Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Revisiting Rogers' Paradox in the Context of Human-AI Interaction

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Mirror-AI offers no long-term boost to collective knowledge.

desk verdict The headline result is the classic Rogers paradox wearing an AI costume — the AI is defined as a mirror of the population mean, so no-improvement follows by construction — but the extensions on critical learning, update schedules, and deskilling are genuinely new and worth a referee's time. read the letter →

arxiv 2501.10476 v1 pith:GRNOV2HJ submitted 2025-01-16 cs.AI cs.LG

Katherine M. Collins , Umang Bhatt , Ilia Sucholutsky This is my paper · ORCID
classification cs.AIcs.LG
keywords Rogers'Paradoxsociallearninghuman-AIinteractioncollectiveworldmodelagent-basedsimulationculturalevolutioncriticalnegativefeedbackloops
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

The paper extends the classic social-learning puzzle known as Rogers' Paradox to networks in which humans can learn from an AI system, and the AI system in turn learns from the human population. It claims that when the AI simply adopts the population's average level of adaptation each step, adding this cheap AI leaves the long-run equilibrium collective world model unchanged relative to a population of individual learners only. The paper then shows that some stakeholder strategies, notably critical social learning where a human overrides the AI output and falls back on individual learning, can raise the equilibrium. It also shows that a negative feedback loop, where learning from the AI degrades one's own ability to learn individually, can depress collective understanding unless humans can choose a more expensive human social source instead.

What carries the argument

The machinery is an extended Rogers-style agent-based network in which each of 1000 agents can learn individually, copy a random human, or copy an abstract AI. The load-bearing object is the AI's adaptation rule: $p_{\mathrm{AI}}^{\mathrm{OK}} := q^{\mathrm{OK}}$, meaning the AI's probability of being adapted at the next step equals the population's mean adaptation from the previous step. This makes the AI a perfect, immediate aggregator of the population's knowledge, i.e., pure social learning from everyone at once. A second mechanism is the negative-feedback parameter $\kappa$: agents who learn from the AI have their individual-learning success multiplied by $\kappa$, starting at 1 and decaying by a factor of 0.9 each time they use the AI, which models deskilling.

What would settle it

Run the same 1000-agent network but give the AI a training lag (its adaptation level reflecting the population several timesteps earlier), aggregate a non-representative subset of the population instead of the mean, or let it explore with moderate success; if any of these changes pushes the equilibrium population world understanding above the individual-learning-only baseline of about 0.58, the central claim that a population-mirroring AI cannot help would be shown to depend on the perfect-aggregation idealization. The paper's own Figure 6 provides a partial check: an AI with high individual-learning success clearly raises collective understanding.

Watch

Extended reading notes

Core claim

The central claim is an 'AI Rogers' Paradox': the widespread availability of AI systems that are cheap, reliable, and trained on all human data may not, on their own, improve the population's collective world model in the long run. In the paper's simulation, a zero-cost AI node whose adaptation level is set to the mean adaptation of the population at the previous timestep behaves like another social-learning channel; the equilibrium population fitness stays at the same value as when only individual learning is available. The finding holds even when the AI is made cheaper relative to human social learning or more reliable in transmitting the population's behavior. The paper also establishes that the paradox can be broken by giving the AI its own individual learning (exploration) when that learning is successful, and that the collective outcome depends on how humans and model builders choose to interact with and update the AI.

Load-bearing premise

The whole no-benefit result rests on the AI being a perfect and immediate aggregator of the population's current knowledge, with its adaptation level each step set to the population mean of the previous step; if real AI is stale, biased, or capable of independent exploration, the conclusion does not follow.

