .001 , 95%-CI [3.56, 3.85]). Assessing these four dimensions could in other words not find that users found the content very humorous or the opposite, but did indicate that participants viewed the content to be engaging and of high quality.
One-way ANOVAs revealed only non-significant diferences across the three prompt styles. There were small diferences for perceived fun: (2, 789) = 2.82 , = .060 , 2 ≈ .007. Perceived engagement showed a similar non-significant trend, (2, 789) = 2.53 , = .080 , while summary perceived quality neither varied across prompt types, (2, 789) = 2.27 , = .105 . However, when restricting our analysis to our “High” news-consumption cluster, consisting of a considerable portion of our data (81%), prompting condition C Free satire led to significantly more perceived fun than the conventional condition A, gentle satire (mean diference = 0.58, 95% CI [0.16, 0.99], = .003 , Tukey-adjusted). We observed no diferences in relation to Condition B, bold satire. This subgroup finding provided the primary empirical basis for selecting free satire as our preferred LLM-prompt moving forward.
Participants’ overall understanding was shown to be well above the neutral benchmark of 4 through
the understanding dimension ( = 5.61 , = 1.45 ): (791) = 31.29 , < .001 . This indicated that
participants still felt informed after receiving the content in a satirical format. Transparency was also
found to afect understanding between the conditions without ( = 5.49 ) and with transparency ( =
5.76) disclosure: Welch’s (780.5) = −2.64 , = .0085 . This suggested that participants’ understanding
was increased upon being notified about the use of AI for a news article summary. Interestingly, this
indicates that transparency about the use of AI and the satirical nature of the article modestly boosts
users’ sense of feeling informed. Participants who found the articles more funny also tended to report
higher understanding ( = .36 , 95% CI [.30, .42], < .001 ) as well as greater perceived summary quality
( = .61 , 95% CI [.56, .65], < .001 ). Perceived enjoyment was also positively linked with both perceived
summary quality ( = .77 , < .001 ) and understanding ( = .60 , < .001 ). These results suggested
that satirical re-framing not only preserved understanding and learning as shown in earlier studies
[
Our findings from Study 1 indicate that while participants did not consistently rate satirical news summaries as highly humorous, they nonetheless found them engaging, of high quality, and efective at maintaining or even enhancing their sense of being informed. Importantly, transparency regarding the AI-generated nature of the summaries did not negatively impact user perceptions. instead, it modestly improved comprehension. Given these promising results, particularly with the free satire prompting style, we proceeded to Study 2 to investigate whether this satirical approach could actively influence user choices and more directly mitigate news avoidance behaviors. 4. Study 2 To address RQ2a and RQ2b, Study 2 was designed as an online experiment to investigate how satirical summaries generated by Large Language Models (LLMs) could influence users’ selection of news content and potentially mitigate news avoidance behaviors. Participants were asked to actively choose between original and satirical versions of news article summaries based on their topical preferences.
4.1.1. Dataset We used the same dataset as in Study 1 (cf. Section 3.1.1), consisting of 90 news article summaries consisting of 80 words. Each summary was re-framed into a satirical version using OpenAI’s ChatGPT-4 model, leveraging the most efective prompting style identified in Study 1 ( free satire), characterized by greater creative freedom and a columnist persona. The complete dataset, including all prompts and original summaries, can be accessed in our repository2 for Study 2.
A total of 160 U.S. residents were recruited through the crowdsourcing platform Prolific. They were lfuent in English, maintained at least a 95% approval rating on the platform, and were compensated at a rate of 9 GBP per hour. The study took approximately 15 minutes to complete. Following the application of exclusion criteria, such as failing attention checks or submitting incomplete data, the ifnal sample for analysis included = 151 participants.
The design of Study 2 was subject to 2 (between) x 2 (within) mixed-factorial design. As in Study 1, the between factor involved Transparency, either disclosing the use of AI to participants or not. The within factor was News Preference, re-framing news articles from either preferred (top-2 selection) or non-preferred (3 least preferred) categories. Each participant evaluated five pairs of articles, which each consisted of one original and one satirical summary. The temperature for generating satirical summaries via ChatGPT-4 was consistently set at 0.6, promoting balance between creativity and coherence. The presentation order of article summaries and categories was randomized to counterbalance order efects.
