Epistemic Injustice in Generative AI

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The rapid advancement of generative AI has been accompanied by growing societal concerns and numerous instances of misuse.

“The rapid advancement of generative AI, marked by accelerated software and hardware innovation and a proliferation of novel applications, has been accompanied by growing societal concerns and numerous instances of misuse (Bender et al. 2021;Weidinger et al. 2022;Bird, Ungless, and Kasirzadeh 2023).”

• cites Bender et al. 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Weidinger et al. 2022 sourcePartly supported

  The Introduction (page unavailable) describes growing language-model research and breakthroughs, and §2.4 (page unavailable) identifies societal risks from malicious uses. But §2.4 largely discusses anticipated uses, while §2.4.1 gives a demonstration that GPT-3 could write fake news; these passages do not establish numerous actual instances of misuse or the claimed acceleration in hardware innovation. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites Bird, Ungless, and Kasirzadeh 2023 sourceSupported

May combine two claims.

The only work with a focus on the epistemic injustice of generative AI is by De Proost and Pozzi (2023), which looks at the potential of conversational AI for hermeneutical ignorance.

“To the best of our knowledge, the only work with a focus on the epistemic injustice of generative AI is by De Proost and Pozzi (2023), which looks at the potential of conversational AI for hermeneutical ignorance.”

• cites De Proost and Pozzi 2023 sourcePartly supported

  In “Conversational Artificial Intelligence and the Potential for Epistemic Injustice” (page unavailable), the authors examine how conversational AI in psychotherapy could cause testimonial and hermeneutical injustice, including patients being unable to make sense of experiences the system cannot recognize. The copy does not identify the systems as generative AI or use the term “hermeneutical ignorance”; its stated focus is hermeneutical injustice in a specific therapeutic setting. The claim that this is the only such work is the citing authors’ own qualification, not a finding attributed to the source. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

De Proost and Pozzi's study presents two significant limitations: its analysis is confined to textual dialogue interactions, and it does not cover multimodal systems.

“However, this study presents two significant limitations.First, its analysis is confined to textual dialogue interactions.”

• cites De Proost and Pozzi 2023 sourcePartly supported

  In “Conversational Artificial Intelligence and the Potential for Epistemic Injustice” (page unavailable), the authors say they examine testimonial and hermeneutical injustice in conversational AI used in psychotherapy. The copy does not establish that those interactions are exclusively textual or identify the omission of multimodal systems as a limitation; searches for text-only and multimodal interactions found no such passage. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

May combine two claims.

Kazenwadel and Steinert (2023) found that GPT-3.5 reported higher fatality numbers in Arabic than in Hebrew for Israeli airstrikes, and in one case was more likely to deny a particular airstrike when asked in Hebrew.

“When asked about Israeli airstrikes, the model reports higher fatality numbers in Arabic than in Hebrew, and in one case, GPT-3.5 was more likely to deny the existence of a particular airstrike when asked about it in Hebrew.”

• cites Kazenwadel and Steinert (2023) sourcePartly supported

  The Introduction and Research Design report higher GPT-3.5 fatality estimates in Arabic than in Hebrew for Israeli airstrikes. Research Design also reports more responses denying an event in Hebrew than in Arabic overall (29 versus 5), but it does not establish that GPT-3.5 was more likely to deny any particular airstrike when asked in Hebrew. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Researchers have studied a backfire effect in which an intervention attempting to change an individual's belief ends up reinforcing their belief.

“Researchers have also studied the backfire effect, in which an intervention attempting to change an individual's belief ends up reinforcing their belief (Swire-Thompson, DeGutis, and Lazer 2020).”

• cites Swire-Thompson, DeGutis, and Lazer 2020 sourceSupported

May combine two claims.

Misinformation infiltrates generative model outputs through two primary pathways: memorization of inaccuracies from the source dataset and generation of high-likelihood sequences containing clear factual errors.