Editorial extensions

If this is right

  • An AI that merely mirrors the population will not dissolve Rogers' Paradox; long-run collective world understanding stays at the individual-learning-only equilibrium.
  • Critical social learning, overriding the AI's output and switching to individual learning when social learning fails, raises collective world understanding and also reduces reliance on the AI.
  • An AI that performs its own (individual) exploration can substantially improve collective understanding when its success rate is high and cost low, but can harm collective understanding when it frequently learns badly.
  • Too-infrequent AI updates degrade the equilibrium, with stronger effects in a faster-changing world; beyond a saturation point, more frequent updates add little.
  • When learning from the AI weakens the learner's own individual-learning ability, the collective equilibrium can drop below baseline unless humans retain a more expensive human social alternative.

Reading between the lines

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

  • An implied testable extension: if the AI aggregates a biased subset of the population rather than the true mean, the equilibrium should shift; the paper's framing suggests equity of data inclusion is not just a fairness issue but a determinant of collective learning outcomes.
  • The negative-feedback result points to a testable empirical prediction: populations that lean heavily on AI assistance for routine problem-solving may show a measurable decline in unaided problem-solving skill over time, and this decline could feed back into the quality of the data the AI learns from.
  • The update-saturation result implies a practical cost-saving guideline for model builders: once the environment-change rate is known, there is an update frequency beyond which additional updates do not improve collective outcomes, so resources can be saved without harming population-level understanding.
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 / 4 minor

Summary. The paper extends Alan Rogers' classic model of cultural learning to a setting in which a single abstract AI system is added to a population of human learners. The AI learns socially from the entire population, and humans may learn individually, from other humans, or from the AI. The authors report that this addition does not raise the equilibrium level of collective world understanding above the individual-learning baseline, which they call an 'AI Rogers' Paradox'. They then simulate several interventions—critical social learning, the option to override AI output, variable AI update schedules, AI individual learning, and a negative-feedback deskilling mechanism—and find that critical appraisal and AI individual learning can improve the equilibrium, while pure reliance on a population-mirroring AI cannot.

Significance. If the central result were a genuinely new phenomenon, the paper would be of broad interest to cultural evolution, AI-assisted decision-making, and human-AI interaction research. The paper has some genuine strengths: it re-instantiates the classic Rogers baseline and the Enquist critical-social-learning framework, it presents analytical derivations for the baseline human-only case in the supplement, it commits to releasing simulation code, and Section 5.6 explicitly acknowledges limitations of the aggregate fitness notion and of the population-mean learning assumption. However, the headline 'AI Rogers' Paradox' is, on close reading, a direct corollary of the defining assumption p^OK_AI := q^OK, rather than a new equilibrium phenomenon. The paper's own Figure 6 shows that an AI with independent individual learning can substantially improve collective understanding, and Section 5.6 concedes that real systems learn from biased data rather than from the population mean. The contribution is therefore best understood as a conditional model exploration, not as a general statement about AIs trained on human data.

