Group Selection as a Safeguard Against AI Substitution

Iyad Rahwan; Julian Garcia; Qiankun Zhong; Thomas F. Eisenmann

REVIEW 4 major objections 6 minor 5 references

Group Selection as a Safeguard Against AI Substitution

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that under individual selection AI substitutes always win, but cultural group selection with strong group boundaries can make AI complements the dominant strategy, preserving the collective variance needed for innovation.

desk verdict A useful possibility argument for group selection and AI use, but the central example contains a parameter inconsistency that undercuts the individual-versus-group tension. read the letter →

arxiv 2602.03541 v2 pith:RYCPY2IU submitted 2026-02-03 cs.AI econ.TH

Qiankun Zhong , Thomas F. Eisenmann , Julian Garcia , Iyad Rahwan This is my paper

classification cs.AIecon.TH

keywords generativeAIculturalevolutioncumulativegroupselectioncomplementvssubstituteevolutionarygametheorymodelcollapsereplicatordynamics

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 asks whether AI use can stall cumulative cultural evolution and whether that outcome is avoidable. It models two ways people use generative AI—as a substitute that produces most of the output, and as a complement that assists while the human remains the main author—and treats the choice between them as a cultural trait under evolutionary competition. The central claim is that under individual-level selection in a well-mixed population, the substitute strategy is the unique evolutionary winner because it improves individual skill fastest, even though it drains the collective variance needed for long-run innovation; but under cultural group selection with strong group boundaries, the complement strategy can spread and become dominant, because groups that preserve variance accumulate cultural improvements faster and other groups copy them. The paper therefore argues that the apparently individual-rational triumph of AI substitutes is not inevitable and can be reversed by population structure.

What carries the argument

The engine is a Gumbel extreme-value model of social learning: each learner copies the highest-skilled model in their neighborhood and draws their post-learning skill from a Gumbel distribution with mode z_j − α and dispersion β, where α is average learning error and β is outcome dispersion. AI strategies rescale these parameters—Complement reduces both moderately, Substitute reduces both more strongly—and the resulting expected skills feed a replicator equation for individual selection. Group selection is added by allowing strategy copying within groups at rate G1 and between groups at rate G2 (G1 >> G2); the group boundary slows the short-term invasion of the substitute strategy and protec

What would settle it

Give two populations the same learning task over many generations, one using AI substitutes and one using AI complements, and measure the spread of output quality and the progress of the best skill; the model predicts complement users preserve noticeably higher variance and overtake substitutes after repeated generations, so observing no variance gap or no reversal would count against it.

Watch

Extended reading notes

Core claim

Under individual selection, AI Substitute is the only strategy that survives evolutionary competition: it invades both no-AI and AI-Complement populations because its greater reduction in learning error yields higher expected skill in every generation. Under group selection with strong group boundaries—within-group learning probability above about 0.9—AI Complement can instead become the dominant strategy across the population: complement-using groups retain more learning variance, overtake substitute-using groups in cumulative skill after roughly 18 generations, and spread their strategy through between-group social learning. The result holds for larger numbers of groups and across a parame

Load-bearing premise

The whole result rests on the assumption that AI use shrinks learning error and learning variance by fixed percentages—substitutes shrink both more than complements—and that these numbers are the same for all users; if real AI use does not compress variance this way, the evolutionary winner could change.

Editorial extensions

If this is right

If individual incentives dominate, AI substitutes will spread to fixation in any well-mixed population, regardless of initial conditions.
Strengthening group boundaries—making within-group learning much more frequent than between-group learning—can let complement use take over, because variance-preserving groups out-accumulate substitutes.
The long-run benefit of complement use appears only after several generations (about 18 in the illustrative run); short-term comparisons favor substitutes.
The group-selection result scales to ten groups and holds across a range of AI error/variance reduction parameters, but only when in-group learning is strong.