We first inquired on participants’ demographic details and news consumption frequency. Afterwards, they rated their preferences across the nine topical news categories using 7-point Likert scales. Based on their responses, the system selected five articles: two from their most preferred categories and three from their least preferred categories. Each participant was then presented with pairs of summaries (original vs. satirical) and was asked to select their preferred version for each of the five articles. Participants were not initially informed about the satirical manipulation. Following their choices, participants in all conditions were informed of the manipulation in a subsequent screen and responded to additional questionnaire items designed to explore their reasons for selecting each summary and their attitudes toward the satirical summaries. Attention-check items were embedded to ensure participants’ attentiveness and validity of responses.
Each participant’s motivation was captured using several measures, all employing 7-point Likert scales. For each chosen summary, participants rated their perceptions across four dimensions: Perceived Veracity (“The article summary I chose seemed more factually accurate”), Perceived Humor (“I found the article summary I chose to be more humorous”), Perceived Entertainment (“I found the article summary I chose more entertaining”), and Perceived Understanding (“The article summary I chose was easier to understand”). Additionally, participants evaluated the satirical summaries specifically, using four measures that later formed a single factor labeled “Satire-based Engagement”: Emotional Lightness (“The satirical article summary made this topic feel less heavy or overwhelming”), Future Interest (“I would like to read about this topic in a satirical format in the future”), Topical Engagement (“The 2https://anonymous.4open.science/r/Study2-anon-9C8E/README.md. satirical version of this article summary increased my interest in the topic”), and Mood Impact (“The satirical version of this article summary had a positive impact on my mood”). Satire-based engagement was formed using a principal component factor analysis; see Table 1: All items loaded strongly on that single factor (.69 ≤ ≤ .94 ) and the factor accounted for 77% of the common variance ( = 3.08 ). Sampling adequacy was excellent (KMO = .83) and internal consistency was high (Cronbach’s = .93 ). Eventually, we formed a factor that consisted of the average across all of the items.
To study news avoidance, we divided participants into two groups based on their news consumption habits: High-frequency and low-frequency news consumers. This was based on participants’ selfreported news consumption frequency, with those reporting consuming news at least once a day categorized as high-frequency consumers, and those consuming news less often categorized as lowfrequency consumers (i.e., multiple times a week or less). We deliberately set the cutof to once daily consumption, rather than using a more strict median split, as categorizing daily news consumers into the low-frequency group would have inaccurately represented their consumption behavior. In our sample, 64 participants (42%) consumed news several times a day, another 44 (29%) consumed news once a day, while only 41 participants (27%) reported consuming news less than daily. The daily threshold resulted in an unequal split: high consumers = 108 vs. low consumers = 43 . We have accounted for the unequal group sizes with the models used in subsequent analyses.
In study 2, we examined to what extent AI-generated satirical summaries could influence the selection of news content of users and potentially mitigate news avoidance. Across the entire data set (755 trials), participants chose the satirical summary 257 times, 34% of all choices ( = 0.34 , ≈ 0.47 ), indicating that roughly one in three summaries selected overall was the AI-generated satirical version.
Before performing the main part of the analysis, we verified that the two types of summary versions were perceived as intended. Across all 755 trials, participants perceived the original summary ( orig = 5.88, SE = 0.06 to be higher in veracity than its satirical counterpart sat = 4.36, SE = 0.10: Wilcoxon = 99,174, < .001 ). The pattern was reversed for humor, where the satirical version ( sat = 5.25, SE = 0.10) was considered funnier than the original version ( orig = 2.48, SE = 0.08): = 19,697 , < .001 ).
We compared the two groups of news users ( = 43 low-frequency vs = 108 high-frequency users). Overall, low-frequency readers selected the satirical version in 43% of their 215 total trials, compared to 31% of the 540 trials among daily news consumers, which was significantly higher: 2(1) = 9.71, = .002 . This is also depicted in Figure 3.
Beyond behavior, low-frequency consumers consistently rated satire more positively. Satire was judged to make the topic feel less heavy (4.74, SD = 1.78 vs 4.27, SD = 1.85, Wilcoxon = .005 ), increased interest in the topics overall (4.36, SD = 1.94 vs 3.65, SD = 2.10, < .001 ), positively afected a participant’s mood (4.41, SD = 1.94 vs 3.77, SD = 2.02, < .001 ), and sparked a desire to read more satirical news summaries on the given topic (median 5 vs 4, ≈ 4 × 10 −6) among low-frequency consumers.