major comments (3)
  1. [§2.2 and Supplement, 'Introducing AI to the Network'] The central no-benefit result is forced by the model definition rather than derived or discovered. The supplement states that the AI's adaptation level is set to the mean adaptation status of the population, i.e., p^OK_AI := q^OK. Under this rule the AI is a zero-lag, unbiased aggregator of the population's previous-timestep knowledge, so it carries no information beyond what is already contained in the population mean. The invariance of the equilibrium at 0.58 is then a direct corollary of the classic Rogers/Enquist result, not a new 'AI Rogers' Paradox'. The supplement derives Eqs. 4-5 only for human social learners and gives no analogous closed-form derivation for Eq. 3; without that derivation the paper cannot separate a genuinely novel network effect from a relabeling of the existing equilibrium.
  2. [§2.2 and Supplement, timing of AI update] The timing of the AI update is under-specified in a way that matters for the claimed result. The text says the AI learns from the entire population at the end of each timestep and then humans learn from the AI, but it does not state whether the environment change at the start of the next timestep is applied to the AI's stored knowledge. In the human social-learning channel, §2.1 explicitly discounts the teacher's knowledge by (1-u), giving p^{OK→OK}_s = 0.99. The AI rule p^OK_AI := q^OK applies no such discount, so the AI is artificially more reliable than a human teacher. The simulation code and the analytic statement should specify whether the AI's snapshot is taken before or after the environment-change step; if the snapshot is taken before, the correct rule would be p^OK_AI := (1-u) q^OK, which would require re-running the simulations and re-deriving the claimed invariance.
  3. [§3.2.2, Figure 6, and §5.6] The paper's own results undermine the generality of the headline claim. Figure 6 shows that when the AI can learn individually with low cost and high success, equilibrium collective understanding rises from 0.58 to about 0.93 in the baseline condition and to about 0.89 in the critical-social-learning condition. Section 5.6 similarly concedes that real AI systems learn from biased or non-representative data rather than from the population mean. Therefore the abstract's statement that cheap AI systems trained on all human data 'may not, on its own in the long-term, improve our collective world model' is not a robust property of AI-trained-on-human-data; it is a conditional statement about a perfect population-mean mirror. The abstract, introduction, and conclusion should present the result as conditional on this mirror assumption, with the same emphasis given to the counterexample in Figure 6.
minor comments (4)
  1. [§5.6 and §6] There are several typos: 'fan intuition nad further thinking' should be 'and', and 'demand work work across' should be 'demand work across'.
  2. [§3.2.1 and Supplement, 'How Often Should an AI System Update Its Model of the World?'] The notation is confusing and inconsistent: the environment change probability is called u in §2.1, but the supplement reuses u for the AI's update probability, and the main text presents the update rule as p^{x→qOK}_AI := 1-cλs and p^{x→x}_AI := cλs, where cλs was introduced as a cost. Please use distinct symbols for update probability, update cost, and environment change rate, and give a clean equation for the AI's state after an update.
  3. [Figure 7 and §4] The caption says 'Learners start in the low equilibrium but phase out the AI after a period of time' without identifying which panel (left, center, or right) this refers to, and the text does not clearly explain how the strategy-replication dynamics implement this phase-out; please clarify.
  4. [References] The citation 'N. S and D. N. S. (2021)' is incomplete and nonstandard; please replace it with the full author or institutional name and report number.

Circularity Check

2 steps flagged · score 8.0 of 10

The headline 'AI Rogers' Paradox' is fixed by the definition p^OK_AI := q^OK: the AI is a zero-lag mirror of the population mean, so the invariance of the equilibrium is an algebraic consequence, not an independent finding.

  1. self definitional [Section 2.2 and Eq. (3); Supplement 'Introducing AI to the Network']
    "succeeds with probability equal to adaptation level of the AI, pOK AI := qOK ... At the end of each timestep, the AI system learns socially from the entire population and matches the corresponding probability distribution of strategies (i.e., the AI's adaptation level is set to the mean adaptation status of the population). Running these extended simulations we again find that the average population fitness does not increase relative to the individual learning-only case."

    The AI's success probability is, by definition, the current population mean q^OK. Substituting p^OK_AI=E[q^OK] into the paper's own equilibrium expression Eq. (3) makes the AI term proportional to E[q^OK] itself; the fixed point is then the individual-learning value regardless of what fraction learns from the AI (any channel whose output is the population mean carries no information beyond that mean). The central 'AI Rogers' Paradox' conclusion therefore follows immediately from the assignment p^OK_AI := q^OK rather than from the dynamics simulated in Figure 3.

  2. self definitional [Section 3.1.1; Supplement Figure 10]
    "when we implement such a strategy – by making the AI system unavailable when the expected adaptation value of learning from it is lower than the expected adaptation value of learning individually (E[pAI ]OK < E[pOK i ]) – into our network model, we find that the population equilibria of world understanding do not change."