Reading between the lines

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

If the model transfers to real organizations, policies that strengthen internal knowledge-sharing and limit cross-organization imitation—proprietary data, nondisclosure norms, specialized in-house models—could make AI-complement use culturally stable; the paper discusses structural pluralism but does not test it empirically.
The core assumption that AI reduces learning variance proportionally (more for substitutes) is the crux; an empirical study measuring the variance of creative output under complement vs. substitute use across repeated tasks could validate or falsify the model's foundation.
The replicator-dynamics result describes deterministic, infinite-population selection; in small populations where skill maxima fluctuate stochastically, the outcome may differ, which the paper does not analyze.
The categorical treatment of AI use could be extended to a continuous trait (degree of substitution); the paper's parameter sweep hints at thresholds but does not model gradual adoption decisions.

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.

Desk Editor's Note

read the letter

The paper is a clean possibility argument: if using AI as a substitute reduces learning variance more than using it as a complement, then group-level selection with strong boundaries can favor complement use even when substitutes pay off better for individuals. That specific application of multilevel cultural selection to AI-use strategies is new, and the model is clearly laid out as an extension of Henrich's framework. The replicator-dynamics treatment and the honest limitation section are also credits.

The soft spots are real, though. The most serious is an internal inconsistency in §4.1. With the parameters stated there—r_Cα = r_Sα = 0.2, r_Cβ = 0.4, r_Sβ = 0.5—the Gumbel expectation is z − α + γβ, so complement actually has a higher expected skill than substitute. The text says the opposite: 'AI Substitutes provide a higher average payoff than AI Complements in each generation.' That claim is load-bearing, because the individual-vs-group dilemma depends on Substitute being individually fitter. Either the parameters in the figure differ from the text, or the example is miscalibrated. Either way, the claimed conflict is not established as a structural result.

Second, the main qualitative outcomes are effectively baked in by the ordering assumptions α_S < α_C and β_S < β_C. The paper never characterizes the region where Substitute is individually fitter and Complement is group-fitter, and the assertion that different learning parameters only change speed, not structure, is unsupported by a sweep or proof. Third, no code is shipped, which makes the ABM hard to check. Finally, the authors themselves admit in §5.4 that the model lacks empirical quantitative validation. That admission is appropriate, but it means the policy implications in §5.2 should be read as conditional, not robust.

None of this kills the core idea. It is a legitimate thought experiment, and the group-selection mechanism is worth airing in the AI-and-culture literature. But the paper needs a corrected or clarified parameter regime, a proper exploration of the tradeoff space, and ideally code before the conclusions can be trusted.

For a referee: yes, send it out. The novelty and clarity justify referee time, even though revision would be substantial. For a reading group, it is a good discussion piece. I would not cite it in its current form.

Referee Report

4 major / 6 minor

Summary. The paper extends Henrich's model of cumulative cultural evolution with an agent-based model in which agents can use AI as a Complement (reduces learning error and dispersion mildly) or as a Substitute (reduces error and dispersion more strongly). Under payoff-biased social learning in a well-mixed population, the authors report that the Substitute strategy is the only evolutionarily stable strategy, despite its stronger variance reduction. In a group-structured population with strong in-group learning, they report that Complement can spread and become dominant, because groups with Complement maintain higher variance and thus accumulate cultural skill faster. The paper concludes that cultural group selection can safeguard against AI substitution by preserving variance. The central claims are that (i) individual selection favors Substitutes, (ii) group selection can favor Complements, and (iii) this leads to a long-term tradeoff captured by the model.

Significance. If the results are robust, the paper would make a timely contribution to the emerging literature on AI and cultural evolution, providing a formal evolutionary argument for why organizational or group boundaries might preserve beneficial AI-use strategies. The model itself is clearly described, extends a well-known framework (Henrich 2004), and combines agent-based simulation with replicator dynamics in a way that is reproducible in principle. The group-selection mechanism, modeled via in-group versus between-group social learning, is a plausible extension of existing cultural evolution theory. However, the significance currently hinges on parameter choices that are inconsistently reported and not justified by empirical calibration, as discussed in the major comments. With a systematic parameter analysis and consistent reporting, the paper could make a useful theoretical contribution; as presented, its conclusions are conditional and not yet fully supported.

major comments (4)