“This analysis brings us to two primary pathways through which misinformation infiltrates the outputs of generative models: the memorization of inaccuracies from the source dataset and the generation of high-likelihood sequences that, despite the model's prediction, contain clear factual errors (Augenstein et al. 2023).”

• cites Augenstein et al. 2023 sourcePartly supported

  The Introduction and §2.1 Factuality Challenges say LLMs can generate factually false content, and the Introduction notes that training datasets can introduce biases. But our copy does not identify memorized dataset inaccuracies and high-likelihood yet erroneous sequences as two primary pathways for misinformation; searches for memorization and likelihood did not establish those mechanisms. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Generative AI's output can be trusted over the lived experiences and knowledge of marginalized individuals, reinforcing existing power imbalances and perpetuating testimonial injustice.

“This can lead to a situation where the generative AI's output is trusted over the lived experiences and knowledge of marginalized individuals, thus reinforcing existing power imbalances and perpetuating testimonial injustice.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The discourse surrounding generative AI's rapid advancement lacks a philosophical account that coherently relates epistemic concerns and explains how they constitute moral violations of a unifying principle.

“Despite these escalating societal concerns and numerous instances of misuse, the discourse surrounding generative AI's rapid advancement lacks a philosophical account that coherently relates these epistemic concerns and explains how they constitute moral violations of a unifying principle.”

May combine two claims.

Marginalized groups experience a decline in trust towards the system itself, the entity that established it, and other institutions tasked with upholding veracity.

“Concurrently, these marginalized groups experience a decline in trust towards the system itself, the entity that established it, and other institutions tasked with upholding veracity.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

In the algorithmic setting, testimonial injustice requires a credibility excess assigned to the algorithm; humans believe the account amplified through the technology over the individual or group who is discredited.

“In the algorithmic setting, the injustice requires a credibility excess assigned to the algorithm; that is, humans believe the account amplified through the technology over the individual or group who is discredited.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The distinctive characteristics of generative AI give rise to novel forms of epistemic injustice that necessitate a dedicated analytical framework.

“However, we argue that the distinctive characteristics of generative AI give rise to novel forms of epistemic injustice that necessitate a dedicated analytical framework.”

May combine two claims.

Generative AI models are more likely to reproduce the voices of those frequently represented and culturally dominant online, while erasing the voices of the socially marginalized.

“Generative AI models are more likely to reproduce the voices of those frequently represented and culturally dominant online, while erasing the voices of the socially marginalized.”

May combine two claims.

Generative AI can produce or amplify misinformation, perpetuate representational harm, and create epistemic inequities, particularly in multilingual contexts.

“We illustrate each dimension with real-world examples that reveal how generative AI can produce or amplify misinformation, perpetuate representational harm, and create epistemic inequities, particularly in multilingual contexts.”

May combine two claims.

Generative AI misuse includes confabulating facts and distorting truth.

“confabulating facts and distorting truth (Ji et al. 2023)”

• cites Ji et al. 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Generative AI models trained on vast datasets scraped from the web inherit and reproduce the prejudices and biases embedded within those sources.

“Trained on vast datasets often scraped from the web, these models inherit and reproduce the prejudices and biases embedded within those sources.”

May combine two claims.

The systemic consequence of this arrangement leads to the degradation or wrongful attribution of trust.

“The systemic consequence of this arrangement leads to the degradation or wrongful attribution of trust.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Epistemic injustice emphasizes how identity-based prejudice within an information ecosystem not only unjustly hinders the expression of marginalized groups, but also significantly impairs the knowledge formation capabilities of all individuals.

“Epistemic injustice emphasizes how identity-based prejudice within an information ecosystem not only unjustly hinders the expression of marginalized groups, but also significantly impairs the knowledge formation capabilities of all individuals.”

May combine two claims.

LLMs lack embodied knowledge and cultural history.

“While LLMs may demonstrate forms of human language skills, they lack embodied knowledge and cultural history.”

May combine two claims.