    Because p^OK_AI = q^OK and the model's baseline equilibrium is q^OK=0.58 while p^OK_i=0.627, the condition E[p_AI]<E[p_i] is always true; the 'critical' strategy simply never activates the AI, reducing the network to the individual-only case. The null result is the same definitional identity restated as an intervention outcome, not an independent discovery about appraisal-based engagement with AI.

full rationale

The paper is self-contained and its base Rogers simulation is a legitimate replication; there is no load-bearing self-citation and the strategy simulations (update schedules, override/critical learning implementation, negative feedback, AI individual learning) are independently coded. However, the central claim of Section 2.2 is definitionally forced: the AI agent is constructed so that its adaptation level equals the population mean, p^OK_AI := q^OK. Under that construction, adding the AI as a learning channel cannot change the population mean at equilibrium by any mechanism other than relabeling; the simulation in Figure 3 is a consistency check, not evidence for the substantive claim that AIs trained on all human data will fail to improve collective understanding. The paper's own Figure 6 and Section 5.6 concede that an AI that individually learns, or that learns from biased rather than mean data, changes or invalidates the result. The no-benefit conclusion is therefore a conditional theorem about a perfect mirror, and the headline paradox is the input assumption wearing the guise of an output.

Assumptions & free parameters 9 free parameters · 6 assumptions · 2 invented entities

The central claims rest on several hand-chosen parameters inherited from the Rogers/Enquist simulations plus new ad hoc mechanisms (AI update rule, negative feedback decay) and one definitional axiom (p^OK_AI := q^OK) that forces the headline paradox. The count of free parameters is substantial relative to the qualitative conclusions, and the paper itself urges caution in interpreting such simple network models.

free parameters (9)
  • environment change rate u = 0.01 (baseline); 0.1, 0.5 (sweeps)
    Probability the optimal behavior changes each timestep; set by hand following the Rogers/Enquist setup. Central to the equilibrium value but the paradox holds across sweeps.
  • individual learning success probability z_i = 0.66
    Chance an individual learner discovers the optimal behavior when attempting individual learning, before cost; taken from Enquist et al. (2007).
  • individual learning cost c_i = 0.05
    Cost of individual learning; taken from Enquist et al. (2007).
  • survival probabilities sOK / sNotOK = 0.93 / 0.85
    Adapted agents survive at higher rate; taken from the Rogers/Enquist model.
  • social learning cost c_s = 0
    Free social learning is the core of Rogers' Paradox.
  • mutation rate for strategy inheritance = 0.005
    Probability offspring flips learning strategy; introduced in the Supplement, chosen by hand, no sensitivity analysis.
  • AI update probability = swept in Fig. 5 (label conflated with u in the text)
    Frequency with which the AI snaps to the population mean; the exact update rule is garbled in Section 3.2.1.
  • AI individual learning cost and success (c_lambda_i, z_AI) = swept over 0-1 grid in Fig. 6
    Swept parameters that determine whether AI individual learning improves or harms population understanding.
  • negative feedback decay factor kappa = 0.9 per AI-learning event
    Each AI social-learning event multiplies the agent's individual-learning efficacy by 0.9; hand-chosen without sensitivity analysis, drives the Section 4 results.
assumptions (6)
  • domain assumption Rogers' (1988) population dynamics: agents survive proportionally to fitness and are replenished, with learned behaviors and strategies inherited with mutation.
    The base model is taken from Rogers (1988) and Enquist et al. (2007); see Section 2.1 and Supplement.
  • domain assumption The world has a single optimal 'OK' behavior that changes with probability u each timestep, and social learning is one timestep delayed.
    Simplified world model; Section 2.1.
  • ad hoc to paper AI adaptation level equals the population mean adaptation at the previous timestep (p^OK_AI := q^OK).
    The AI is modeled as a perfect aggregator of the population; Section 2.2. This assumption forces the 'AI Rogers' Paradox' and is the weakest assumption.
  • domain assumption Critical social learning success is p^OK_cs = 1 - (1-p^OK_s)(1-p^OK_i); agents override the AI output after observing failure.
    Taken from Enquist et al. (2007); Section 3.1.2.
  • ad hoc to paper Negative feedback: each AI social-learning event multiplies the agent's individual-learning efficacy by 0.9, with no recovery mechanism.
    Section 4 and Supplement; hand-chosen mechanism, no empirical grounding or sensitivity analysis.
  • domain assumption A single 'collective world model' exists and fitness is measured by the expected proportion of adapted agents q^OK.
    Section 5.6 acknowledges this is an oversimplification.
invented entities (2)
  • Abstract 'AI agent' that learns socially from the entire population at each timestep
    purpose: Model human social learning from an AI system in the Rogers' framework
    A modeling abstraction; its learning rule is set equal to the population mean (Section 2.2), so it is not independently evidenced.
  • Individual learning penalty kappa_j (deskilling/negative feedback)
    purpose: Capture the intuition that reliance on AI degrades the human's own learning ability
    Introduced in Section 4 with kappa^{t+1} = 0.9 kappa^t; no empirical calibration or falsifiable prediction is offered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Rogers' Paradox in the Context of Human-AI Interaction." pith.science (2026). https://pith.science/paper/GRNOV2HJ