[§4.1, parameter definitions in §2] The headline group-selection example uses r(C)α = 0.2, r(S)α = 0.2, r(C)β = 0.4, r(S)β = 0.5. With the Gumbel learning model, E[z'] = z_j − α_s + γβ_s (γ ≈ 0.577). Since α_S = α_C and β_S < β_C, Complement has strictly higher expected skill than Substitute (by γ·(0.1β) ≈ 0.029 for β=0.5). The text states 'AI Substitutes provide a higher average payoff than AI Complements in each generation,' which is false under these numbers. This undermines the premise that group selection is rescuing an individually costly strategy. The example needs to use parameters where Substitute is individually fitter, or the individual-fitness region must be characterized explicitly.
[§3.2 and Figure 5] The claim that 'Different cultural learning parameters change the speed of selection, but not the structural outcome' is asserted without proof or a full sweep. The replicator dynamics are shown for one parameter set. For the Gumbel model, the expected payoff difference between Substitute and Complement is E_S − E_C = (α_C − α_S) + γ(β_S − β_C). The sign depends on the relative magnitudes of α and β reductions. For example, with equal α reductions, Complement is fitter, not Substitute. The manuscript must either provide an analytical characterization of the region where Substitute dominates or supply a systematic parameter sweep demonstrating that the individual-selection outcome is invariant across the full supported parameter space.
[Figures 4, 6, 7 and Supplementary Table 2] The parameters used for the same qualitative claims are inconsistent across exhibits. Figure 4 uses (r(C)α=0.2, r(C)β=0.05, r(S)α=0.5, r(S)β=0.5); §4.1/Figure 6 uses (r(C)α=0.2, r(S)α=0.2, r(C)β=0.4, r(S)β=0.5); Figure 7 uses (r(C)α=0.2, r(S)α=0.3, r(C)β=0.5, r(S)β=0.75); the SI table lists yet another set (AIα1=0.2, AIβ1=0.2, AIα2=0.5, AIβ2 varying). This makes it impossible to verify whether the reported outcomes are robust or are selected calibrations. The manuscript needs to state one canonical parameter set for each claim, justify it, and include sensitivity analyses showing that the qualitative conclusions do not depend on exact values.
[§2 and §5.4] The core qualitative prediction—that AI substitution reduces variance and slows cumulative cultural evolution—is built into the model by construction: §2 defines Substitute as having lower β (and often lower α) than Complement, and §5.4 admits the scaling parameters lack empirical validation. The paper should explicitly separate the assumed input (AI reduces error/dispersion, with Substitute reducing more) from the emergent evolutionary outcomes (which strategy spreads under individual vs. group selection). As written, the abstract and conclusions present the variance-reduction effect as a finding rather than as a modeling assumption. This distinction is essential for interpreting the 'safeguard' conclusion as a conditional theoretical result.

minor comments (6)

[Figure 5 caption] The caption refers to 'AI Help' but the strategy is called 'AI Complement' throughout the paper. Please unify terminology.
[Figure 6 caption and §4.1] The caption says 'same parameters except for the group structure and in-group learning rate,' but the parameters in §4.1 differ from those in Figure 4. Clarify whether 'same' refers to the two panels of Figure 6 or to the earlier figures, and state all parameter values in each caption.
[Supplementary Table 2] The parameter table is hard to parse: 'α=0.2' appears twice, and the labels AIα1, AIβ1, AIα2, AIβ2 are not defined in the main text. A clear table with columns for Complement and Substitute reductions (r(C)α, r(C)β, r(S)α, r(S)β) would remove ambiguity.
[§2, model description] The model lacks a specification of initial skill values and the exact number of learning attempts represented by the Gumbel 'best-of-several-attempts' interpretation. Please state these details in the main text or the supplementary material.
[§3.1, last sentence] The sentence 'The rate of population convergence to the full adoption of AI strategies (lower panel) also influences overall cumulative development' is vague. Specify how convergence rate affects cumulative development and which figure supports this.
[References] Some references are cited as preprints (e.g., Wan & Kalman 2025, Sourati et al. 2025) and might have been updated; please verify the final publication status and add DOIs where available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: explicit model assumptions produce simulation outcomes; no prediction reduces to a fit or self-citation.

