From prosthetic memory to prosthetic denial: Auditing whether large language models are prone to mass atrocity denialism

Daniel Bultmann; David J. Simon; Eve M. Zucker; Mykola Makhortykh; Roberto Ulloa

REVIEW 3 major objections 6 minor 10 references

From prosthetic memory to prosthetic denial: Auditing whether large language models are prone to mass atrocity denialism

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Audit: LLMs get 10% of genocide answers wrong

desk verdict A transparent comparative audit with a real Cambodian/Khmer finding, but the headline metric measures yes/no label agreement, not the paper's own concept of prosthetic denial, because justifications of correct answers are never coded. read the letter →

arxiv 2505.21753 v1 pith:4HKHH67Q submitted 2025-05-27 cs.CY cs.CL

Roberto Ulloa , Eve M. Zucker , Daniel Bultmann , David J. Simon , Mykola Makhortykh This is my paper

classification cs.CYcs.CL

keywords largelanguagemodelsprostheticmemorydenialgenocidedenialismAIauditcross-lingualbiasmassatrocityKhmer

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 large language models preserve or distort collective memory of mass atrocities, and answers with a comparative audit of five LLMs across four genocide cases and two languages per case. The authors propose that LLMs can act as sites of 'prosthetic memory'—mediated encounters with the past—but can also generate what they call 'prosthetic denial,' the AI-mediated erasure or distortion of atrocity memory. Across 370 prompts, 37 responses (10%) contradicted the expert baseline, with the worst performance on the Cambodian Genocide (20 of 100 incorrect, most in Khmer) and near-perfect performance on the Holodomor. The paper argues this asymmetry tracks the amount of documentation and the representation of the relevant language in LLM training data, and that even correct labels are sometimes undercut by denialist framing in the justification. The stakes are practical: if LLMs increasingly stand in for search engines and fact-checkers, undetected revisionist narratives in their outputs can shape how millions remember—or deny—atrocities.

What carries the argument

The load-bearing mechanism is the comparative audit itself, built on expert-defined baselines: for each of four mass atrocities, nine prompts encode common denialist claims, and each prompt is asked in English and in a language central to the case (Ukrainian, German, Khmer, French). A uniform template forces a JSON response with a yes/no label and a justification, so the study can separate label accuracy from narrative framing. The concept doing the theoretical work is 'prosthetic denial,' defined as the AI-mediated erasure or distortion of atrocity memory, set against 'prosthetic memory,' the mediated experience of the past. Language representation statistics from the web serve as a proxy for the training data likely available to the models, linking observed performance gaps to data availability.

What would settle it

Code all 370 responses' justifications—including those attached to correct labels—for denialist tropes, and compare rates across cases; if the Holocaust and Holodomor justifications contain denialist framing at rates comparable to Cambodia, the claim that only underrepresented cases are susceptible would fail.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central finding is that LLM accuracy about mass atrocities is uneven and consequential: models answer reliably for widely documented, legally shielded events such as the Holocaust and the Holodomor, but they reproduce revisionist and denialist narratives for less-documented cases. The Cambodian Genocide is the clearest example: 18 of the 20 incorrect responses to Khmer prompts echoed nationalist tropes blaming Vietnam, portrayed Khmer Rouge leaders as patriotic defenders, or treated contested claims as open questions, and Mixtral failed to produce coherent answers in Khmer at all. The audit also found inconsistencies between binary yes/no labels and their justifications, such as Gemini affirming that Ukraine suffered no differently from other Soviet republics while its explanation stated the opposite. These contradictions matter because users may absorb denialist framing even when the headline answer is correct.

Load-bearing premise

The load-bearing premise is that matching the expert yes/no baseline measures whether a model contributes to denialism, even though denialist framing can hide inside the justifications of correct answers and the paper acknowledges this possibility.

Editorial extensions

If this is right

LLM outputs about mass atrocities are not uniformly reliable; accuracy degrades for less-documented events and low-resource languages, so users in Khmer face materially worse information about the Cambodian Genocide.
A correct yes/no label does not guarantee a safe response, because justifications can contradict the label or echo denialist tropes; evaluating model safety on label accuracy alone will miss this.
Unmoderated LLM use risks reinforcing existing asymmetries in historical representation, amplifying nationalist or revisionist narratives for events that already receive less attention.
The small replication variation (96.2% match over four months) suggests the patterns are stable in the short term, but model updates can shift performance, as Mixtral's Khmer answers improved.
If LLMs replace search engines and fact-checkers, the opacity of their sourcing means users cannot easily detect when a confident answer mixes fact with denialist framing.

