AI evidence extraction

Epistemic Injustice in Generative AI: Probabilistic Generation, Trust Erosion, and the Structural Conditions of Algorithmic Knowledge Harm

Shenghui Bao · Open MIND · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.17613/gtkt3-3t116

Methodology & findings

Study design

Philosophical analysis combined with illustrative case analysis.

Main result

The paper identifies three interrelated mechanisms through which LLMs generate structural conditions for epistemic injustice: "Automation-Induced Testimonial Injustice (AITI), in which algorithmically mediated distortion of the epistemic field produces credibility deflation of divergent human testimony; Interpretive Erasure, through which under-represented epistemic frameworks are systematically marginalized as LLMs function as dominant infrastructure; and Epistemic Debt, whereby the systematic gap between AI-assisted epistemic outputs and independently achievable competencies generates structural fragility." The authors emphasize that "these mechanisms constitute a mutually reinforcing feedback system, such that their compound effect is substantially more intractable than any individual mechanism considered independently."

Reports effect sizes.

Research paradigm

Critical epistemology; philosophical analysis of sociotechnical systems

Author conclusions

The authors conclude that "large language models, when deployed as epistemic infrastructure, generate structural conditions for epistemic injustice through three interrelated mechanisms: automation-induced testimonial injustice, interpretive erasure, and epistemic debt accumulation. These mechanisms are not contingent features of current LLM technology amenable to resolution through incremental technical refinement; they are structural properties of probabilistic language generation deployed at epistemic scale." They further argue that "Alignment techniques including RLHF and DPO, rather than resolving this structural problem, may in certain respects deepen it: by optimising for outputs that are confident, fluent, and evaluator-preferred, they risk amplifying the credibility-conferring surface form that drives the AITI mechanism."

Risk of bias

Selection bias in case exemplars (purposively selected rather than systematically sampled); Limited direct empirical evidence for generative LLM behavior in medical and historical domains; Reliance on secondary literature and documented findings rather than primary empirical investigation; Potential confirmation bias in theoretical framework development (building framework then illustrating with cases); Case selection bias: Cases selected through purposive (non-random) sampling for analytical illustration rather than representativeness; Confirmation bias: Illustrative cases selected to demonstrate mechanisms rather than test them; Limited empirical grounding: Medical and historical cases supplement primary material with literature rather than primary data; Philosophical framework dependency: Analysis depends on acceptance of Fricker's epistemic injustice framework and extensions; Case selection bias: Cases selected as paradigmatic illustrations rather than representative samples; Limited empirical grounding: Medical and historical cases supplement primary material with literature findings rather than direct measurement; Confirmation bias risk: Analysis focused on theoretical consistency rather than falsification attempts; Interpretive bias: Philosophical analysis may reflect author's frameworks and assumptions; Incomplete empirical validation: Interpretive erasure in clinical/historical contexts described as 'nascent' rather than established

Limitations

The authors explicitly acknowledge: "The argument is primarily philosophical and proceeds through illustrative case analysis rather than systematic empirical investigation." They identify two critical limitations: "First, quantifying epistemic debt accumulation at the population level faces substantial methodological obstacles: LLM systems are partially opaque, and their long-term effects on users' epistemic competencies are difficult to disentangle from the effects of other co-occurring social and technological changes
Second, the empirical documentation of interpretive erasure specifically attributable to generative LLMs in clinical and historical contexts remains nascent
the cases examined in Section V are analytical exemplars of structural risks rather than fully established empirical facts."

Open questions raised

Validated longitudinal metrics for epistemic debt accumulation
Systematic cross-cultural audits of hermeneutical adequacy in LLMs deployed in education and healthcare
Comparative analysis of epistemic injustice mechanism interactions across different governance regimes and low-resource epistemic contexts
Empirical documentation of interpretive erasure specifically attributable to generative LLMs in clinical and historical contexts
The authors identify a research agenda comprising: validated longitudinal metrics for epistemic debt accumulation, systematic cross-cultural audits of hermeneutical adequacy in LLMs deployed in education and healthcare, and comparative analysis of epistemic injustice mechanism interactions across different governance regimes and low-resource epistemic contexts.

Data: Database maintained by researcher Damien Charlotin documenting over 230 distinct legal proceedings involving AI-generated hallucinations submitted to courts (as of mid-2025)Extracted from: pdfAgreement 72%

Explore related topics

Epistemic Risk Research Integrity Conceptual