Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

AI Humor Generation: Cognitive, Social and Creative Skills for Effective Humor

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Giving LLMs human-like skills produces meme captions that blind raters find almost as funny as top human-written Instagram comments.

desk verdict A solid, modest empirical advance in LLM humor generation, undercut by an equivalence claim that p=0.053 cannot support. read the letter →

arxiv 2502.07981 v1 pith:MGNVLLJW submitted 2025-02-11 cs.HC

Sean Kim , Lydia B. Chilton This is my paper · ORCID
classification cs.HC
keywords AIhumorgenerationlargelanguagemodelsInstagramcaptionsGenZcomputationalcreativitynarrativeextrapolationfine-tuningevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that humor is not too human for AI once it is decomposed into learnable skills: careful observation of an image, generating multiple angles, extrapolating relatable social conflicts, and ranking outputs with an audience-aware judge. The authors build HumorSkills, a pipeline that combines these steps with fine-tuning on Gen Z Instagram caption humor, and test it in blind rating studies. In the main study, HumorSkills captions were rated 2.27 out of 5, significantly funnier than a GPT-4o baseline (about 0.21 points lower) and only 0.08 points below the top-rated human Instagram captions, a gap the paper treats as not statistically significant (p=0.053). If correct, the result matters because it suggests AI can produce audience-specific humor that resonates with people, with implications for creative tools, marketing, and the social risks of AI-generated intimacy and deception.

What carries the argument

The mechanism is a staged pipeline, HumorSkills, rather than a single prompt. First, a vision-language model performs Visual Detail Extraction and Visual Humor Ideation to find non-obvious, funny visual elements. Next, Narrative and Conflict Extrapolation generates relatable Gen Z conflicts, such as work, school, and relationship struggles, that are analogous to the image, expanding the joke space beyond the literal content. A fine-tuned GPT-3.5 then generates 30 image-focused and narrative-driven captions, and a fine-tuned GPT-4o 'Gen Z humor expert' agent ranks them and returns the top five. This divergent-then-convergent structure mirrors cognitive models of creativity and is what carries the paper's argument: each skill adds a distinct source of jokes or filtering, and the ranking agent models the target audience's taste.

What would settle it

Re-run the main comparison with a pre-registered non-inferiority test specifying a margin, say 0.1 points on the 1-5 scale; if the 95% confidence interval for the IG-minus-HumorSkills difference lies above that margin, or if a larger sample makes the 0.078-point gap statistically significant, the parity claim fails. A simpler check is to collect additional ratings for the same 20 images and see whether the p-value for the IG coefficient drops below 0.05.

Watch

Extended reading notes

Core claim

The central claim is that adding human-like skills to LLM generation makes the difference between generic AI humor and humor that a specific audience accepts as nearly human-level. Concretely, for 20 images taken from popular Instagram humor accounts, the HumorSkills pipeline produced captions that blind Gen Z raters scored as funnier than captions from GPT-4o with prompt engineering (2.27 vs. about 2.06, p<0.0001) and statistically indistinguishable from the five most-upvoted human comments on the same images (0.08-point gap, p=0.053). The paper also shows the system beats GPT-4o on out-of-domain images, including camera-roll photos and museum art, arguing the effect is not simply overfitting to the training domain.

Load-bearing premise

The conclusion that HumorSkills is as funny as the top Instagram captions rests on treating a non-significant p-value (p=0.053) as evidence of equivalence, which a standard null-hypothesis test cannot do.