Reading between the lines

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

If justifications were coded for denialist rhetoric regardless of label correctness, the share of problematic outputs would likely exceed the 10% label-error rate, because the paper itself notes similar statements might appear in justifications of correct answers.
The same audit design could be extended to other contested histories—for example colonial atrocities or the Armenian genocide—with the expectation that documentation volume and language representation will predict error patterns.
The label-justification contradiction points to a distinct failure mode: safety fine-tuning may suppress overt denial at the answer level while leaving narrative content to be generated probabilistically, so denialism can leak through in softer form.
As chatbots become multimodal and more affectively expressive, the same revisionist framings could carry greater emotional credibility, making the prosthetic-denial risk larger than the text-only audit captures.

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

Two things to know. First, this is the first multi-case, two-language audit of LLM atrocity denialism, and the main empirical pattern — degraded performance for the Cambodian Genocide, especially in Khmer — is probably real. Second, the quantitative core measures whether the parsed yes/no label matches an expert baseline, not whether the output contains denialist framing. The paper defines prosthetic denial as AI-mediated erasure or distortion, but then operationalizes it as label accuracy. That is a real gap, and the authors know it: they write in the Discussion that similar statements might appear in justifications of correct answers.

What is genuinely new: the comparative design across four genocides, five models, and two languages, with expert-built baselines and a replication round four months later. The data are on OSF. The qualitative analysis of the incorrect answers is careful and specific, showing how outputs reproduce nationalist and revisionist tropes about Vietnam, US bombing, and Khmer Rouge patriots. The replication showing 96.2% label agreement is a useful stability check. The concept of prosthetic denial is a reasonable extension of Landsberg's framework, even if it is a framing label rather than a fitted parameter.

Where the soft spots are: sample size is 9 prompts per model-language-case, with no statistical inference or uncertainty quantification. That is minor for an audit that is transparently descriptive, but it limits how strongly the cross-case differences can be stated. The bigger issue is the justification gap. The paper counts a label-correct answer as accurate even when the justification contains revisionist framing — and its own examples show this is not hypothetical: Gemini Q1 English deflects responsibility to Vietnam, Gemini Q3 Khmer attributes the genocide to US bombing. So the reported 20/100 Cambodia error rate is a lower bound on denialist output, not a measure of prosthetic denial. The Holocaust/Cambodia contrast might be even starker if justifications were coded, or it might shift for some models. The paper should either code justifications for all 370 responses or clearly reframe the claims as label accuracy rather than susceptibility to denialist framing.

Who this is for: researchers in AI auditing, memory studies, and genocide studies who want a systematic, reproducible snapshot of how LLMs handle contested atrocity history. The central argument — that underrepresented events and low-resource languages are more vulnerable — holds up as a direction, even if the precise rates need refinement.

Recommendation: send it to peer review. It deserves a serious referee. The authors have done the hard empirical work honestly; the requested revision is feasible and would make the paper significantly stronger.

Referee Report

3 major / 6 minor

Summary. The paper proposes the concept of 'prosthetic denial' to describe AI-mediated erasure or distortion of atrocity memory and reports a comparative audit of five LLMs (Claude, GPT, Llama, Mixtral, Gemini) across four cases (Holodomor, Holocaust, Cambodian Genocide, genocide against the Tutsi in Rwanda). Each model was prompted with nine questions per case in English and a case-relevant language, and the parsed yes/no answers were compared against expert-defined baselines. The authors report 37 incorrect answers out of 370 prompts (10%), with the highest error concentration in the Cambodian Genocide case, especially for Khmer prompts, and qualitatively describe how some outputs reproduce revisionist or denialist framings. The paper argues that LLMs can act as sites of prosthetic memory but that unmoderated use risks reinforcing historical denialism, particularly for underrepresented cases and low-resource languages.