@misc{pith2026250110476,
  author       = {Pith},
  title        = {Pith review of: Revisiting Rogers' Paradox in the Context of Human-AI Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GRNOV2HJ}},
  note         = {Machine review of arXiv:2501.10476}
}
read the original abstract

Humans learn about the world, and how to act in the world, in many ways: from individually conducting experiments to observing and reproducing others' behavior. Different learning strategies come with different costs and likelihoods of successfully learning more about the world. The choice that any one individual makes of how to learn can have an impact on the collective understanding of a whole population if people learn from each other. Alan Rogers developed simulations of a population of agents to study these network phenomena where agents could individually or socially learn amidst a dynamic, uncertain world and uncovered a confusing result: the availability of cheap social learning yielded no benefit to population fitness over individual learning. This paradox spawned decades of work trying to understand and uncover factors that foster the relative benefit of social learning that centuries of human behavior suggest exists. What happens in such network models now that humans can socially learn from AI systems that are themselves socially learning from us? We revisit Rogers' Paradox in the context of human-AI interaction to probe a simplified network of humans and AI systems learning together about an uncertain world. We propose and examine the impact of several learning strategies on the quality of the equilibrium of a society's 'collective world model'. We consider strategies that can be undertaken by various stakeholders involved in a single human-AI interaction: human, AI model builder, and society or regulators around the interaction. We then consider possible negative feedback loops that may arise from humans learning socially from AI: that learning from the AI may impact our own ability to learn about the world. We close with open directions into studying networks of human and AI systems that can be explored in enriched versions of our simulation framework.

Figures

Figures reproduced from arXiv: 2501.10476 by the authors.

Figure 1
Figure 1. Traditional Rogers’ Paradox. In the leftmost panel, we depict individual learning (blue arrows). In the middle panel, we depict social learning (purple arrows). Note that social learning is delayed by one timestep. In the rightmost panel, we depict an example of humans oscillating between individual and social learning. Agents are colored by the behavior they adopt in the current timestep. Agents are considered “ada… view at source ↗
Figure 2
Figure 2. AI Rogers’ Paradox. At each time step, humans can perform individual learning or learn from the AI system, which reverts to the population mean of the previous time step. In this world, individual learning is often considered costly and risky: the environment is stochastic, and the agent has some chance of failing to adapt to individual learning. Social learning may be cheaper and uncertainty-reducing, if many other… view at source ↗
Figure 3
Figure 3. Comparing the collective world understanding over time, in a network where agents can only learn individually and do so at a cost (left) versus a network where agents can learn individually at the same cost or learn socially for free from an AI system that socially learns from all agents in the network (right). Each network attains the same baseline expected collective world understanding: recovering the classic Rog… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Impact of critical social learning (thatched) from the AI over the baseline learning strategy in the presence of AI (filled). Critical social learning leads to increased population world understanding, across varying rates of world change (u = 0.01, 0.1, 0.5); however,…
Figure 5
Figure 5. Figure 5: Impact of the update schedule of AI on the collective world understanding. Each dot represents the average population world understanding quality attained with the specific update schedule, in a world of a particular change rate. We consider three different rates of en…
Figure 6
Figure 6. Figure 6: Allowing the AI system to either individually learn or socially learn on each turn. Impact on collective world understanding (color of cells), depending on the cost to the AI individually learning (x axis) and the expected quality of the AI individually learning about …
Figure 7
Figure 7. Figure 7: Left: Critical social learning with access only to AI and there is negative feedback (learning from AI makes you worse at individual learning). Center: Critical social learning with access to both AI and humans (with negative feedback from AI). Right: Learners start in…
Figure 8
Figure 8. Figure 8: Quality of population world understanding in the original baseline Rogers’ Paradox setting: only individual or social learning from other humans allowed. This is identical, in equilibrium, to a network of only individual learners, or a network of humans and an AI socia…
Figure 9
Figure 9. Figure 9: Social learning frequency versus population world understanding for the baseline AI Rogers’ Paradox case from Section 2. The dashed line indicates the equilibrium of collective world understanding from a network of purely individual learners. Black indicates mean under…
Figure 10
Figure 10. Figure 10: Impact of permitting an agent to decide whether to learn individually or socially from the AI; there is no net impact on collective world understanding. social learn or individual learn before seeing the output of the teacher (in this case the AI). This particular str…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Human-AI Synergy Supports Collective Creative Search