The Senate Intelligence report on interference in the 2016 US Presidential election concluded that Russian information operatives disproportionately targeted African Americans, and that race and related issues were the preferred target of the information warfare campaign.

“the Senate Intelligence report on interference in the 2016 US Presidential election concluded that Russian information operatives disproportionately targeted African Americans, and”

• cites Committee 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Each configuration of generative epistemic injustice has an equivalent counterpart outside of the realm of generative AI.

“however, we note that each configuration has an equivalent counterpart outside of the realm of generative AI.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

When novel social experiences emerge and mainstream cultural narratives fail to grasp them, hermeneutical injustices inevitably arise.

“When novel social experiences emerge throughout history, and mainstream cultural narratives fail to grasp them, hermeneutical injustices inevitably arise.”

Classified as probably not a substantial claim. May combine two claims.

Hermeneutical injustices can arise when the system lacks sufficient sociocultural understanding of humans, or when the system obstructs access to knowledge itself.

“This may occur when the system lacks sufficient sociocultural understanding of humans, or when the system obstructing the access to knowledge itself.”

Probably not a substantial claim: may not pass S4 (based).

Generative AI systems can erase or misportray marginalized groups due to a lack of contextual and cultural understanding.

“we propose the term “generative hermeneutical ignorance” to describe how these systems can erase or misportray marginalized groups due to a lack of contextual and cultural understanding.”

Probably not a substantial claim: may not pass S2 (checkable).

These phenomena constitute hermeneutical injustice in Fricker's sense because they perpetuate a gap in collective interpretive resources, obstructing the understanding of marginalized experiences.

“These phenomena constitutes hermeneutical injustice in Fricker's sense because it perpetuates a gap in collective interpretive resources, obstructing the understanding of these marginalized experiences.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Testimonial injustice can arise when algorithms are prioritized over human credibility, potentially amplifying existing societal biases.

“Testimonial injustice can arise when algorithms are prioritized over human credibility, potentially amplifying existing societal biases.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Generative models lack the nuanced comprehension of human experience necessary for accurate and equitable representation.

“This occurs when generative models, despite their appearance of world knowledge and language understanding, lack the nuanced comprehension of human experience necessary for accurate and equitable representation.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Image generators can produce aesthetically pleasing visuals but may struggle with grounded physical concepts.

“For example, image generators can produce aesthetically pleasing visuals but may struggle with grounded physical concepts.”

Classified as probably not a substantial claim.

The Parkland polemic is a testimonial injustice to the eyewitnesses of the attack and to the parents who lost their children, due to their statements in favor of firearms regulation in the wake of the tragedy.

“This polemic is a testimonial injustice to the eyewitnesses of the attack and to the parents who lost their children, due to their statements in favor of firearms regulation in the wake of the tragedy.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Hateful propaganda foments further prejudice against marginalized groups, stripping them of credibility and leaving them vulnerable to testimonial injustice.

“Hateful propaganda foments further prejudice against marginalized groups, stripping them of credibility and leaving them vulnerable to testimonial injustice.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Hermeneutical injustices can occur when algorithms independently construct meanings and interpretive frameworks, often in automated setups without direct human oversight.

“Additionally, hermeneutical injustices can occur when algorithms independently construct meanings and interpretive frameworks, often in automated setups without direct human oversight.”

May combine two claims.

In reality, there is extensive evidence and eyewitness accounts that Uyghur ethnic minorities have been detained en masse and subjected to forced labor, forced birth control, separation of families, and Islamophobic religious suppression.

“In reality, there is extensive evidence and eyewitness accounts that Uyghur ethnic minorities have been detained en masse and subjected to forced labor, forced birth control, separation of families, and Islamophobic religious suppression (OHCHR 2022).”

• cites OHCHR 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Hermeneutical injustice extends to novel sociotechnical experiences, where human-technology interactions can lead to misunderstandings and misrepresentations of lived experiences.

“This phenomenon also extends to novel sociotechnical experiences, where interactions between humans and new technologies can lead to misunderstandings and misrepresentations of lived experiences.”

