arxiv: 2604.21853 · v1 · submitted 2026-04-23 · ⚛️ physics.soc-ph

Recognition: unknown

Generative artificial intelligence reduces social welfare through model collapse

Erol Ak\c{c}ay, Fabian Baumann, Joshua B. Plotkin

Authors on Pith no claims yet

Pith reviewed 2026-05-08 13:22 UTC · model grok-4.3

classification ⚛️ physics.soc-ph

keywords generative AImodel collapsesocial welfaresocial dilemmahabit formationcollective action problem

0 comments

The pith

Generative AI reduces social welfare for important tasks through model collapse and habit spillover.

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

This paper builds a simple model of how people decide whether to do work themselves, use generative AI, or skip tasks when collaborating with others. It shows that while relying on AI can help an individual complete a task faster now, training new AI systems on those AI-made outputs causes the models to produce less varied and less accurate results over time. The welfare effects depend on how much people value doing the task without AI and how strongly model collapse affects performance. For tasks that are most valuable to society, the long-term losses from poorer AI outputs outweigh the short-term individual benefits. Additionally, once people get used to using AI on easy tasks, they start using it on hard ones too, spreading the damage.

Core claim

The introduction of generative AI, while initially beneficial at the individual level, will reduce social welfare for the most important types of tasks. Habit formation around genAI use can couple otherwise separate domains, so that adoption in low-stakes tasks spills over into high-value tasks and amplifies welfare losses.

What carries the argument

A parsimonious model of behavior choices in collaborative interactions, organized by a two-dimensional taxonomy of the strength of the incentive to perform the task without AI and the severity of model collapse.

If this is right

For the most important tasks, introducing generative AI reduces overall social welfare.
Habit formation links use across task types, causing losses to spread from minor to major domains.
Individually rational adoption leads to collective welfare reduction without intervention.
The welfare impact is determined by task-specific incentives and collapse severity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This suggests a need for policies that encourage human effort or restrict AI data in training to avoid the welfare decline.
The effect could be measured in real-world settings like creative industries or scientific collaboration where output quality can be tracked over time.
Similar dynamics might appear in other technologies that offer individual shortcuts but degrade shared resources.

Load-bearing premise

Delegating tasks to generative AI is individually beneficial in the short term, but widespread use degrades model performance enough to outweigh those benefits for important tasks, and habit formation reliably connects low-stakes and high-value domains.

What would settle it

Tracking the quality and diversity of outputs in a collaborative task environment before and after high rates of generative AI adoption, particularly for tasks requiring high accuracy and originality, to see if performance declines as predicted.

read the original abstract

Generative artificial intelligence (genAI) is rapidly reshaping how knowledge and culture are produced and consumed. Yet generative models are vulnerable to model collapse: when trained on data generated by earlier versions of themselves, their outputs can lose diversity and accuracy. This creates a social dilemma, because delegating tasks to genAI can be individually beneficial in the short term even as widespread adoption degrades future model performance. Here we develop a parsimonious model of behavior in collaborative interactions in which individuals can either exert human effort, rely on genAI, or refrain from work altogether. The welfare consequences of genAI are organized by a simple two-dimensional taxonomy: the strength of the incentive to perform the task without AI, and the severity of model collapse. Within this framework, the introduction of genAI -- while initially beneficial at the individual level -- will reduce social welfare for the most important types of tasks. In addition, habit formation around genAI use can couple otherwise separate domains, so that adoption in low-stakes tasks spills over into high-value tasks and amplifies welfare losses. Together, these results identify a general pathway by which, in the absence of intervention, individually rational adoption of genAI will assuredly and profoundly reduce collective welfare.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper models how rational genAI adoption reduces welfare for important tasks through model collapse, with habit formation as the key link between task types, though the spillover strength is a potential weak point.