Significance. If the central empirical claim is fully supported, the paper addresses an important and under-studied question: whether LLM outputs about mass atrocities are uniformly reliable or degrade for less-documented events and lower-resource languages. The audit design is transparent and valuable: it uses expert-derived baselines, covers multiple models and languages, includes a replication wave, and makes incorrect-answer justifications available via an OSF appendix. The conceptual framing of 'prosthetic denial' is a useful contribution to the memory-studies literature and to AI auditing discussions. The main quantitative headline—that Cambodia and Khmer prompts produce more errors—is plausible, but the evidence as currently coded does not fully measure the paper's own construct of prosthetic denial, because denialist framing inside label-correct justifications is not systematically assessed. The paper's strengths are its cross-case design, the clear prompt set, and the transparency about replication; its main weakness is the gap between the measured outcome (yes/no label agreement) and the claimed outcome (susceptibility to denialist framings).

major comments (3)

[Methods and Results (quantitative analysis; Discussion)] The central quantitative outcome is only whether the parsed yes/no label matches the expert baseline: the Methods state 'we compared the parsed outputs obtained from all prompts against the baseline answer,' and the Results state 'we only reviewed justifications for incorrect answers.' Yet the paper's core construct, 'prosthetic denial,' is defined as AI-mediated erasure or distortion of atrocity-related past. A label-correct response that still reproduces a revisionist frame—such as the Gemini Q1 English Cambodia response invoking Vietnam's later invasion, or the ChatGPT Q2y Khmer response attributing atrocities to a 'bad faction' of the Khmer Rouge—is counted as accurate despite being denialist in framing. The Discussion explicitly concedes that 'similar statements might appear in the justifications of correct answers as well.' Because justifications for correct answers were never coded, the reported 20/100 Cambodia error rate and the Cambodia-versus-Holocaust contrast cannot be interpreted as rates of denialist output. This is load-bearing for the abstract's claim about 'susceptibility to denialist framings' and needs to be fixed, for example by coding all 370 justifications with a rubric for denialist framing or by restricting the quantitative claim to label accuracy.
[Results, Figure 1 and overall counts] The paper reports raw proportions (37/370, 20/100 for Cambodia, 2/90 for Holodomor) without any statistical inference, confidence intervals, or adjustment for the nested structure of the data (9 prompts per model-language-case, repeated across five models). With only nine prompts per model-language-case, the cross-case differences could be influenced by prompt selection and stochastic variation; the replication wave shows 14 mismatches between the original and replication labels, further indicating variability. To support the comparative claim that Cambodia is significantly worse than the Holocaust or Holodomor, the authors should provide at least exact binomial confidence intervals or a simple regression/test that accounts for the grouping of prompts by case, model, and language. This is a fixable but necessary addition to make the central quantitative contrast interpretable.
[Results, Rwanda Q8 and Discussion, Bagosora] There is an internal inconsistency in the reporting of the Bagosora question (Q8). The Results state that 'six out of the eight answers incorrectly affirmed the proposition that Bagosora stands convicted of conspiracy,' while the Discussion states that 'several models answered correctly that no conspiracy conviction occurred, yet their justifications oversimplified or ambiguously described the legal rationale.' These two statements cannot both describe the same set of responses: the first says most answers were label-incorrect, the second implies most answers were label-correct. This contradiction affects the interpretation of the Rwanda case and the general point about label-justification mismatches; the authors should clarify which models gave which labels and reconcile the two passages.

minor comments (6)

[Methods, Table 5 note] The text below Table 5 says the placeholder [QUESTION] 'was replaced by the corresponding question in Table 1,' but the question list is Table 3; this cross-reference should be corrected.
[Results, replication paragraph] The replication paragraph contains a typo: 'Mixtra' should be 'Mixtral' in the sentence about the Khmer-language improvements.
[Results, section header] The section header 'Holomodor' is misspelled; it should be 'Holodomor'.
[Methods, model selection] The Methods say the 'considered best model' of each provider was used, but the selection criterion is not defined; specifying the criterion (e.g., provider documentation, benchmark rankings) would strengthen reproducibility.
[Methods, prompt template and API settings] The paper does not report API parameters such as temperature or decoding settings, which are known to affect LLM output variability; reporting these would improve the replicability of the audit beyond the two collection dates.
[Figure 1 caption] The caption says 'Number of incorrect answers per model per genocide,' but the study includes the Holodomor and the Holocaust, which are not all termed 'genocide' in the same way; 'per case study' would be more accurate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the audit compares LLM outputs against externally defined expert baselines, and the acknowledged 'correct justifications' gap is a measurement-validity limitation rather than a circular derivation.