Probably not a substantial claim: may not pass S2 (checkable).

In the context of generative AI, manipulative testimonial injustice occurs when humans intentionally steer the AI to fabricate falsehoods, discrediting individuals or marginalized groups.

“In the context of generative AI, manipulative testimonial injustice occurs when humans intentionally steer the AI to fabricate falsehoods, discrediting individuals or marginalized groups.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The quality and veracity of generative AI outputs varies, posing risks of epistemic contamination.

“However, the quality and veracity of these outputs varies, posing risks of epistemic contamination.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

By unifying the epistemic injustices of AI misinformation, bias, representational harms, and power imbalances within the AI industry under a single four-dimensional theoretical framework, we can identify commonalities in their mitigation strategies and build solidarity among affected groups.

“By unifying the epistemic injustices of AI misinformation, bias, representational harms, and power imbalances within the AI industry under a single four-dimensional theoretical framework, we can identify commonalities in their mitigation strategies and build solidarity among affected groups.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Generative algorithmic epistemic injustice is not a speculative theory, but a real danger that has exploded with recent advances and investment in the field.

“Note that generative algorithmic epistemic injustice is not a speculative theory, but a real danger that has exploded with recent advances and investment in the field.”

Classified as probably not a substantial claim. The quote only partly states this claim.

The four configurations of generative epistemic injustice are: amplified testimonial injustice, manipulative testimonial injustice, hermeneutical ignorance, and hermeneutical access injustice.

“Therefore our four configurations of generative epistemic injustice are: 1. Generative amplified testimonial injustice: when generative AI magnifies and produces socially biased viewpoints from its training data.2. Generative manipulative testimonial injustice: when humans fabricate testimonial injustices with generative AI.”

• Also stated in: § 4 Generative Epistemic Injustice · plain

  “3. Generative hermeneutical ignorance: when generative AI lacks the interpretive frameworks to understand human experiences.4. Generative hermeneutical access injustice: when unequal access to information and knowledge is facilitated by generative AI.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Unlike amplified testimonial injustice, which emerges from memorized patterns in data, manipulative testimonial injustice involves deliberate manipulation through techniques like prompting or jailbreaking.

“Unlike amplified testimonial injustice, which emerges from memorized patterns in data, manipulative testimonial injustice involves deliberate manipulation through techniques like prompting or jailbreaking.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

While classification AI is set to delineate true categories and false ones, generative AI offers statements that play the role of testimonies, explanations, and interpretations.

“While classification AI is set to delineate true categories and false ones, generative AI offers statements that play the role of testimonies, explanations, and interpretations.”

May combine two claims. The quote only partly states this claim.

The generative AI's sycophantic fulfilment of the request to spread misinformation reflects how testimonial injustices are memorized and the potential for their amplification by generative models.

“Although the NewsGuard study was a simulation of misinformation, rather than an”

Probably not a substantial claim: may not pass S3 (load-bearing).

A 2023 report found that 98% of deepfake videos online were porn.

“with a 2023 report finding that 98% of deepfake videos online were porn (Heroes 2023).”

• cites Heroes 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

When algorithms make decisions in bureaucratic systems such as governments, businesses, or healthcare providers, they exert power via their contributions to knowledge.

“When algorithms make decisions, particularly in our bureaucratic systems such as governments, businesses, or healthcare providers, they exert power via their contributions to knowledge.”

May combine two claims.

Generative AI can either amplify testimonial injustices due to biases acquired in pretraining and finetuning processes, or it can be manipulated by human users to create harmful content.

“Generative AI can either amplify testimonial injustices due to biases acquired in the pretraining and finetuning processes, or it can be manipulated by human users to create harmful content.”

May combine two claims.

Systems for misinformation detection are vulnerable to epistemic injustices and may automate discrediting of marginalized people if epistemic virtues are not exercised in their creation.