full rationale

The derivation chain is transparent. Section 2 states the Gumbel learning model and the alpha/beta reductions for Complement and Substitute as explicit assumptions, including the orderings r(S)_alpha > r(C)_alpha and r(S)_beta > r(C)_beta. The results in Sections 3.1, 3.2, and 4.1 are simulation and replicator-dynamics consequences of these assumptions, not fitted parameters renamed as predictions. The conclusion that larger beta preserves variance and supports cumulative cultural evolution follows from the extreme-value structure (larger dispersion gives a wider upper tail above the current max), but that is a derived property of the model, not an input identical to the output. The group-selection result is an application of multilevel selection with explicit group boundaries (G1 = 0.85), and the mechanism is cited to McElreath et al. 2003, an external source. There is no load-bearing self-citation: the only overlapping-author reference (Brinkmann et al. 2023, if Rahwan is a co-author) provides background, not the argument's premise. Section 5.4 acknowledges the lack of empirical calibration, but a limitation on external validity is not circularity. One internal inconsistency exists: Section 4.1 uses r(C)_alpha = r(S)_alpha = 0.2, which makes Complement's expected skill higher (since beta_C > beta_S and the Gumbel mean is mode + gamma*beta), while the text claims 'AI Substitutes provide a higher average payoff than AI Complements in each generation.' This is a correctness/consistency issue, not a circular reduction, so it does not raise the circularity score.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The paper relies on several hand-chosen parameters, especially the AI effect ratios and the in-group learning probability. The axioms are standard modeling choices from cultural evolution, but the core AI variance-reduction assumption is not empirically calibrated. No new physical or conceptual entities are postulated.

free parameters (7)

baseline learning error α = 1 (AI-influence experiments); 0.2 (multi-level experiments)
Hand-chosen; the supplement says parameters were picked so the system is sensitive to the AI modification.
baseline learning dispersion β = 0.5
Hand-chosen; not estimated from data.
AI effect parameters rCα, rCβ, rSα, rSβ = vary across figures: (0.2, 0.05, 0.5, 0.5) in Fig 4; (0.2, 0.4, 0.2, 0.5) in Fig 6; (0.2, 0.5, 0.3, 0.75) in Fig 7
Hand-selected and inconsistent across simulations; these directly determine which strategy has short-run versus long-run advantages.
selection strength δ = 10
Steepness of the logistic adoption function; chosen by hand.
in-group learning probability G1 = 0.85 in main group-selection run; threshold above 0.9 in Fig 7c
Critical for the group-selection result; the desired outcome only appears when group boundaries are strong.
initial early-adopter fraction p = 0.1
Initial fraction of AI users; chosen by hand.
population size N and group count m = N=1000; m=3 or 10
Simulation scale and structure; not fitted to any empirical population.

assumptions (5)

domain assumption Learning outcomes are drawn from a Gumbel distribution with mode z_j − α and dispersion β (§2).
A reduced-form model of best-of-several-attempts learning; no empirical evidence that cultural learning follows this distribution.
ad hoc to paper AI Complement and Substitute reduce learning error and dispersion, with Substitute reducing both more strongly (§2).
This is the core assumption of the paper; it is qualitatively supported by cited empirical work but is not quantitatively calibrated.
domain assumption Payoff equals post-learning skill and strategy adoption is pay-off-biased social learning with Pr(i←k) = 1/(1+e^{-δ(z_k-z_i)}) (§2).
A standard evolutionary-game-theory assumption mapping skill to fitness.
domain assumption Multi-level selection is mediated by in-group vs out-group interaction probabilities G1 and G2 (§4).
The group-selection result depends on this boundary mechanism and specifically on G1 being large.
domain assumption Henrich's model of cumulative cultural evolution correctly describes how variance enables cumulative gains (§1, §2).
The authors themselves note in Limitations (§5.4) that the base model is 'well-cited, but often used as a reference to cumulative cultural evolution rather than validated directly.'

