Satirical news is a unique form of misinformation: it is intended to be humorous by the original publication source but can be misinterpreted. Explanation of satirical news articles is challenging because recent and local context needs to be incorporated. This is a pilot study that explores explanation generation for satirical news articles from Australia and compares them alongside those from USA. The paper demonstrates that framing explanations on unreliable narrator theory on open-weights models can be a feasible, cost-effective alternative to frontier models. We find that models identify the satirical tropes more accurately on the American source than the Australian one. This work results in a novel dataset and an evaluation benchmark for satirical explanation.
Figures & tables
Figure 1: Unreliable narrator framing of an Australian satirical news article. The satire emerges from the unreliable narrator’s surface account diverging from the implied author’s critique of performative allyship.
Figure 2: Satirical gap model based on Booth (1961) .
Figure 3: Both prompting conditions share the same retrieval scaffold. The model may emit a SEARCH: directive and receive the top three DuckDuckGo results before producing its REAL_EVENT and FINAL blocks.
Betoota (AU)
The Onion (US)
Model
Prompt
target
expl.
target
expl.
Qwen2-VL-7B
base
2.00
2.00
2.01
1.86
un
1.93
2.46
1.96
2.38
Mistral-7B
base
1.76
2.75
2.09
2.45
un
1.97
3.02
2.14
3.02
Qwen2.5-7B
base
2.03
2.31
2.28
2.19
Table 1: Mean judge scores (1–5) under the baseline ( base ) and unreliable-narrator ( un ) prompts. Open-weight models were run with n≤500 , generations that did not produce a well-formed "FINAL" block are excluded. ∗ Frontier models were run on Betoota only with n=100 .
Model
trope
surface
author
mech.
Betoota Advocate (AU)
Qwen2-VL-7B
0.29
3.37
2.54
2.27
Mistral-7B
0.25
3.94
3.55
2.84
Qwen2.5-7B
0.35
3.80
3.16
2.75
Llama-3.1-8B
0.46
3.88
3.87
3.25
GPT-5.6 Luna
0.95
4.96
5.00
5.00
Table 2: Mean scores on the four unreliable-narrator ( un ) dimensions: trope accuracy (0–1), surface accuracy, implied author belief, and satirical mech anism (each 1–5). Bold marks the best open-weight model per column. Frontier models were run on Betoota only, n=100 .
Dimension
n
human
judge
ICC (2,k)
ρ
satire target
50
2.70
1.98
0.78
0.68 ∗
explanation
48
3.24
3.00
0.36
0.68 ∗
trope
24
0.70
0.44
0.22
0.76 ∗
surface
24
4.01
3.64
-0.11
0.30
implied author
23
3.49
3.36
0.49
0.62 ∗
mechanism
23
3.40
2.84
0.52
0.44 ∗
Table 3: Human evaluation. human is the mean of three annotators; ICC (2,k) is the two-way random effects intraclass correlation for that mean; ρ is Spearman correlation between the aggregated human score and the judge, with ∗ marking p<0.05 .
Input Headline: “Girlfriend Takes A Pause From Scrolling To Declare She’s Going To Read 20 Books This Year”
Baseline Explanation (2 / 5). The satire is about the superficial nature of New Year’s resolutions and the tendency to prioritize social media over personal growth.
UN Explanation (5 / 5). …The surface account presents Bianca Sutton as a woman who has decided to break away from her TikTok addiction and read 20 books this year. However, the implied author, who is likely a satirist, actually sees this as a continuation of her usual behaviour, where she spends most of her time on social media. The gap between the narrator’s presentation of a serious change in behaviour and the implied author’s recognition of it as a superficial and ultimately unchanged pattern creates the satirical effect.
Table 4: An example where a baseline prompt fails to produce a response that explains the satire.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Source
Naif
Clown
Madman
Picaro
Llama-3.1-8B
AU
354
65
47
34
US
411
39
37
3
Qwen2.5-7B
AU
307
82
77
31
US
347
56
76
20
Mistral-7B
AU
35
68
295
98
US
44
92
343
21
Appendix
Table 5: Narrator tropes assigned under the unreliable narrator prompt, n=500 per source. Each model applies a consistent preference across both sources, but the preferences differ between models.
Input Headline: “Struggling Winter Olympics Minnow Nation Surprises With First Gold Medal”
Satire target (1 / 5). Mikaël Kingsbury, a Canadian freestyle skier, won the gold medal in men’s moguls at the 2026 Winter Olympics in Milano Cortina, Italy.