read the letter

The paper's core point is that without intervention, people will rationally choose genAI in ways that degrade models over time and cut collective welfare, especially on tasks that matter most, because habit formation carries the behavior over from low-stakes work to high-value work. The new element is the two-dimensional taxonomy that sorts outcomes by incentive strength to do the task without AI and by collapse severity, plus the addition of habit formation to connect otherwise separate domains. This extends earlier model collapse work into a behavioral setup with three choices per person: human effort, genAI, or opting out. It does a solid job of framing the social dilemma in clear terms and showing how individual short-term gains can conflict with long-term group losses. The taxonomy keeps the analysis organized without unnecessary complexity. The soft spots center on the habit formation mechanism. The abstract presents it as a coupling feature that amplifies losses, but it is not clear from the description how this spillover is derived from the individual choice rules or what its strength and functional form are. If the link turns out weak or absent, adoption could stay confined to low-incentive tasks and the welfare reduction for important work would not follow as directly. The two free parameters for incentive strength and collapse severity are explicit, yet the model would be stronger with explicit checks on how sensitive the main result is to habit strength. This paper is for social scientists and policy readers who want a structured way to think about AI adoption effects. It deserves a serious referee because it formalizes a real concern with a usable framework, even if the derivations and robustness details need closer scrutiny in review. I would send it out for peer review with requests for the full equations and sensitivity analysis on the spillover.

Referee Report

2 major / 2 minor

Summary. The manuscript develops a parsimonious model of individual choices in collaborative tasks, where agents select human effort, genAI delegation, or abstention. Welfare outcomes are organized via a two-dimensional taxonomy parameterized by incentive strength for non-AI performance and model collapse severity. The central claims are that genAI adoption is individually beneficial in the short term yet reduces social welfare for high-incentive (important) tasks, and that habit formation produces spillover adoption from low-stakes to high-value domains, yielding assured and profound collective welfare losses absent intervention.

Significance. If the derivations hold, the work supplies a compact organizing framework for the social dilemma created by genAI and model collapse, clarifying conditions under which individually rational behavior produces collective harm. The taxonomy and explicit treatment of spillover identify a general mechanism that could inform policy on AI governance and data provenance.

major comments (2)

[Habit formation subsection] Habit formation subsection: the spillover mechanism that couples low-stakes and high-value domains is introduced as a qualitative feature rather than derived from the individual choice rules or the two-parameter taxonomy; without a functional form or threshold for the coupling strength, the extension of welfare losses to important tasks does not follow from the model alone.
[Welfare consequences section] Welfare consequences section: the assertion that adoption 'assuredly and profoundly' reduces welfare for the most important tasks requires that long-term collective losses from collapse outweigh short-term individual gains specifically in the high-incentive regime, yet the manuscript provides no explicit comparison or threshold condition showing this inequality holds across the taxonomy.

minor comments (2)

[Abstract] The abstract states the three behavioral options clearly but does not preview the exact parameterization of the taxonomy.
[Model section] Notation for incentive strength and collapse severity should be introduced with explicit symbols or equations in the model section to facilitate reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and insightful report. The comments correctly identify points where the derivations can be made more explicit. We address each major comment below and will revise the manuscript to incorporate the suggested clarifications.

read point-by-point responses

Referee: [Habit formation subsection] Habit formation subsection: the spillover mechanism that couples low-stakes and high-value domains is introduced as a qualitative feature rather than derived from the individual choice rules or the two-parameter taxonomy; without a functional form or threshold for the coupling strength, the extension of welfare losses to important tasks does not follow from the model alone.

Authors: We agree that the spillover is currently presented as a qualitative extension rather than derived from the core choice rules. The primary welfare results in the high-incentive regime follow directly from the individual optimization and taxonomy without requiring spillovers. To strengthen the habit-formation claim, we will revise the subsection to introduce a minimal functional form (e.g., a spillover probability linear in the fraction of low-stakes tasks delegated to genAI) and derive the resulting threshold condition on coupling strength under which welfare losses propagate to high-value domains. revision: yes
Referee: [Welfare consequences section] Welfare consequences section: the assertion that adoption 'assuredly and profoundly' reduces welfare for the most important tasks requires that long-term collective losses from collapse outweigh short-term individual gains specifically in the high-incentive regime, yet the manuscript provides no explicit comparison or threshold condition showing this inequality holds across the taxonomy.