how reviews work

0 comments

Cite this review

Pith. "Pith review of Group Selection as a Safeguard Against AI Substitution." pith.science (2026). https://pith.science/paper/RYCPY2IU

@misc{pith2026260203541,
  author       = {Pith},
  title        = {Pith review of: Group Selection as a Safeguard Against AI Substitution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYCPY2IU}},
  note         = {Machine review of arXiv:2602.03541}
}

read the original abstract

Reliance on generative AI can reduce cultural variance and diversity, especially in creative work. This reduction in variance has already led to problems in model performance, including model collapse and hallucination. In this paper, we examine the long-term consequences of AI use for human cultural evolution and the conditions under which widespread AI use may lead to "cultural collapse", a process in which reliance on AI-generated content reduces human variation and innovation and slows cumulative cultural evolution. Using an agent-based model and evolutionary game theory, we compare two types of AI use: complement and substitute. AI-complement users seek suggestions and guidance while remaining the main producers of the final output, whereas AI-substitute users provide minimal input, and rely on AI to produce most of the output. We then study how these use strategies compete and spread under evolutionary dynamics. We find that AI-substitute users prevail under individual-level selection despite the stronger reduction in cultural variance. By contrast, AI-complement users can benefit their groups by maintaining the variance needed for exploration, and can therefore be favored under cultural group selection when group boundaries are strong. Overall, our findings shed light on the long-term, population-level effects of AI adoption and inform policy and organizational strategies to mitigate these risks.

Figures

Figures reproduced from arXiv: 2602.03541 by the authors.

**Figure 1.** Comparing AI Complement and Substitute in writing tasks. Here we illustrate how people can use GenAI as a substitute or a complement in creative work. For an essay writing task on a new topic, users could use GenAI as a complement to assist their writing or ask for suggestions on the given topic, ultimately still writing the essay themselves based on the materials and structure provided by LLMs. Users could also use… view at source ↗

**Figure 2.** Effects of AI on Social Learning Outcome. Building on empirical work, we assume that AI affects the mean and variance of social learning outcomes. Greater reliance on AI leads to higher average social learning outcomes (lower α) and reduced dispersion in outcomes (lower β). The right panel demonstrates how AI Complements (blue) and AI Substitutes (green) change the initial distribution of social learning outcomes … view at source ↗

**Figure 3.** Model Iteration. The model runs in three steps iteratively: 1. Searching and learning; 2. Use of AI. 3. Selection on AI strategy. In well-mixed population, Step 3 selects for of AI strategies based on fitness. With group structure, we assume individuals interact more frequently with in-group than out-group members 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Cumulative outcome and adoption rate of AI Complement and AI Substitute. We run the two different conditions (with 100 repetitions) to contrast a population starting with 10% of AI Complement early adopters and one starting with 10% AI Substitute early adopters. The li…

**Figure 5.** Figure 5: Selection gradient on the simplex under replicator dynamics. We plot the replicator vector field on the simplex of strategy frequencies. Payoffs πs(x) are estimated from repeated simulations of the searching-and-learning process in a well-mixed population and substitu…

**Figure 6.** Figure 6: AI Complement is selected for through group-level selection but not in individual-level selection. As a group-beneficial but individually costly strategy, using AI as a Complement can be selected for in group-level selection. We run two simulations with the same param…

**Figure 7.** Figure 7: Conditions for Group Selection. The differential payoffs enable group-level selection for the group-beneficial AI strategy. Panel (a) shows the median skill level of the population in a single run, varying by the level of AI influence on learning accuracy Dα and learn…

**Figure 1.** Figure 1: Effects of different AI strategies on Cumulative Cultural Evolution. We compare the results of a single run for the AI Substitute condition and a single run for the AI Complement condition. The box plot zooms into one repetition to show in detail the skill distributio…

**Figure 2.** Figure 2: Multi-level selection of AI Complement across 10 groups We generalize the group-selection mechanism to 10 groups and show that AI Complement can spread across more groups and become the dominant strategy. 2 [PITH_FULL_IMAGE:figures/full_fig_p025_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 2 linked inside Pith

[1]