Editorial extensions

If this is right

  • For the tested format and audience, a skill-decomposed LLM pipeline can produce humor that users like almost as much as top human-written captions, suggesting humor generation can be engineered rather than left to chance.
  • The system generalizes beyond its training images: it beats the GPT-4o baseline on camera-roll photos and museum art, so the approach is not simply memorizing Instagram posts.
  • A fine-tuned LLM-as-judge can act as a proxy for audience taste, selecting the most relatable captions from a larger generated pool.
  • If these results hold, AI could generate captions, memes, and social media content for a specific demographic at scale, changing who produces viral humor.
  • The near-parity with human captions also brings a cautionary corollary the paper draws: AI humor could be used to fabricate social bonding or enhance scams, making genuineness harder to verify.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the p=0.053 result is not proof of parity; a pre-specified equivalence test with a small margin, such as 0.1 points, would settle whether the gap is practically negligible, and the current conclusion leans on a null result.
  • Editorial inference: the paper does not isolate which skill contributes most; an ablation study removing narrative extrapolation or the ranking agent would determine whether all three are needed or whether one dominates.
  • Editorial inference: the approach likely transfers to other subjective creative domains where audience taste is knowable, such as memes, advertising slogans, or product names, but only if a corpus of audience-vetted examples exists.
  • Editorial inference: because raters were university students self-identifying as Gen Z and were paid, the audience model is a convenience sample, and the parity claim may shift with a broader or older Gen Z sample.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents HumorSkills, a pipeline for generating Gen Z meme captions that combines GPT-4o-based visual detail extraction, visual humor ideation, narrative and conflict extrapolation, fine-tuned caption generation with GPT-3.5, and a fine-tuned LLM-as-judge ranking step. The authors evaluate the system on Instagram images by having human raters score captions from three conditions: top-voted Instagram comments, a prompt-engineered GPT-4o baseline, and HumorSkills. They also report two out-of-domain studies on Flickr camera-roll images and museum art, comparing only HumorSkills against GPT-4o. The main results are that HumorSkills significantly outperforms the GPT-4o baseline (p<0.0001 on the target images) and is described as 'as funny as' the top Instagram captions, based on a non-significant difference (p=0.053) in the opposite direction.

Significance. If the parity claim were supported, the paper would make a noteworthy contribution to computational humor and HCI, showing that a skill-augmented LLM pipeline can approach top human-written captions in a blind rating study. The system design is thoughtful, combining visual observation, divergent narrative generation, and audience modeling through fine-tuning; the three-study structure and the use of external human raters are strengths. The out-of-domain comparisons add useful evidence about generalizability. However, the central parity claim is not established by the reported statistics, and a potential overlap between fine-tuning and evaluation data creates a validity risk that the paper does not address. The contributions are therefore conditional on re-analysis and additional reporting.

major comments (3)
  1. [§4.2.2, Table 1]
  2. [§3.1.6 vs. §4.1]
  3. [§4.2.2, sentence on effect size]
minor comments (6)
  1. [§4.2.1]
  2. [§2.1, reference [27]]
  3. [§4.2 vs. Table 1 heading]
  4. [§4.5]
  5. [§5.2]
  6. [§7]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the HumorSkills evaluation rests on external human ratings, and the p=0.053 equivalence concern is a statistical inference issue, not a circular reduction.

full rationale

This paper makes no mathematical derivation claim; its central evidence is a blind humor-rating survey. The HumorSkills pipeline is fine-tuned on top Instagram comments and uses an LLM-as-judge to select captions, but the outcome measure—human funniness ratings on 6,015 observations—is external to the system and is not defined in terms of the system's own outputs. The closest candidate for circularity, that fine-tuning on human-written IG captions makes the later comparison to IG captions tautological, is not a circular reduction: the system generates new captions, and human raters score them independently. The paper's H2 conclusion ('rated as funny as the top-rated Instagram captions') rests on a non-significant p-value (p=0.053), which is a statistical inference error (a null result cannot establish equivalence), but this is not a circularity pattern covered by the schema. There are no self-citations that carry the argument, no imported uniqueness theorems, and no fitted parameter renamed as a prediction. Therefore the paper is self-contained with respect to circularity, even though its statistical interpretation may be contestable.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities and no fitted theory parameters. Its load-bearing assumptions are empirical: the validity of upvotes as ground truth, the representativeness of the raters, the treatment of ordinal ratings, and the reliability of the fine-tuned judge.