full rationale

The derivation chain in this paper is empirical and externally grounded. Each prompt is scored by comparing the parsed yes/no answer to a baseline answer fixed by the authors' expertise (Methods: 'For the quantitative analysis, we compared the parsed outputs obtained from all prompts against the baseline answer defined by the experts'). The central quantitative claim (37/370 incorrect, 20/100 Cambodian Genocide prompts) is therefore a measurement against external historical consensus, not a quantity defined by the model outputs or by the paper's own concept. The concept 'prosthetic denial' is introduced as a framing label for the phenomenon under study; it is not fitted, estimated, or defined in terms of the reported label-mismatch rate. There are self-citations (e.g., Makhortykh et al., 2023b), but they provide conceptual context and prior motivation only; no uniqueness theorem, fitted parameter, or load-bearing premise is imported from them. The paper does contain one acknowledged measurement gap: justifications were reviewed only for answers that already mismatched the baseline, and the Discussion concedes, 'Notably, similar statements might appear in the justifications of correct answers as well.' This means the label-mismatch rate may undercount denialist framing inside otherwise correct outputs. That is a construct-validity caveat about what the quantitative outcome measures, not a case of a result being equivalent to its input by construction or by definition. No circular step can be exhibited from the paper's own equations or citation chain, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The audit is an empirical comparison against an external expert baseline, so it introduces no fitted parameters. The main assumptions concern the validity of the expert baselines, the representativeness of the questions, the proxy for training data availability, and the use of binary labels as the accuracy metric.

assumptions (4)

domain assumption The expert-defined baseline answers are historically correct and unambiguous.
The audit relies on the authors' expertise to set the expected answer for each of the 36 questions. For contested questions such as whether 'genocide' accurately describes the Holodomor or the Holocaust, the baseline is a scholarly judgment, not an uncontested fact.
domain assumption The nine questions per case are representative of denialist narratives and the translations preserve the intended meaning.
The authors selected questions to address 'themes found in denialist narratives.' Translation quality matters, especially for Khmer, where the models already struggle; the paper includes two versions for two Khmer questions to account for terminology.
domain assumption Web language prevalence statistics approximate the amount of LLM training data available for each language.
The paper uses Observatory of Linguistic and Cultural Diversity on the Internet statistics to infer that Khmer is underrepresented in training data (0.12% of web content). Training data composition is not publicly disclosed, so this proxy is used to explain the cross-lingual performance gap.
domain assumption A binary yes/no label that matches the baseline is a valid indicator of whether the model contributes to denialism.
The quantitative analysis compares only parsed yes/no outputs to baselines. The paper does not systematically code justifications for correct answers, even though the discussion states that denialist statements may also appear in correct-answer justifications.

invented entities (1)

prosthetic denial
purpose: A concept describing AI-mediated erasure or distortion of atrocity memories, used to frame the audit's findings and distinguish them from prosthetic memory.
The term is introduced and operationalized within this paper. It does not yet have independent falsifiable predictions or measures outside the audit design; it is a label for observed patterns of incorrect or misleading LLM responses.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From prosthetic memory to prosthetic denial: Auditing whether large language models are prone to mass atrocity denialism." pith.science (2026). https://pith.science/paper/4HKHH67Q

@misc{pith2026250521753,
  author       = {Pith},
  title        = {Pith review of: From prosthetic memory to prosthetic denial: Auditing whether large language models are prone to mass atrocity denialism},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HKHH67Q}},
  note         = {Machine review of arXiv:2505.21753}
}

read the original abstract

The proliferation of large language models (LLMs) can influence how historical narratives are disseminated and perceived. This study explores the implications of LLMs' responses on the representation of mass atrocity memory, examining whether generative AI systems contribute to prosthetic memory, i.e., mediated experiences of historical events, or to what we term "prosthetic denial," the AI-mediated erasure or distortion of atrocity memories. We argue that LLMs function as interfaces that can elicit prosthetic memories and, therefore, act as experiential sites for memory transmission, but also introduce risks of denialism, particularly when their outputs align with contested or revisionist narratives. To empirically assess these risks, we conducted a comparative audit of five LLMs (Claude, GPT, Llama, Mixtral, and Gemini) across four historical case studies: the Holodomor, the Holocaust, the Cambodian Genocide, and the genocide against the Tutsis in Rwanda. Each model was prompted with questions addressing common denialist claims in English and an alternative language relevant to each case (Ukrainian, German, Khmer, and French). Our findings reveal that while LLMs generally produce accurate responses for widely documented events like the Holocaust, significant inconsistencies and susceptibility to denialist framings are observed for more underrepresented cases like the Cambodian Genocide. The disparities highlight the influence of training data availability and the probabilistic nature of LLM responses on memory integrity. We conclude that while LLMs extend the concept of prosthetic memory, their unmoderated use risks reinforcing historical denialism, raising ethical concerns for (digital) memory preservation, and potentially challenging the advantageous role of technology associated with the original values of prosthetic memory.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 6 canonical work pages