    cs.SI 2026-02 conditional novelty 6.0 of 10

    Hybrid human-AI groups outperform all-human and all-AI groups on a semantic word-search task while keeping guess diversity as high as all-human groups.

  2. Documenting Deployment with Fabric: A Repository of Real-World AI Governance

    cs.CY 2025-08 conditional novelty 6.0 of 10

    Fabric is a public repository of 20 deployed AI systems with co-designed workflow diagrams, plus a four-level human oversight and five-level institutional oversight taxonomy.

Reference graph

Works this paper leans on

25 extracted references · 10 canonical work pages · cited by 2 Pith papers

  1. [4]

    A. Chan, R. Salganik, A. Markelius, C. Pang, N. Rajkumar, D. Krasheninnikov, L. Langosco, Z. He, Y. Duan, M. Carroll, et al. Harms from increasingly agentic algorithmic systems. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , pages 651–666,

    work page 2023

  2. [10]

    individual learning penalty

    When Should You Override from an AI System? We next imagine that the agent can tell whether they became adapted after attempting to learn: that is, the agent can evaluate how successful a learning interaction was, in line with [Enquist et al., 2007]. We extend the simulation to consider the case where agents can override the outputs of the AI system by le...

    work page 2007

  3. [11]

    Gurnee and M

    W. Gurnee and M. Tegmark. Language models represent space and time. arXiv preprint arXiv:2310.02207,

    arXiv

  4. [12]

    S. Hao, Y. Gu, H. Ma, J. J. Hong, Z. Wang, D. Z. Wang, and Z. Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992,

    arXiv

  5. [14]

    URL http:// arxiv.org/abs/1706.06551. J. Hernández-Orallo and K. Vold. Ai extenders: the ethical and societal implications of humans cognitively extended by ai. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 507–513,

    arXiv 2019

  6. [15]

    Hinton, O

    G. Hinton, O. Vinyals, and J. e. a. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7),

    arXiv

  7. [16]

    Jaech, A

    A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720,

    arXiv

  8. [17]

    V . Lai, C. Chen, A. Smith-Renner, Q. V . Liao, and C. Tan. Towards a science of human-ai decision making: An overview of design space in empirical human-subject studies. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1369–1385,

    work page 2023

Show all 25 references
  1. [18]

    R. Liu, J. Geng, J. C. Peterson, I. Sucholutsky, and T. L. Griffiths. Large language models assume people are more rational than we really are. arXiv preprint arXiv:2406.17055,

    arXiv

  2. [19]

    doi: 10.2760/376602(online)

    ISSN 1831-9424 (online). doi: 10.2760/376602(online). A. Salinas, P . Shah, Y. Huang, R. McCormack, and F. Morstatter. The unequal opportunities of large language models: Examining demographic biases in job recommendations by chatgpt and llama. In Proceedings of the 3rd ACM Co...