assumptions (5)
  • domain assumption Upvoted Instagram comments are a valid ground truth for humorousness.
    The main study compares HumorSkills to the top five most-upvoted comments on the same images, assuming upvotes reflect funniness for the Gen Z audience.
  • domain assumption University students who self-identify as Gen Z and like Instagram caption humor represent the target audience.
    Raters were recruited from clubs and classes at a local university with self-reported Gen Z identity and interest in caption humor, per §4.1.
  • domain assumption A 1 to 5 Likert rating and a mixed linear model on ordinal ratings adequately measure funniness.
    The GLMM treats ordinal ratings as continuous responses, which is a common simplification but not statistically exact, see §4.2.
  • domain assumption Fine-tuning on 80 comments from three Instagram meme pages is sufficient to teach Gen Z humor style.
    The fine-tuning dataset is small and no validation is reported, see §3.1.6.
  • ad hoc to paper The fine-tuned LLM judge can reliably rank captions by Gen Z humor quality.
    The caption ranking agent is introduced without any validation of its ranking quality, see §3.1.5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI Humor Generation: Cognitive, Social and Creative Skills for Effective Humor." pith.science (2026). https://pith.science/paper/MGNVLLJW

@misc{pith2026250207981,
  author       = {Pith},
  title        = {Pith review of: AI Humor Generation: Cognitive, Social and Creative Skills for Effective Humor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGNVLLJW}},
  note         = {Machine review of arXiv:2502.07981}
}
read the original abstract

Humor is a social binding agent. It is an act of creativity that can provoke emotional reactions on a broad range of topics. Humor has long been thought to be "too human" for AI to generate. However, humans are complex, and humor requires our complex set of skills: cognitive reasoning, social understanding, a broad base of knowledge, creative thinking, and audience understanding. We explore whether giving AI such skills enables it to write humor. We target one audience: Gen Z humor fans. We ask people to rate meme caption humor from three sources: highly upvoted human captions, 2) basic LLMs, and 3) LLMs captions with humor skills. We find that users like LLMs captions with humor skills more than basic LLMs and almost on par with top-rated humor written by people. We discuss how giving AI human-like skills can help it generate communication that resonates with people.

Figures

Figures reproduced from arXiv: 2502.07981 by the authors.

Figure 1
Figure 1. HumorSkills System Diagram. Given an image, the system first extracts visual details with a visual language model, then [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. A diagram for how narrative extrapolation works [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Top Rated Image Captions (marked in green) for Instagram, GPT-4o, and HumorSkills and corresponding top scorers for each [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Images across all 3 datasets with Gen Z slang. Instagram (left), Flickr (center), Museum Art (right) [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Images across all 3 datasets with narrative generation. Instagram (left), Flickr (center), Museum Art (right) [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Images across all 3 datasets with notable visual extractions. Instagram (left), Flickr (center), Museum Art (right) [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. HumorRank: A Tournament-Based Leaderboard for Evaluating Humor Generation in Large Language Models

    cs.CL 2026-03 unverdicted novelty 7.0 of 10

    HumorRank ranks nine LLMs on textual humor using GTVH-grounded pairwise tournaments and Adaptive Swiss aggregation on the SemEval-2026 MWAHAHA dataset, finding that comedic mechanism mastery matters more than scale.

Reference graph

Works this paper leans on

37 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Anonymous. 2024. FunLMs: Methods for Fine-tuning LLMs to Generate Humor. In Submitted to ACL Rolling Review - June 2024 . https://openreview. net/forum?id=JMwjgXHLCj under review

    work page 2024

  2. [2]

    Kim Binsted and Graeme Ritchie. 1997. Computational humor. In Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI), Vol. 97. 1089–1094

    work page 1997

  3. [3]

    Sascha Brodsky. 2024. Who watches the AI watchers? The challenge of self-evaluating AI. https://www.ibm.com/think/news/ai-testing-advances. Accessed: 2025-02-06

    work page 2024

  4. [4]

    Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. 2024. A Survey on Mixture of Experts. arXiv:2407.06204 [cs.LG] https://arxiv.org/abs/2407.06204

    arXiv 2024

  5. [5]

    Judy Carter. 2001. The Comedy Bible: From Stand-up to Sitcom–The Comedy Writer’s Ultimate "How To" Guide . Touchstone, New York, New York, USA. 368 pages

    work page 2001

  6. [6]

    Megan Cerullo. 2024. https://www.cbsnews.com/news/elder-scams-family-safe-word/. https://www.cbsnews.com/news/elder-scams-family-safe- word/. Accessed: 2025-02-06

    work page 2024

  7. [7]

    Lee, Gagan Bansal, Yuan Cao, Shuyuan Zhang, Justin Lu, Jackie Tsay, Yinan Wang, Andrew M

    Mia Xu Chen, Benjamin N. Lee, Gagan Bansal, Yuan Cao, Shuyuan Zhang, Justin Lu, Jackie Tsay, Yinan Wang, Andrew M. Dai, Zhifeng Chen, Timothy Sohn, and Yonghui Wu. 2019. Gmail Smart Compose: Real-Time Assisted Writing. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (Anchorage, AK, USA) (KDD ’19). Associ...

    arXiv 2019

  8. [9]

    Greg Dean. 2000. Step by Step to Stand-Up Comedy . Heinemann Drama, Portsmouth, New Hampshire. 208 pages

    work page 2000

Show all 37 references
  1. [10]

    Goldstein

    Renee DiResta and Josh A. Goldstein. 2024. How Spammers and Scammers Leverage AI-Generated Images on Facebook for Audience Growth. arXiv:2403.12838 [cs.CY] https://arxiv.org/abs/2403.12838

    2024 arXiv

  2. [11]

    Drew Gorenz and Norbert Schwarz. 2024. How funny is ChatGPT? A comparison of human- and A.I.-produced jokes. PLOS ONE 19, 7 (07 2024), 1–13. https://doi.org/10.1371/journal.pone.0305364

    2024 doi

  3. [12]

    He He, Nanyun Peng, and Percy Liang. 2019. Pun Generation with Surprise. arXiv:1904.06828 [cs.CL] https://arxiv.org/abs/1904.06828

    2019 arXiv

  4. [13]

    Sally Holloway. 2010. The Serious Guide to Joke Writing: How To Say Something Funny About Anything . Bookshaker, Great Yarmouth, UK. 207 pages

    2010

  5. [14]

    Zachary Horvitz, Jingru Chen, Rahul Aditya, Harshvardhan Srivastava, Robert West, Zhou Yu, and Kathleen McKeown. 2024. Getting Serious about Humor: Crafting Humor Datasets with Unfunny Large Language Models. arXiv:2403.00794 [cs.CL] https://arxiv.org/abs/2403.00794

    2024 arXiv

  6. [15]

    Hurley, Daniel C

    Matthew M. Hurley, Daniel C. Dennett, and Reginald B. Adams Jr. 2011. Inside Jokes: Using Humor to Reverse-Engineer the Mind . The MIT Press, Cambridge, MA, USA

    2011

  7. [16]

    Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson, and Mor Naaman. 2023. Co-Writing with Opinionated Language Models Affects Users’ Views. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI ’23). Association for C...

    2023

  8. [17]

    Sophie Jentzsch and Kristian Kersting. 2023. ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models. arXiv:2306.04563 [cs.AI] https://arxiv.org/abs/2306.04563

    2023 arXiv

  9. [18]

    Steve Kaplan. 2013. The Hidden Tools of Comedy: The Serious Business of Being Funny . Michael Wiese Productions, Studio City, CA. 280 pages. 18 Sean Kim and Lydia B. Chilton

    2013

  10. [19]

    Chloé Kiddon and Yuriy Brun. 2011. That’s What She Said: Double Entendre Identification. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers - Volume 2 (Portland, Oregon) (HLT ’11). Association f...