[1]

L., Corney, D., DiResta, R., Ferrara, E., Hale, S., Halevy, A., Hovy, E., Ji, H., Menczer, F., Miguez, R., Nakov, P., Scheufele, D., Sharma, S., & Zagni, G

Augenstein, I., Baldwin, T., Cha, M., Chakraborty, T., Ciampaglia, G. L., Corney, D., DiResta, R., Ferrara, E., Hale, S., Halevy, A., Hovy, E., Ji, H., Menczer, F., Miguez, R., Nakov, P., Scheufele, D., Sharma, S., & Zagni, G. (2024). Factuality challenges in the era of large language models and opportunities for fact-checking. Nature Machine Intelligence...

arXiv 2024
[2]

Mierwald, M. (2024). Chatting about the Past with Artificial Intelligence: A Case Study of Pupils’ Interaction with ChatGPT while Completing a History-Learning Task. Journal of Educational Media, Memory, and Society, 16(2), 143-173. Newton, C. (2025, May 9). Stats from a dying web. Retrieved from https://www.platformer.news/safari-search-decline-apple-goo...

work page 2024
[6]

(1976, October 4)

Lewis, A. (1976, October 4). Menu for disaster. The New York Times. Retrieved from https://www.nytimes.com, May 19, 2025 Li, Z., Shi, Y., Liu, Z., Yang, F., Payani, A., Liu, N., & Du, M. (2025, April). Language ranker: A metric for quantifying llm performance across high and low-resource languages. In Proceedings of the AAAI Conference on Artificial Intel...

work page 2025
[7]

https://doi.org/10.3389/frai.2024.1341697 Reading, A. (2009). Memobilia: The Mobile Phone and the Emergence of Wearable Memories. In: Garde-Hansen, J., Hoskins, A., Reading, A. (eds) Save As … Digital Memories. Palgrave Macmillan, London. https://doi.org/10.1057/9780230239418_5 Richardson-Walden, V. G., & Makhortykh, M. (2024). Imagining Human-AI Memory S...

work page arXiv 2009
[28]

Makhortykh, M., Sydorova, M., Baghumyan, A., Vziatysheva, V., & Kuznetsova, E. (2024). Stochastic lies: How LLM-powered chatbots deal with Russian disinformation about the war in Ukraine. Harvard Kennedy School Misinformation Review . Manning, C. D. (2022). Human Language Understanding & Reasoning. Daedalus, 151(2), 127–138. https://doi.org/10.1162/daed_a...

doi:10.1162/daed_a_01905 2024
[305]

Gadzins’ka, I. (2021). Desiat’ milioniv piatsot tysh. I krapka. Ukraintsiam nav’iazyit’ superechlivi dani pro 10,5 mil’ioniv zhertv Golodomory. Istorycha Pravda, 26 November. Available at: https://www.istpravda.com.ua/articles/2021/11/26/160564/ (accessed 18 May

work page 2021
[1936]

Bloch, M

New York . Bloch, M. (1998). Autobiographical Memory and the Historical Memory of the More Distant Past, in How We Think They Think: Anthropological Approaches to Cognition, Memory, and Literacy , Westview Press, 114-130. DeVerna, M. R., Yan, H. Y., Yang, K.-C., & Menczer, F. (2024). Fact-checking information from large language models can decrease headli...

doi:10.1073/pnas.2322823121 1998
[2018]

25 Landsberg, A. (2004). Prosthetic memory: The transformation of American remembrance in the age of mass culture . Columbia University Press. Lee, J. (2008). Rwanda: No Conspiracy, No Genocide Planning … No Genocide? JURIST. Retrieved from https://www.jurist.org/commentary/2008/12/rwanda-no-conspiracy-no-genocide/, May 26,

work page 2004

Show all 10 references

[2024]

https://www.obdilci.org/Base/en/search Pisanty, V

[Database]. https://www.obdilci.org/Base/en/search Pisanty, V. (2023). Identification and Make-Believe: the Fallacies of Prosthetic Memory. Witnessing the Witness of War Crimes, Mass Murder, and Genocide . Quelle, D., & Bovet, A. (2024). The perils and promises of fact-checkin...