Authors: We acknowledge that the manuscript does not supply an explicit inequality or threshold comparing long-term collective losses to short-term individual gains within the high-incentive regime. The claim rests on the equilibrium welfare expressions under the two-parameter taxonomy, but an explicit side-by-side comparison is absent. We will revise the welfare consequences section to add this comparison, deriving the boundary in (incentive strength, collapse severity) space where collective welfare with genAI falls below the no-genAI baseline. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation is self-contained from individual choice rules and taxonomy

full rationale

The paper constructs a new parsimonious model of individual choices (exert effort, use genAI, or refrain) whose welfare outcomes are organized by a two-dimensional taxonomy of incentive strength and collapse severity. The central result that genAI reduces welfare for important tasks follows directly from comparing short-term individual benefits against long-term collective degradation within that taxonomy; habit formation is introduced as an additional qualitative coupling mechanism rather than a fitted parameter or renamed input. No equations or claims reduce by construction to prior self-citations, fitted subsets of data, or self-definitional loops. The derivation therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The paper rests on standard assumptions from game theory and AI degradation literature rather than new fitted constants or invented entities; the two taxonomy dimensions organize outcomes but are not presented as fitted parameters.

free parameters (2)

incentive strength to perform task without AI
One axis of the two-dimensional taxonomy that organizes welfare consequences; treated as a variable parameter rather than a fitted constant.
severity of model collapse
Second axis of the taxonomy; determines how much performance degrades with AI use.

axioms (2)

domain assumption Individuals choose between human effort, genAI assistance, or refraining based on short-term individual payoffs in collaborative settings.
Core behavioral premise of the parsimonious model described in the abstract.
domain assumption Widespread genAI use produces model collapse that reduces output diversity and accuracy.
Premise drawn from existing genAI literature and invoked to link individual choices to collective outcomes.

pith-pipeline@v0.9.0 · 5518 in / 1430 out tokens · 59654 ms · 2026-05-08T13:22:00.694755+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

64 extracted references · 19 canonical work pages · 2 internal anchors

[1]

S. Noy, W. Zhang, Experimental evidence on the productivity effects of generative artificial intelligence.Science381(6654), 187–192 (2023)

2023
[2]

Assael,et al., Contextualizing ancient texts with generative neural networks.Nature 645(8079), 141–147 (2025)

Y. Assael,et al., Contextualizing ancient texts with generative neural networks.Nature 645(8079), 141–147 (2025)

2025
[3]

Daniotti, J

S. Daniotti, J. Wachs, X. Feng, F. Neffke, Who is using AI to code? Global diffusion and impact of generative AI.Science0(0), eadz9311, doi:10.1126/science.adz9311

work page doi:10.1126/science.adz9311
[4]

Zhang,et al., Deep generative models design mRNA sequences with enhanced translational capacity and stability.Science390(6773), eadr8470 (2025)

H. Zhang,et al., Deep generative models design mRNA sequences with enhanced translational capacity and stability.Science390(6773), eadr8470 (2025)

2025
[5]

V. M. Rao,et al., Multimodal generative AI for medical image interpretation.Nature639(8056), 888–896 (2025)

2025
[6]

Swanson,et al., Generative AI for designing and validating easily synthesizable and struc- turally novel antibiotics.Nature Machine Intelligence6(3), 338–353 (2024)

K. Swanson,et al., Generative AI for designing and validating easily synthesizable and struc- turally novel antibiotics.Nature Machine Intelligence6(3), 338–353 (2024)

2024
[7]

Q. Hao, F. Xu, Y. Li, J. Evans, Artificial intelligence tools expand scientists’ impact but contract science’s focus.Nature649(8099), 1237–1243 (2026), doi:10.1038/s41586-025-09922-y

work page doi:10.1038/s41586-025-09922-y 2026
[8]

J. Gao, D. Wang, Quantifying the use and potential benefits of artificial intelligence in scientific research.Nature Human Behaviour8(12), 2281–2292 (2024)

2024
[9]

Epstein,et al., Art and the science of generative AI.Science380(6650), 1110–1111 (2023)

Z. Epstein,et al., Art and the science of generative AI.Science380(6650), 1110–1111 (2023)

2023
[10]

Brinkmann,et al., Machine culture.Nature Human Behaviour7(11), 1855–1868 (2023)

L. Brinkmann,et al., Machine culture.Nature Human Behaviour7(11), 1855–1868 (2023)

2023
[11]