Explanation (4 / 5). The Betoota Advocate employs the naif narrator trope to present a heartwarming story of a struggling nation, Canada, finally achieving a rare Winter Olympics victory. The implied author, however, recognises the true significance of Kingsbury’s achievement, as he is one of the most decorated athletes in the history of moguls skiing. The gap between the narrator’s portrayal of a fairy tale underdog story and the implied author’s recognition of Kingsbury’s true talent and legacy creates a satirical effect.
Appendix
Table 6: An unreliable narrator output in which the explanation is coherent and scores 4, while the satire target scores 1. The model names a real athlete and a real event, but accepts the article’s premise that Canada is a minnow nation rather than recognising it as the joke.
target
expl.
Model
Prompt
raw
+ sc
raw
+ sc
Betoota Advocate (AU)
Qwen2-VL-7B
base
2.00
1.64
2.00
2.04
un
1.93
1.74
2.46
2.48
Mistral-7B
base
1.76
1.55
2.75
3.13
un
1.97
1.61
3.02
2.98
Appendix
Table 7: Effect of the sentence-classification pre-pass (+ sc ) against raw article text. Bold marks the better of each pair.
Score
Satire Target Score ( satire_target )
Explanation Score ( satire_explanation )
1
No real satire target identified, or entirely wrong / hallucinated.
Irrelevant, hallucinated, fails to explain, or merely restates the headline.
2
Correct general topic, but wrong specific people, dates, or details.
Identifies the target, but fails to explain how or why it is satirical.
3
Correct target with minor factual inaccuracies (e.g., wrong date, right person).
Explains the mechanism, but misses important cultural or political nuance.
4
Correct target with all key people and context; trivial differences only.
Explains target, mechanism, and most cultural context with minor gaps.
5
Fully accurate and specific: correct target (including, as applicable, event, people, dates, and AU context.)
Complete explanation: target, satirical device, and cultural critique.
Appendix
Table 8: Evaluation rubric for satirical article comprehension across satire target grounding and satirical mechanism explanation.
Input Headline: “Corporations That Plaster Themselves In Rainbows Every Year Not Keen To Bail Out Mardi Gras Sadly”
Baseline Prompt
Unreliable Narrator
Score: 2 / 5
Score: 5 / 5
The satire is satirizing the hypocrisy of large corporations, such as banks, tech companies, and food giants, who use the pride flag for branding purposes but are unwilling to support the Mardi Gras event financially. The article suggests that these corporations are more interested in profit than in genuine support for the LGBTQ+ community.
The Betoota Advocate employs the naif narrator trope to present the situation as a simple failure of corporate sponsorship , while the implied author is likely satirizing the hypocrisy of corporations using rainbow-themed branding for profit while not actually supporting the community. The gap between the narrator’s surface account and the implied author’s critique of performative rainbow capitalism creates the satirical effect.
Names the target, then stops without explaining the satirical device
Names all four parts: trope, surface, author, gap
Appendix
Table 9: Comparison of qualitative outputs and evaluation scores between the baseline prompt and the unreliable narrator framework.
Input Headline: “Girlfriend Takes A Pause From Scrolling To Declare She’s Going To Read 20 Books This Year”
Baseline Prompt
Unreliable Narrator
Target: 3 / 5 Explanation: 2 / 5
Target: 1 / 5 Explanation: 5 / 5
Satire target. The article is a satire about a woman named Bianka Sutton who claims she will read 20 books this year, but instead of actually reading, she continues to scroll through her TikTok feed.
Satire target. Bianca Sutton is a social media influencer with over 38K followers on Instagram, where she posts content about her personal life, including her relationship with her boyfriend, Peter Evans. She is also a professional in the recruitment industry, as indicated by her LinkedIn profile.
Explanation. The satire is about the superficial nature of New Year’s resolutions and the tendency to prioritize social media over personal growth.
Explanation. The Betoota Advocate employs the naif narrator trope in this article. The surface account presents Bianca Sutton as a woman who has decided to break away from her TikTok addiction and read 20 books this year. However, the implied author , who is likely a satirist, actually sees this as a continuation of her usual behaviour, where she spends most of her time on social media. The gap between the narrator’s presentation of a serious change in behaviour and the implied author’s recognition of it as a superficial and ultimately unchanged pattern creates the satirical effect.
Names the target but does not explain the device; summarises the article’s plot rather than the phenomenon.