    doi

  3. [20]

    Silver, T

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Ku- maran, T. Graepel, et al. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815,

    arXiv

  4. [21]

    Snell, J

    C. Snell, J. Lee, K. Xu, and A. Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314,

    arXiv

  5. [22]

    Sucholutsky, R

    I. Sucholutsky, R. M. Battleday, K. M. Collins, R. Marjieh, J. Peterson, P . Singh, U. Bhatt, N. Jacoby, A. Weller, and T. L. Griffiths. On the informativeness of supervision signals. In Uncertainty in Artificial Intelligence, pages 2036–2046. PMLR, 2023a. 19 I. Sucholutsky, L...

    arXiv

  6. [24]

    K. Vafa, J. Y. Chen, A. Rambachan, J. Kleinberg, and S. Mullainathan. Evaluating the world model implicit in a generative model. In Neural Information Processing Systems, 2024a. K. Vafa, A. Rambachan, and S. Mullainathan. Do large language models perform the way people expect?...

    arXiv

  7. [25]

    L. Wong, G. Grand, A. K. Lew, N. D. Goodman, and V . K. e. a. Mansinghka. From word models to world models: Translating from natural language to the probabilistic language of thought. arXiv preprint arXiv:2306.12672, pages arXiv–2306,

    arXiv

  8. [1993]

    doi: 10.1017/S0140525X0003123X. M. Tomasello, M. Carpenter, J. Call, T. Behne, and H. Moll. Understanding and sharing intentions: The origins of cultural cognition. Behavioral and brain sciences, 28(5):675–691,

    doi

  9. [2007]

    Gabriel, A

    I. Gabriel, A. Manzini, G. Keeling, L. A. Hendricks, and V . e. a. Rieser. The ethics of advanced ai assistants. arXiv preprint arXiv:2404.16244,

    arXiv

  10. [2017]

    Ehsan, Q

    U. Ehsan, Q. V . Liao, M. Muller, M. O. Riedl, and J. D. Weisz. Expanding explainability: Towards social transparency in ai systems. In Proceedings of the 2021 CHI conference on human factors in computing systems, pages 1–19,

    2021

  11. [2018]

    Doshi-Velez and B

    F. Doshi-Velez and B. Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608,

    arXiv

  12. [2019]

    Z. Guo, Y. Wu, J. D. Hartline, and J. Hullman. A decision theoretic framework for measuring ai reliance. In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 221–236,

    2024

  13. [2020]

    URL https://www

    doi: 10.1073/pnas.1912341117. URL https://www. pnas.org/doi/abs/10.1073/pnas.1912341117. 15 J. Andreas. Language models as agent models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5769–5779,

    2022 doi

  14. [2021]

    ISBN 9781450384735. U. Bhatt, V . Chen, K. M. Collins, P . Kamalaruban, E. Kallina, A. Weller, and A. Talwalkar. Learning personalized decision support policies. arXiv preprint arXiv:2304.06701,

    arXiv

  15. [2022]

    K. M. Collins, M. Barker, M. Espinosa Zarlenga, N. Raman, and U. e. a. Bhatt. Human uncertainty in concept-based ai systems. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 869–889,

    2023

  16. [2023]

    K. M. Collins, V . Chen, I. Sucholutsky, H. R. Kirk, M. Sadek, H. Sargeant, A. Talwalkar, A. Weller, and U. Bhatt. Modulating language model experiences through frictions. In NeurIPS 2024 Workshop on Behavioral Machine Learning, 2024a. K. M. Collins, A. Q. Jiang, S. Frieder, L...

    2024 arXiv

  17. [2024]

    Bhatt, J

    U. Bhatt, J. Antorán, Y. Zhang, Q. V . Liao, and P . e. a. Sattigeri. Uncertainty as a form of transparency: Measuring, communicating, and using uncertainty. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 401–413, New York, NY, USA,

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.