    2011

  11. [20]

    Xiao Ma, Swaroop Mishra, Ahmad Beirami, Alex Beutel, and Jilin Chen. 2023. Let’s Do a Thought Experiment: Using Counterfactuals to Improve Moral Reasoning. arXiv:2306.14308 [cs.CL] https://arxiv.org/abs/2306.14308

    2023 arXiv

  12. [21]

    Josipa Majic. 2024. Rizz App: How the 5th Most Downloaded Dating App is Redefining Digital Relationships. Forbes (2024). https://www.forbes. com/sites/josipamajic/2024/09/09/rizz-app-how-the-5th-most-downloaded-dating-app-is-redefining-digital-relationships/ Accessed: 2024-09-13

    2024

  13. [22]

    Mirowski, Juliette Love, Kory Mathewson, and Shakir Mohamed

    Piotr W. Mirowski, Juliette Love, Kory Mathewson, and Shakir Mohamed. 2024. A Robot Walks into a Bar: Can Language Models Serve as Creativity Support Tools for Comedy? An Evaluation of LLMs’ Humour Alignment with Comedians. In The 2024 ACM Conference on Fairness, Accountabilit...

    2024

  14. [23]

    Stanford Encyclopedia of Philosophy. 2015. Philosophy of Humor. http://plato.stanford.edu/entries/humor/. Accessed: 2015-12-01

    2015

  15. [24]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442 [cs.HC] https://arxiv.org/abs/2304.03442

    2023 arXiv

  16. [25]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. OpenAI Blog (2019). https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf

    2019

  17. [27]

    Victor Raskin. 2009. The Primer of Humor Research . De Gruyter, Berlin, Germany. 673 pages

    2009

  18. [28]

    Dafna Shahaf, Eric Horvitz, and Robert Mankoff. 2015. Inside Jokes: Identifying Humorous Cartoon Captions. InProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Sydney, NSW, Australia) (KDD ’15). ACM, New York, NY, USA, 1065–1074...

    2015

  19. [29]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. Advances in Neural Information Processing Systems 36 (2023). Publisher Copyright: © 2023 Neural information processing system...

    2023

  20. [30]

    Minhyang (Mia) Suh, Emily Youngblom, Michael Terry, and Carrie J Cai. 2021. AI as Social Glue: Uncovering the Roles of Deep Generative AI during Social Music Composition. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21...

    2021

  21. [31]

    Taylor and Lawrence J

    Julia M. Taylor and Lawrence J. Mazlack. 2004. Computationally recognizing wordplay in jokes. In In Proceedings of CogSci 2004

    2004

  22. [32]

    Alexey Tikhonov and Pavel Shtykovskiy. 2024. Humor Mechanics: Advancing Humor Generation with Multistep Reasoning. arXiv:2405.07280 [cs.CL] https://arxiv.org/abs/2405.07280

    2024 arXiv

  23. [33]

    Joe Toplyn. 2023. Witscript 2: A System for Generating Improvised Jokes Without Wordplay. arXiv:2302.03036 [cs.CL] https://arxiv.org/abs/2302. 03036

    2023 arXiv

  24. [34]

    Joe Toplyn. 2023. Witscript 3: A Hybrid AI System for Improvising Jokes in a Conversation. arXiv:2301.02695 [cs.CL] https://arxiv.org/abs/2301.02695

    2023 arXiv

  25. [35]

    Joe Toplyn. 2023. Witscript: A System for Generating Improvised Jokes in a Conversation. arXiv:2302.02008 [cs.CL] https://arxiv.org/abs/2302.02008

    2023 arXiv

  26. [36]

    John Vorhaus. 1994. The Comic Toolbox: How to Be Funny Even If You’re Not . Silman James Press, Los Angeles, CA. 191 pages

    1994

  27. [37]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2024. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processi...

    2024

  28. [38]

    Tongshuang Wu, Michael Terry, and Carrie Jun Cai. 2022. AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22). Asso...

    2022

  29. [39]

    Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. In Transactions of the Association for Computational Linguistics , Vol. 2. 67–78

    2014

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.