2023
[2025]

Garde-Hansen, J., Hoskins, A., & Reading, A. (Eds.). (2009). Save as... digital memories. Springer. 24 Hirsch, M. (2008). The Generation of Postmemory. Poetics Today, 29(1), 103–128. https://doi.org/10.1215/03335372-2007-019 Hitchcott, N. (2021). Seeing the Genocide against th...

2009 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

L., Corney, D., DiResta, R., Ferrara, E., Hale, S., Halevy, A., Hovy, E., Ji, H., Menczer, F., Miguez, R., Nakov, P., Scheufele, D., Sharma, S., & Zagni, G

Augenstein, I., Baldwin, T., Cha, M., Chakraborty, T., Ciampaglia, G. L., Corney, D., DiResta, R., Ferrara, E., Hale, S., Halevy, A., Hovy, E., Ji, H., Menczer, F., Miguez, R., Nakov, P., Scheufele, D., Sharma, S., & Zagni, G. (2024). Factuality challenges in the era of large language models and opportunities for fact-checking. Nature Machine Intelligence...

arXiv 2024

[2] [2]

Mierwald, M. (2024). Chatting about the Past with Artificial Intelligence: A Case Study of Pupils’ Interaction with ChatGPT while Completing a History-Learning Task. Journal of Educational Media, Memory, and Society, 16(2), 143-173. Newton, C. (2025, May 9). Stats from a dying web. Retrieved from https://www.platformer.news/safari-search-decline-apple-goo...

work page 2024

[3] [6]

(1976, October 4)

Lewis, A. (1976, October 4). Menu for disaster. The New York Times. Retrieved from https://www.nytimes.com, May 19, 2025 Li, Z., Shi, Y., Liu, Z., Yang, F., Payani, A., Liu, N., & Du, M. (2025, April). Language ranker: A metric for quantifying llm performance across high and low-resource languages. In Proceedings of the AAAI Conference on Artificial Intel...

work page 2025

[4] [7]

https://doi.org/10.3389/frai.2024.1341697 Reading, A. (2009). Memobilia: The Mobile Phone and the Emergence of Wearable Memories. In: Garde-Hansen, J., Hoskins, A., Reading, A. (eds) Save As … Digital Memories. Palgrave Macmillan, London. https://doi.org/10.1057/9780230239418_5 Richardson-Walden, V. G., & Makhortykh, M. (2024). Imagining Human-AI Memory S...

work page arXiv 2009

[5] [28]

Makhortykh, M., Sydorova, M., Baghumyan, A., Vziatysheva, V., & Kuznetsova, E. (2024). Stochastic lies: How LLM-powered chatbots deal with Russian disinformation about the war in Ukraine. Harvard Kennedy School Misinformation Review . Manning, C. D. (2022). Human Language Understanding & Reasoning. Daedalus, 151(2), 127–138. https://doi.org/10.1162/daed_a...

doi:10.1162/daed_a_01905 2024

[6] [305]

Gadzins’ka, I. (2021). Desiat’ milioniv piatsot tysh. I krapka. Ukraintsiam nav’iazyit’ superechlivi dani pro 10,5 mil’ioniv zhertv Golodomory. Istorycha Pravda, 26 November. Available at: https://www.istpravda.com.ua/articles/2021/11/26/160564/ (accessed 18 May

work page 2021

[7] [1936]

Bloch, M

New York . Bloch, M. (1998). Autobiographical Memory and the Historical Memory of the More Distant Past, in How We Think They Think: Anthropological Approaches to Cognition, Memory, and Literacy , Westview Press, 114-130. DeVerna, M. R., Yan, H. Y., Yang, K.-C., & Menczer, F. (2024). Fact-checking information from large language models can decrease headli...

doi:10.1073/pnas.2322823121 1998

[8] [2018]

25 Landsberg, A. (2004). Prosthetic memory: The transformation of American remembrance in the age of mass culture . Columbia University Press. Lee, J. (2008). Rwanda: No Conspiracy, No Genocide Planning … No Genocide? JURIST. Retrieved from https://www.jurist.org/commentary/2008/12/rwanda-no-conspiracy-no-genocide/, May 26,

work page 2004