“Furthermore, systems for misinformation detection are also vulnerable to epistemic injustices and may automate the discrediting of the marginalized (Neumann, De-Arteaga, and Fazelpour 2022) if epistemic virtues are not exercised in their creation.”

• cites Neumann, De-Arteaga, and Fazelpour 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

AI conversations can be a tool for bolstering the confidence and resources of those who might be isolated from their communities.

“Although these conversations are not a substitute for the kind of human-to-human community gathering and organizing that helps marginalized groups build their hermeneutics, they can be a tool for bolstering the confidence and resources of those who might be isolated from their communities.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Hermeneutical access injustice can lead to echo chambers and epistemic fragmentation, isolating identity groups from each other on an informational level.

“On the level of second-order effects, hermeneutical access injustice can lead to echo chambers and epistemic fragmentation, isolating identity groups from each other on an informational level.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The Midjourney images are evidence of the hermeneutical ignorance of Midjourney.

“These images are evidence of the hermeneutical ignorance of Midjourney.”

Probably not a substantial claim: may not pass S2 (checkable).

AI can be directed to expand our cultural understanding of each other and ourselves.

“Although we have so far emphasized their hermeneutical injustice, we can alternatively direct these systems to expanding our cultural understanding of each other and ourselves.”

Probably not a substantial claim: may not pass S2 (checkable).

In generative AI, hermeneutical access injustice centers on generative AI's control over access to information, leading to denial of knowledge based on identity-driven bias or misrecognition.

“In the context of generative AI, this injustice takes another specific form: it centers on the generative AI's control over access to information, leading to a denial of knowledge based on identity-driven bias or misrecognition.”

Probably not a substantial claim: may not pass S4 (based).

Generative hermeneutical ignorance stems directly from the limitations of generative AI models themselves.

“generative hermeneutical ignorance is unique in that it stems directly from the limitations of generative AI models themselves.”

Probably not a substantial claim: may not pass S4 (based).

Simulating the experiences of others can build empathy across identity lines.

“Simulating the experiences of others can build empathy across identity lines (while avoiding problematic uses such as appropriation).”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

If we remain ignorant about the inner workings of AI, affected groups will have no recourse to contesting the reasoning behind unfair decisions or generations.

“If we remain ignorant about the inner workings of AI, affected groups will have no recourse to contesting the reasoning behind unfair decisions or generations.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Epistemic injustices are human-computer interactions.

“Epistemic injustices are human-computer interactions”

May combine two claims.

Generative hermeneutical ignorance diverges from traditional forms of hermeneutical injustice and from those perpetuated by other algorithmic systems.

“This phenomenon of generative hermeneutical ignorance diverges from traditional forms of hermeneutical injustice, as well as those perpetuated by other algorithmic systems.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The challenges posed by generative AI are rooted in the inherent limitations of current technology.

“the challenges posed by generative AI are rooted in the inherent limitations of current technology.”

Probably not a substantial claim: may not pass S4 (based).

If generative models become a common epistemic tool, they will shape the structure of our collective body of knowledge.

“If generative models become a common epistemic tool-if we treat them like web search or encyclopedias, or like infinite firehoses of information-they will shape the structure of our collective body of knowledge.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

In armed conflict, when the attacking side downplays fatality rates, they are most likely trying to deflate the credibility of critics of this violence, motivated by political interests and prejudice against the harmed group.

“when the attacking side downplays fatality rates, particularly civilian fatalities, they are most likely trying to deflate the credibility of critics of this violence, who may be members of the targeted group, or third parties who simply oppose acts of war.”

• AssumptionRests on an assumption: In armed conflict, the attacking side downplays fatality rates to deflate the credibility of critics of the violence.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Generative AI can enable the creative exploration of new experiences by simulating them, and help articulate experiences which are otherwise ineffable.

“Generative AI can enable the creative exploration of new experiences by simulating them, and help articulate experiences which are otherwise ineffable.”

May combine two claims.

Systemic effects of generative AI can hamper information flow and erode trust between disparate groups.