Names all four parts, but invents a real-world identity for a fictional character.
Appendix
Table 10: A second qualitative comparison. The unreliable narrator prompt raises explanation from 2 to 5, but satire target falls from 3 to 1 because the model fabricates biographical detail for a character the article invented.
Input Headline: “Disappointed Man Reaches Bottom Of Ice Cream Carton Right When He Was Hitting His Stride” (The Onion)
Baseline Prompt
Unreliable Narrator
Satire target. Billy Crenshaw, a local man in Clemson, SC, is disappointed after reaching the bottom of an ice cream carton just as he was hitting his stride.
Satire target. A man, Billy Crenshaw, is disappointed after finishing an ice cream carton because he had just started feeling satisfied with the dessert.
Explanation. The article is satirizing the idea that people can become so engrossed in eating ice cream that they become disappointed when they finish the carton.
Explanation. The satire uses the naif narrator trope , presenting Billy Crenshaw’s disappointment as a serious and relatable experience. The implied author , however, is likely to see the situation as humorous and exaggerated. The gap between the narrator’s surface account of Billy’s disappointment and the implied author’s recognition of the absurdity of his reaction creates the satirical effect.
Judge feedback on the satire target: “Restates the fictional character and his ice cream carton as if it were literal fact, instead of identifying the real satirical target: sports-metaphor language applied to overeating.”
Appendix
Table 11: Both prompts restate the article’s invented character as the satire target and score 1 on that dimension. The unreliable narrator prompt still produces a structured explanation, but the structure is applied to the fiction rather than to the phenomenon the article is mocking.
News outlets covering the same event differ in what they emphasize and omit. Same Story, Different Lens is an experimental research prototype for comparing these choices. It collects world and politics coverage from 17 outlets, groups reports about the same event, and produces structured analyses of sentiment, political lean, framing, loaded language, possible omissions, and model-estimated proportions of factual claims and opinion. Loaded-language entries include quotations and neutral rewordings. Three model roles (agents) handle article analysis, assessment of source support, and cross-outlet briefing generation. Analyses are stored in a JSONL archive. The evaluation compares outlet-level lean estimates with approximate, predefined reference ratings and summarizes the Critic's model-assigned grounding scores. These measures provide preliminary diagnostics. A grounding score reflects the model's assessment of support in the supplied text and does not establish factual accuracy or constitute independent verification.
Affan Yasin, Yueqi Lyu, Rubia Fatima +2
Academy of AI and Advanced Technology (AIAT), Xi’an Jiaotong-Liverpool University (XJTLU), Suzhou, China · Department of Computer Science and Technologies, Emerson University, Multan, Pakistan.
Humor generation remains challenging task for Large Language Models (LLMs), due to their subjective nature. We focus on satire, a form of humor strongly shaped by context. In this work, we present a novel pipeline for grounded satire generation that uses Retrieval-Augmented Generation (RAG) over current news to produce satirical dictionary definitions in the Finnish context. We also introduce a new task-specific evaluation framework and annotate 100 generated definitions with six human annotators, enabling analysis across multiple experimental conditions, including cultural background, source-word type, and the presence or absence of RAG. Our results show that the generated definitions are perceived as more political than humorous. Both topic-based word selection and RAG improve the political relevance of the outputs, but neither yields clear gains in humor generation. In addition, our LLM-as-a-judge evaluation of five state-of-the-art models indicates that LLMs correlate well with human judgments on political relevance, but perform poorly on humor. We release our code and annotated dataset to support further research on grounded satire generation and evaluation.
Natural-language explanations are often treated as a unified interface for understanding model behavior, but different explanation sources may support simulation in different ways. This paper compares two families of explanations for question answering models: verbalized feature attributions and self-generated rationales. We evaluate them under a shared counterfactual simulation setting, using an LLM judge as predictor and measuring whether it can better predict a model's answers to follow-up questions when given its explanation. Across multiple instruction-tuned models, we analyze how explanation source, verbalization strategy, and feature granularity affect the simulatability of explanations. Our results show that explanation format and granularity affect simulatability: attribution-based explanations and self-generated rationales differ in how much they improve counterfactual prediction, with effects that vary across models and formats.
Pingjun Hong, Benjamin Roth
Faculty of Computer Science, University of Vienna, Vienna, Austria · UniVie Doctoral School Computer Science, University of Vienna, Vienna, Austria · Faculty of Philological and Cultural Studies, University of Vienna, Vienna, Austria