Clark, Extending minds with generative AI.Nature Communications16(1), 4627 (2025)

A. Clark, Extending minds with generative AI.Nature Communications16(1), 4627 (2025)

2025
[12]

Farrell, A

H. Farrell, A. Gopnik, C. Shalizi, J. Evans, Large AI models are cultural and social technologies. Science387(6739), 1153–1156 (2025)

2025
[13]

Wang,et al., Scientific discovery in the age of artificial intelligence.Nature620(7972), 47–60 (2023)

H. Wang,et al., Scientific discovery in the age of artificial intelligence.Nature620(7972), 47–60 (2023)

2023
[14]

A. Bick, A. Blandin, D. J. Deming, The Rapid Adoption of Generative AI.Management Science 0(0) (2026), doi:10.1287/mnsc.2025.02523

work page doi:10.1287/mnsc.2025.02523 2026
[15]

Anthropic economic index report: Uneven geographic and enterprise ai adoption.arXivpreprintarXiv:2511.15080, 2025

R. Appel,et al., Anthropic economic index report: Uneven geographic and enterprise ai adoption.arXiv:2511.15080(2025)

work page arXiv 2025
[16]

Brooks, S

C. Brooks, S. Eggert, D. Peskoff, The rise of ai-generated content in wikipedia, inProceedings of the first workshop on advancing natural language processing for Wikipedia(2024), pp. 67–79

2024
[17]

AI use in American newspapers is widespread, uneven, and rarely disclosed

J. Russell,et al., AI use in American newspapers is widespread, uneven, and rarely disclosed. arXiv:2510.18774(2025). 12

work page internal anchor Pith review Pith/arXiv arXiv 2025
[18]

Liang,et al., Quantifying large language model usage in scientific papers.Nature Human Behaviourpp

W. Liang,et al., Quantifying large language model usage in scientific papers.Nature Human Behaviourpp. 1–11 (2025)

2025
[19]

K. Z. Cui,et al., The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.Management Science(2026)

2026
[20]

Measuring the impact of early-2025 AI on experienced open-source developer productivity.CoRR, abs/2507.09089, 2025

J. Becker, N. Rush, E. Barnes, D. Rein, Measuring the impact of early-2025 ai on experienced open-source developer productivity.arXiv:2507.09089(2025)

work page arXiv 2025
[21]

S. Peng, E. Kalliamvakou, P. Cihon, M. Demirer, The impact of ai on developer productivity: Evidence from github copilot.arXiv:2302.06590(2023)

work page internal anchor Pith review arXiv 2023
[22]

F. Dell’ Acqua,et al., Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality.Harvard Business School Technology & Operations Mgt. Unit Working Paper(24-013) (2023)

2023
[23]

Lin,et al., Persuading voters using human–artificial intelligence dialogues.Naturepp

H. Lin,et al., Persuading voters using human–artificial intelligence dialogues.Naturepp. 1–8 (2025)

2025
[24]

Messeri, M

L. Messeri, M. J. Crockett, Artificial intelligence and illusions of understanding in scientific research.Nature627(8002), 49–58 (2024)

2024
[25]

T. H. Costello, G. Pennycook, D. G. Rand, Durably reducing conspiracy beliefs through dialogues with AI.Science385(6714), eadq1814 (2024)

2024
[26]

Cheng,et al., Sycophantic AI decreases prosocial intentions and promotes dependence

M. Cheng,et al., Sycophantic AI decreases prosocial intentions and promotes dependence. Science391(6792), eaec8352 (2026)

2026
[27]

M. H. Tessler,et al., AI can help humans find common ground in democratic deliberation. Science386(6719), eadq2852 (2024)

2024
[28]

Baum,et al., From fear to action: AI governance and opportunities for all.Frontiers in Computer Science5, 1210421 (2023)

K. Baum,et al., From fear to action: AI governance and opportunities for all.Frontiers in Computer Science5, 1210421 (2023)

2023
[29]

Peter, K

S. Peter, K. Riemer, J. D. West, The benefits and dangers of anthropomorphic conversational agents.Proceedings of the National Academy of Sciences122(22), e2415898122 (2025)