Ratings for perceived understanding were found to be similar across both groups of frequency of news use. However, both means were higher than the neutral benchmark (4) (5.50, SD = 1.20 vs 5.41, SD = 1.47, test for diference: = .88 ). Collapsing the four items into a reliable ‘satire-based engagement’ scale ( = .93 ) led to similar results: Low-frequency participants averaged 4.51, SD = 1.66 versus 3.85, SD = 1.86 for daily readers (Wilcoxon < .001 ). This can also be seen in Figure 2. In addition, Figure 4 shows how this trend continues across all categories in the study, with low frequency readers reporting larger positive efects of satire across the news spectrum. This pattern reinforce our ifndings even further.
Taken together, the behavior and attitudinal data conveyed a clear answer to RQ2a. AI-generated satirical summaries can measurably reduce news avoidance by making more disengaged readers more inclined to read news articles.
We compared ‘preferred’ and ‘avoided’ categories across participant responses. A single-parameter logistic model (prefScore = +1 for preferred, −1 for avoided) showed a reliable efect on the odds of choosing the satirical version ( = −0.16, SE = 0.08, = −2.00 , = .045 ; OR ≈ 0.85). Put plainly, participants were 7 percentage points less likely to pick satire for topics they already liked (30%) than for topics they normally avoid (37%). A confirmatory one-way ANOVA on the two groups repeated the finding ( = 4.04 , = .045 ), and a mixed-efect model with random intercepts for participants and categories again found the avoided > preferred contrast ( = −0.21, SE = 0.09, = −2.35 , = .019 ). Although, an exploratory interaction revealed that the efect is carried almost in its entirety by the high frequency news group ( = −0.52, SE = 0.20, = .008 ), whereas low-frequency readers showed no reliable diference ( ≈ 0 , > .70 ). In other words, the data supports [H1], as satire was found to not be particularly attractive when a user preferred a topic, instead it was more persuasive for less favorable categories.
We tested whether trust in generative AI difered across transparency conditions. It was found to be consistent across the non-transparent ( = 4.84, SD = 1.56; = 80 ) and transparent conditions ( = 4.72, SD = 1.58; = 71 ). Hence, further testing considered both conditions to be a single group when testing [H2], examining whether trust in AI correlated with choice. The tests revealed that higher levels of trust in AI among participants correlated with a higher probability to select the satirical version of an article: A logistic model with -scored AI-trust showed that every 1 SD increase in trust raised the odds of choosing the satirical summary by about 4% ( = 0.21, SE = 0.08, = 2.65 , = .008 , OR ≈ 1.24). Performing a median split showed the same pattern: high-trust participants selected satire in 40% of their 275 trials, compared to the 30% of 480 trails among low-trust readers ( 2(1) = 7.27, = .007 ). A full three-way model confirmed that neither the transparency, AI-trust nor news-consumption frequency efected the relationship (all interaction -values > .14). These findings suggested that greater confidence in AI reliably predicts a higher chance of embracing satirical articles.
We have investigated to what extent satirical summaries of news articles, generated by LLMs, could help news avoiders to read more news. Our findings suggest that LLM-generated news satire can help re-engage individuals that might otherwise avoid news, or some news topics. Participants in our studies generally reported the re-framed summaries to be of high quality and engaging, an efect we saw strengthened withing the users who consumed news less frequently in the second study. Notably, the appeal seemed to be strongest for articles the participants reported to dislike, and we saw positive efects when recommending disliked articles in a more satirical format. This implies how humor can in fact lower some of the psychological barriers that have been shown to lead to intentional news avoidance [14]. By presenting these disliked topics in a more lighthearted way, the satirical framing seemed to draw readers into content they might normally skip. With the common reason for avoidance being due to feeling overwhelmed or disinterested, or results indicate that more positive humorous framing might counteract those feelings.