“The resulting sys-temic effects can further hamper the flow of information and erode trust between disparate groups.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The watermark merely tells us the origin of content but does not shore up our tools for critically assigning credibility and trust.

“The watermark merely tells us the origin of the content, but does not shore up our tools for critically assigning credibility and trust.”

May combine two claims.

Willful hermeneutical ignorance leads to complacency and reinforces hermeneutical oppression.

“This willful hermeneutical ignorance-the continued misunderstanding or misinterpretation of marginalized experiences despite their articulation (Pohlhaus 2012)-leads to complacency and reinforces hermeneutical oppression.”

• cites Pohlhaus 2012 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

LLMs are Englishcentric and have variable quality across languages, particularly under-resourced languages.

“LLMs are notoriously Englishcentric and have variable quality across languages, particularly so-called “under-resourced” languages.”

Probably not a substantial claim: may not pass S2 (checkable).

The outputs or results of AI systems for eliciting or detecting testimonial injustice must be evaluated by humans, particularly members of the affected group.

“Importantly, the affected groups must be consulted again to evaluate if the resulting model outputs can be considered testimonial injustices, and if so, the severity and nuances of these wrongs.”

• Also stated in: § Generative AI for Epistemic Justice · hedged

  “The results of these methods would again need to be evaluated by humans, particularly those of the affected group.”

• AssumptionRests on an assumption: Affected groups can reliably identify and evaluate testimonial injustices and their severity.

Probably not a substantial claim: may not pass S2 (checkable).

A study of epistemic injustice must include discourse with the humans who provide the basis for the meaning of the information generated and spread by machines.

“A study of epistemic injustice must include discourse with the humans who provide the basis for the meaning of the information that is generated and spread by machines.”

Probably not a substantial claim: may not pass S2 (checkable).

Image generation can be used to re-imagine and express oneself.

“Image generation can be used to re-imagine and express oneself.”

Classified as probably not a substantial claim.

AI conversations are not a substitute for human-to-human community gathering and organizing that helps marginalized groups build their hermeneutics.

“Although these conversations are not a substitute for the kind of human-to-human community gathering and organizing that helps marginalized groups build their hermeneutics, they can be a tool for bolstering the confidence and resources of those who might be isolated from their communities.”

Classified as probably not a substantial claim.

AI can be an interactive tool for exploring one's own experiences.

“AI can be an interactive tool for exploring one's own experiences.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Testimonial smothering occurs when a speaker is so inured to being silenced or misunderstood that they hold back their testimony entirely, hindering others from accessing their truth.

“Another relevant concept is testimonial smothering when a speaker is so inured to being silenced or misunderstood that they hold back their testimony entirely, which hinders everyone else from accessing their truth (Dotson 2011).”

• cites Dotson 2011 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Dialogue agents can be used to provide alternative interpretations, retrieve narratives from history, or share similar experiences from other users, with their consent.

“Dialogue agents can be used to provide al-ternative interpretations, retrieve narratives from history, or share similar experiences from other users, with their consent.”

Probably not a substantial claim: may not pass S4 (based).

Any algorithm that tries to automatically assign credibility is at risk of committing testimonial injustice.

“Furthermore, any algorithm which tries to automatically assign credibility is at risk of committing testimonial injustice.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Memorization and synthesis are the key processes by which generative models formulate outputs, which can reproduce unjust testimonies and formulate new hermeneutics.

“Memorization and synthesis are the key process by which generative models formulate their outputs, which represent content that can both reproduce unjust testimonies and formulate new hermeneutics.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The utility of a system for detecting testimonial injustice in recorded data is identifying the contexts in which testimonial injustice occurs, who perpetuates it, and the prevalence relative to opposing narratives, if those are also measured.

“Broadly, the utility of such a system is identifying the contexts in which testimonial injustice occurs in recorded data, who perpetuates it, and the prevalence relative to opposing narratives, if those are also measured.”

May combine two claims.