2025
[30]

Binz,et al., How should the advancement of large language models affect the practice of science?Proceedings of the National Academy of Sciences122(5), e2401227121 (2025)

M. Binz,et al., How should the advancement of large language models affect the practice of science?Proceedings of the National Academy of Sciences122(5), e2401227121 (2025)

2025
[31]

K ¨obis,et al., Delegation to artificial intelligence can increase dishonest behaviour.Nature pp

N. K ¨obis,et al., Delegation to artificial intelligence can increase dishonest behaviour.Nature pp. 1–9 (2025)

2025
[32]

A. R. Doshi, O. P. Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content.Science Advances10(28), eadn5290 (2024), doi:10.1126/ sciadv.adn5290

2024
[33]

P. R. Kalluri,et al., Computer-vision research powers surveillance technology.Nature 643(8070), 73–79 (2025). 13

2025
[34]

Haas,et al., A roadmap for evaluating moral competence in large language models.Nature 650(8102), 565–573 (2026)

J. Haas,et al., A roadmap for evaluating moral competence in large language models.Nature 650(8102), 565–573 (2026)

2026
[35]

Alessandra Toniato, Alain C

I. Shumailov,et al., AI models collapse when trained on recursively generated data.Nature 631(8022), 755–759 (2024), doi:10.1038/s41586-024-07566-y

work page doi:10.1038/s41586-024-07566-y 2024
[36]

M. E. A. Seddik, S.-W. Chen, S. Hayou, P. Youssef, M. Debbah, How bad is training on synthetic data? a statistical analysis of language model collapse.arXiv:2404.05090(2024)

work page arXiv 2024
[37]

Dohmatob, Y

E. Dohmatob, Y. Feng, P. Yang, F. Charton, J. Kempe, A tale of tails: Model collapse as a change of scaling laws.arXiv:2402.07043(2024)

work page arXiv 2024
[38]

R. Boyd, P. J. Richerson,Culture and the evolutionary process(University of Chicago press) (1988)

1988
[39]

P. D. Taylor, L. B. Jonker, Evolutionary stable strategies and game dynamics.Mathematical biosciences40(1-2), 145–156 (1978)

1978
[40]

Materials and methods are available as supplementary material
[41]

X. Chen, S. Li, S. Liu, R. Fowler, X. Wang, MeetScript: Designing Transcript-based Interactions to Support Active Participation in Group Video Meetings.Proceedings of the ACM on Human- Computer Interaction7(CSCW2), 1–32 (2023), doi:10.1145/3610196

work page doi:10.1145/3610196 2023
[42]

S. S. Dvivedi, V. Vijay, S. L. R. Pujari, S. Lodh, D. Kumar, A comparative analysis of large language models for code documentation generation, inProceedings of the 1st ACM international conference on AI-powered software(2024), pp. 65–73

2024
[43]

G ¨ond¨ocs, V

D. G ¨ond¨ocs, V. D¨orfler, AI in medical diagnosis: AI prediction & human judgment.Artificial Intelligence in Medicine149, 102769 (2024), doi:10.1016/j.artmed.2024.102769

work page doi:10.1016/j.artmed.2024.102769 2024
[44]

L. G. McCoy, A. K. Manrai, A. Rodman, Large language models and the degradation of the medical record.The New England journal of medicine391(17), 1561–1564 (2024)

2024
[45]

K ¨obis, L

N. K ¨obis, L. D. Mossink, Artificial intelligence versus Maya Angelou: Experimental evidence that people cannot differentiate AI-generated from human-written poetry.Computers in Human Behavior114, 106553 (2021), doi:10.1016/j.chb.2020.106553

work page doi:10.1016/j.chb.2020.106553 2021
[46]

Porter, E

B. Porter, E. Machery, AI-generated poetry is indistinguishable from human-written po- etry and is rated more favorably.Scientific Reports14(1), 26133 (2024), doi:10.1038/ s41598-024-76900-1

2024
[47]

Hitsuwari, Y

J. Hitsuwari, Y. Ueda, W. Yun, M. Nomura, Does human–AI collaboration lead to more creative art? Aesthetic evaluation of human-made and AI-generated haiku poetry.Computers in Human Behavior139, 107502 (2023), doi:10.1016/j.chb.2022.107502

work page doi:10.1016/j.chb.2022.107502 2023
[48]