Our outcomes resonate with previous media research on the power of humor. Prior studies had shown
that adding humor to news can boost audience engagement and in some studies was shown to even
improve the retention of information [31]. In our online experiments, participants reported improved
understanding when reading the satirical summaries, suggesting that the use of satire improved how
informed participants felt without sacrificing the informational content of the article summary. This
also suggests that information and entertainment can indeed co-exist, which adheres to some definitions
of satire [
We have also observed that humor can make the news more memorable. Another notable insight we gained, is the role of user attitudes towards AI in mediating these efects. We observed how individuals with higher trust in AI were significantly more likely to prefer the re-framed satirical versions. In direct contrast, those more skeptical towards AI tended to be less receptive to the manipulated content. Our interpretation for this would be that those who report more trust in AI are more open to its more creative and non-traditional outputs, while the skeptics might question the quality and the trustworthiness. This does highlight that beyond the content itself, audience bias towards AI will have an efect on the overall success of similar applications of LLMs on online news content. These insights show us how applications of similar services will need to take into account the individual diferences of users, while some might embrace this untraditional way of receiving news content, others might not.
Another interesting finding was how our transparency condition about the use of AI, was in most cases found to not have an efect. In both studies, clearly disclosing that the summaries was re-framed by AI to be more satirical had minimal efect of any of our main measurement dimensions. These results to some extent counters earlier research, which found that when the readers know AI is involved in producing news, their trust in the content often drops [ 32]. One study reported that simply suspecting that the content was generated by AI lowered readers’ credibility judgments [32]. However, in our context, informing participants that the content was “re-framed to be more satirical by a large language model such as ChatGPT” has not significantly afected their engagement or trust. In a similar vein, we have also found that the transparency condition has had no efect on the users’ reported humor, also contradicting earlier research that found that users reported jokes and humorous content to be less funny when they were told that it was created by AI [12].
One possible explanation for this could be the nature of our sample and content. Participants sourced online could perhaps be more accustomed to AI in their daily lives, as it is in the time the studies were conducted highly relevant and spoken about. This could lead to the AI-disclosure tag not being as alarming as originally anticipated. The satirical tone of the articles might also have signalized some form of departure from traditional journalism, efectively “resetting” expectations to some degree. Another potential impact factor could be the fact that the LLMs role was only to amplify the entertainment value, not to deliver hard facts on its own, maybe making users less critical of the use of AI. Lastly, it should be noted that perceptions are prone to evolving over time. The use of AI and LLMs has become much more common for the average user in recent years. As the field evolves at a staggering speed, it is natural to expect corresponding increase in users trust and acceptance, a trend our results indicate may already have begun within certain groups.
We would like to discuss a few limitations to our study. First and foremost, our participant sample mostly consisted of active news consumers, while very persistent news avoiders were underrepresented. We encouraged participation by news avoiders in the second study by explicitly mentioning this in the crowdsourcing invitation, but this has likely only made a small diference. However, the fact that low-frequency news readers in our sample found the satire to be especially engaging illuminates the positive efects of satire, but it could be that this strategy is less efective among even less frequent news consumers (i.e., less than once a week).
Another important issue is the factual accuracy of the manipulations. While we explicitly instructed
the LLM to maintain “every factual detail”, we lacked a formal method to verify that every important
detail was in fact preserved and no misinformation slipped in. Satire, by its nature, exaggerates and
twists aspects of reality to raise the entertainment value [
To validate our results, we would propose a study where users can actively choose to consume ‘regular’ or satirical news. We envision a news recommender website where users can choose between reading an original news article or a satirical version, and even switch back and forth. Multiple news recommendations like this could be combined into a single interface, presenting a more realistic, out-of-lab scenario. Ultimately, determining whether news avoiders would like to remain on such a news website is a long-term objective.
This research was supported by the Research Council of Norway with funding to MediaFutures: Research Centre for Responsible Media Technology and Innovation, through the Centre for Research-based Innovation scheme, project number 309339.