R., Shah, J

Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homoge- nization eﬀects of large language models on human creative ideation. Proceedings of the 16th conference on creativity & cognition, 413–425. Aoki, K. (2018). On the absence of a correlation between popula- tion size and ‘toolkit size’in ethnographic hunter–gatherers. Philosophical Transactions o...

arXiv 2024
[30]

Wan, Y., & Kalman, Y. M. (2025). Using generative ai personas in- creases collective diversity in human ideation.arXiv preprint arXiv:2504.13868. Winters, J. (2019). Escaping optimization traps: The role of cultural adaptation and cultural exaptation in facilitating open-ended cumulative dynamics. Palgrave Communications, 5(1). 23 Supplementrary Materials...

arXiv 2025
[623]

https://doi.org/10.1145/3442188.3445922 Blind, K. (2012). The inﬂuence of regulations on innovation: A quan- titative assessment for oecd countries.Research policy, 41(2), 391–400. Brinkmann, L., Baumann, F., Bonnefon, J.-F., Derex, M., M¨ uller, T. F., Nussberger, A.-M., Czaplicka, A., Acerbi, A., Grif- ﬁths, T. L., Henrich, J., et al. (2023). Machine cu...

arXiv 2012
[2407]

Evans, J., & Rzhetsky, A. (2010). Machine science.Science, 329(5990), 399–400. Fan, D., Messmer, B., Doikov, N., & Jaggi, M. (2024). On-device collaborative language modeling via a mixture of generalists and specialists. arXiv preprint arXiv:2409.13931 . Ganguli, D., Askell, A., Bai, Y., & et al. (2022). Predictability and surprise in large generative mod...

arXiv 2010
[3440]

King, A. (2024). Your favorite lo-ﬁ chill music station is probably ai. Digital Music News . Retrieved June 12, 2025, from https: //www.digitalmusicnews.com/2024/11/26/is-this-youtube- channel-lo-ﬁ-music-created-with-ai/ Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Gold - stein, T. (2023). A watermark for large language models. In- terna...

arXiv 2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

R., Shah, J

Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homoge- nization eﬀects of large language models on human creative ideation. Proceedings of the 16th conference on creativity & cognition, 413–425. Aoki, K. (2018). On the absence of a correlation between popula- tion size and ‘toolkit size’in ethnographic hunter–gatherers. Philosophical Transactions o...

arXiv 2024

[2] [30]

Wan, Y., & Kalman, Y. M. (2025). Using generative ai personas in- creases collective diversity in human ideation.arXiv preprint arXiv:2504.13868. Winters, J. (2019). Escaping optimization traps: The role of cultural adaptation and cultural exaptation in facilitating open-ended cumulative dynamics. Palgrave Communications, 5(1). 23 Supplementrary Materials...

arXiv 2025

[3] [623]

https://doi.org/10.1145/3442188.3445922 Blind, K. (2012). The inﬂuence of regulations on innovation: A quan- titative assessment for oecd countries.Research policy, 41(2), 391–400. Brinkmann, L., Baumann, F., Bonnefon, J.-F., Derex, M., M¨ uller, T. F., Nussberger, A.-M., Czaplicka, A., Acerbi, A., Grif- ﬁths, T. L., Henrich, J., et al. (2023). Machine cu...

arXiv 2012

[4] [2407]

Evans, J., & Rzhetsky, A. (2010). Machine science.Science, 329(5990), 399–400. Fan, D., Messmer, B., Doikov, N., & Jaggi, M. (2024). On-device collaborative language modeling via a mixture of generalists and specialists. arXiv preprint arXiv:2409.13931 . Ganguli, D., Askell, A., Bai, Y., & et al. (2022). Predictability and surprise in large generative mod...

arXiv 2010

[5] [3440]

King, A. (2024). Your favorite lo-ﬁ chill music station is probably ai. Digital Music News . Retrieved June 12, 2025, from https: //www.digitalmusicnews.com/2024/11/26/is-this-youtube- channel-lo-ﬁ-music-created-with-ai/ Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Gold - stein, T. (2023). A watermark for large language models. In- terna...

arXiv 2024