Generative AI models have the potential to create entirely new forms of misinterpretation, unlike classification algorithms that perpetuate biases present in training data.

“While these systems can perpetuate biases present in their training data, generative AI models have the potential to create entirely new forms of misinterpretation and misinterpretation.”

Probably not a substantial claim: may not pass S2 (checkable).

Generative hermeneutical access injustice is a distinct form of hermeneutical injustice within generative AI.

“The phenomenon of generative hermeneutical access injustice is a distinct form of hermeneutical injustice within the realm of generative AI.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The common body of knowledge can be thought of as an epistemic ecosystem vulnerable to pollution by false narratives.

“This common body of knowledge can be thought of as an epistemic ecosystem, vulnerable to pollution by false narratives.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Generative AI outputs have persuasive and performative potential, enabling AI to participate in defining semantic concepts and shaping individual and collective knowledge.

“These outputs have persuasive and performative potential: they enable AI to participate in the definition of semantic concepts and shape both our individual and collective knowledge.”

Probably not a substantial claim: may not pass S4 (based).

Generative AI models lack the embodied and cultural knowledge that humans acquire through lived experiences.

“Generative AI models lack the embodied and cultural knowledge that humans acquire through lived experiences.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The model has misrecognized these groups and through doing so erased a part of their cultural experience, indeed of any non-American culture.

“The model has misrecognized these groups, and through doing so erased a part of their cultural experienceindeed, of any non-American culture.”

Probably not a substantial claim: may not pass S2 (checkable).

Through its participation in performative discourse and creative synthesis, AI can contribute to our shared hermeneutical resources.

“Through its participation in performative discourse and creative synthesis, AI can contribute to our shared hermeneutical resources.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Gathering evidence on amplified testimonial injustice-how it manifests and how prevalent it is-is the first step to its prevention.

“Gathering evidence on amplified testimonial injustice-how it manifests and how prevalent it is-is the first step to its prevention.”

• AssumptionRests on an assumption: Popular narratives that discredit or undermine affected identity groups circulate on social media and the Internet.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Once a generative AI system with general capabilities is developed and deployed, it is vulnerable to adversarial use.

“Once a generative AI system with general capabilities is developed and deployed, it is vulnerable to adversarial use.”

Probably not a substantial claim: may not pass S2 (checkable).

Input from affected groups is crucial to both design criteria for identifying testimonial injustices and measure the impact of these injustices.

“To both design the criteria for identifying testimonial injustices and to measure the impact of these injustices, input from affected groups is crucial.”

• AssumptionRests on an assumption: Affected groups can reliably identify and evaluate testimonial injustices and their severity.

May combine two claims.

This exacerbates conventional hermeneutical injustice by widening the information gap between identity groups and detracting from shared understanding and access to knowledge about marginalized experiences.

“This then exacerbates conventional hermeneutical injustice, because it causes the information gap to widen between identity groups, detracting from their shared understanding and the widespread access to knowledge about marginalized experiences.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Generative AI systems can be re-engineered to surface injustices and enrich our cultural resources.

“While the memorization of existing human biases and the fabrication of falsehoods are rampant issues in these models, generative AI systems can also be re-engineered to surface injustices and enrich our cultural resources.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The alien nature of these models means they are less likely to follow our preexisting hermeneutics, for better or worse.

“The”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Memorization of existing human biases and fabrication of falsehoods are rampant issues in generative AI models.

“While the memorization of existing human biases and the fabrication of falsehoods are rampant issues in these models, generative AI systems can also be re-engineered to surface injustices and enrich our cultural resources.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The ideal and its realization for fine-tuning is produced by AI's dominant groups, which restricts the diversity and depth of experiences the model can portray.

“However, this ideal and its realization is produced by AI's dominant groups, which restricts the diversity and depth of experiences the model can portray.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Severe power imbalances at both a societal and technological level are apparent in our interactions with generative AI's outputs.

“Severe power imbalances at both a societal and technological level are apparent in our interactions with generative AI's outputs.”