During the preparation of this work, the authors used ChatGPT in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content. [9] V. L. Rubin, N. Conroy, Y. Chen, S. Cornwell, Fake news or truth? using satirical cues to detect potentially misleading news, in: Proceedings of the Second Workshop on Computational Approaches to Deception Detection, 2016, pp. 7–17. [10] N. J. Stroud, A. Muddiman, Selective exposure, tolerance, and satirical news, International Journal of Public Opinion Research 25 (2013) 271–290. [11] J. H. Jeng, G. Kasangu, A. Starke, K. M. A. Seddik, C. Trattner, The role of gpt as an adaptive technology in climate change journalism, in: Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization, 2025, pp. 214–223. [12] N. N. Joshi, Evaluating human perception and bias in ai-generated humor, in: Proceedings of the 1st Workshop on Computational Humor (CHum), Association for Computational Linguistics, 2025, pp. 79–87. [13] D. Gorenz, N. Schwarz, How funny is chatgpt? a comparison of human- and a.i.-produced jokes,
PLOS ONE 19 (2024) e0305364. doi:10.1371/journal.pone.0305364. [14] R. Vliegenthart, K. de Bruin, Y. de Haan, S. Kruikemeier, The news avoidance paradox? exploring the relationship between news repertoires and intentional news avoidance, Communications (2025). doi:10.1515/commun- 2024- 0062, published online March 15, 2025. [15] A. B. Becker, L. Bode, Satire as a source for learning? the diferential impact of news versus satire exposure on net neutrality knowledge gain, Information, Communication & Society 21 (2018) 612–625. [16] C. Burgers, B. Brugman, How satirical news impacts afective responses, learning, and persuasion:
A three-level random-efects meta-analysis, Communication Research 49 (2022) 966–993. [17] M. Karimi, D. Jannach, M. Jugovac, News recommender systems–survey and roads ahead, Information Processing & Management 54 (2018) 1203–1227. [18] S. Raza, C. Ding, News recommender system: a review of recent progress, challenges, and opportunities, Artificial Intelligence Review (2022) 1–52. [19] X. Meng, H. Huo, X. Zhang, W. Wang, J. Zhu, A survey of personalized news recommendation,
Data Science and Engineering 8 (2023) 396–416. [20] N. Helberger, On the democratic role of news recommenders, in: Algorithms, Automation, and
News, Routledge, 2021, pp. 14–33. [21] S. Vrijenhoek, L. Michiels, J. Kruse, A. Starke, N. Tintarev, J. Viader Guerrero, Normalize: The first workshop on normative design and evaluation of recommender systems, in: Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 1252–1254. [22] S. Vrijenhoek, G. Bénédict, M. Gutierrez Granada, D. Odijk, Radio*–an introduction to measuring normative diversity in news recommendations, ACM Transactions on Recommender Systems 3 (2024) 1–29. [23] S. Kim, L. B. Chilton, Ai humor generation: Cognitive, social and creative skills for efective humor, arXiv preprint arXiv:2502.07981 (2025). URL: https://arxiv.org/abs/2502.07981. [24] S. O. Larsen, Developing and Comparing Similarity Functions for the News Recommender Domain Using Human Judgments, Master’s thesis, University of Bergen, 2024. URL: https://bora.uib.no/ bora-xmlui/bitstream/handle/1956/23059/Master-thesis-final-sebastian-overhaug-larsen.pdf? sequence=1&isAllowed=y, available via BORA (University of Bergen Open Research Archive). [25] R. Misra, News category dataset, arXiv preprint arXiv:2209.11429v3, 2022. doi:10.48550/arXiv.
2209.11429, submitted September 23, 2022; revised October 6, 2022. [26] N. Newman, R. Fletcher, A. Schulz, S. Andı, C. T. Robertson, R. K. Nielsen, News readers really want to read: Relevance works—news audiences in an era of choice, https://www.digitalnewsreport. org/publications/2019/news-readers-really-want-read-relevance-works-news-audiences/, 2019. [27] A. D. Starke, V. R. Solberg, S. Øverhaug, C. Trattner, Examining the merits of feature-specific similarity functions in the news domain using human judgments, User Modeling and User-Adapted Interaction 34 (2024) 995–1042. [28] Ground News, Methodology: Bias ratings & factuality scores, https://ground.news/rating-system#, 2025. Retrieved June 25, 2025. [29] NewsCatcher API, Product overview, https://www.newscatcherapi.com/product-overview, 2025.
Web page. Retrieved June 25, 2025. [30] M. A. Xenos, A. B. Becker, Moments of zen: Efects of the daily show on information seeking and political learning, Political Communication 26 (2009) 317–332. [31] J. C. Coronel, M. B. O’Donnell, P. Pandey, M. X. Delli Carpini, E. B. Falk, Political humor, sharing, and remembering: Insights from neuroimaging, Journal of Communication 71 (2021) 129–161. doi:10.1093/joc/jqaa041. [32] H. Jia, A. Appelman, M. Wu, S. Bien-Aimé, News bylines and perceived ai authorship: Efects on source and message credibility, Computers in Human Behavior: Artificial Humans 2 (2024) 100093. doi:10.1016/j.chbah.2024.100093.