A. Yuan, A. Coenen, E. Reif, D. Ippolito, Wordcraft: story writing with large language models, inProceedings of the 27th International Conference on Intelligent User Interfaces(2022), pp. 841–852. 14

2022
[49]

M. A. M. Molla, M. M. Ahsan, Artificial intelligence and journalism: A systematic bibliometric and thematic analysis of global research.Computers in Human Behavior Reportsp. 100830 (2025)

2025
[50]

K. A. Appiah, I’m a Screenwriter. Is It All Right if I Use A.I.?The New York Times Mag- azine(2025), the Ethicist column,https://www.nytimes.com/2025/10/04/magazine/ magazine-email/screenwriter-ai-ethics.html

2025
[51]

J. Wind, M. Pandya, D. Yao, Creativity in the Age of AI (2025),https://knowledge. wharton.upenn.edu/article/creativity-in-the-age-of-ai/, knowledge at Whar- ton

2025
[52]

Vaccaro, A

M. Vaccaro, A. Almaatouq, T. Malone, When combinations of humans and AI are useful: A systematic review and meta-analysis.Nature Human Behaviour8(12), 2293–2303 (2024)

2024
[53]

Reiter, M

L. Reiter, M. J ¨orling, C. Fuchs, Working group ‘Artificial Intelligence in Higher Education’, R. B¨ohm, Student (Mis) Use of Generative AI Tools for University-Related Tasks.International Journal of Human–Computer Interactionpp. 1–14 (2025)

2025
[54]

Biswas, J

M. Biswas, J. Murray, The impact of education level on AI reliance, Habit formation, and usage, in2024 29th International Conference on Automation and Computing (ICAC)(IEEE) (2024), pp. 1–6

2024
[55]

C. Zhai, S. Wibowo, L. D. Li, The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: a systematic review.Smart Learning Environments11(1), 28 (2024)

2024
[56]

Kleinberg, M

J. Kleinberg, M. Raghavan, Algorithmic monoculture and social welfare.Proceedings of the National Academy of Sciences118(22), e2018340118 (2021)

2021
[57]

Braess, ¨Uber ein Paradoxon aus der Verkehrsplanung.Unternehmensforschung12(1), 258–268 (1968)

D. Braess, ¨Uber ein Paradoxon aus der Verkehrsplanung.Unternehmensforschung12(1), 258–268 (1968)

1968
[58]

Taitler, O

B. Taitler, O. Ben-Porat, Braess’s paradox of generative ai, inProceedings of the AAAI Con- ference on Artificial Intelligence, vol. 39 (2025), pp. 14139–14147

2025
[59]

Taitler, O

B. Taitler, O. Ben-Porat, Selective response strategies for genai.arXiv:2502.00729(2025)

work page arXiv 2025
[60]

Taitler, O

B. Taitler, O. Ben-Porat, Collaborating with GenAI: Incentives and Replacements. arXiv:2508.20213(2025)

work page arXiv 2025
[61]

Acemoglu, D

D. Acemoglu, D. Kong, A. Ozdaglar,AI, Human Cognition and Knowledge Collapse, Tech. rep., National Bureau of Economic Research (2026)

2026
[62]

Y. Guo, G. Shang, M. Vazirgiannis, C. Clavel, The curious decline of linguistic diversity: Training language models on synthetic text.arXiv:2311.09807(2023)

work page arXiv 2023
[63]

Dohmatob, Y

E. Dohmatob, Y. Feng, A. Subramonian, J. Kempe, Strong model collapse.arXiv:2410.04840 (2024)

work page arXiv 2024
[64]

Empirical evidence of Large Language Model's influence on human spoken communication , journal =

H. Yakura,et al., Empirical evidence of Large Language Model’s influence on human spoken communication.arXiv:2409.01754(2024). 15 Supplementary Materials for Generative artificial intelligence reduces social welfare through model collapse Fabian Baumann, Erol Akc ¸ay, Joshua B. Plotkin University of Pennsylvania This supplement provides the formal derivat...

work page arXiv 2024