Large Language Models Threaten Language’s Epistemic and Communicative Foundations

Shashank SrivastavaEMNLP 2025

TL;DR

Large Language Models (LLMs) are fundamentally altering human communication by decoupling language from genuine thought, leading to semantic drift and eroding accountability. This paper proposes novel concepts like "authorship entropy" and "epistemic doppelgängers" to understand these risks and suggests technical and educational solutions to preserve the epistemic and communicative foundations of language.

What Problem Does This Paper Address?

For centuries, the bedrock of written language has been the assumption that behind every text lies a human mind, implying intention, accountability, and a unique authorial voice. The rapid advancement of LLMs, capable of generating highly coherent and stylistically nuanced text, challenges this fundamental covenant. This paper addresses the systemic risk LLMs pose to language's epistemic foundations, arguing that they can lead to semantic drift (language meaning changing over time due to AI influence), model collapse (LLMs becoming less diverse and accurate as they train on their own outputs), and a general erosion of trust and accountability in written communication. The authors aim to understand and mitigate these threats to preserve the value and integrity of human language.

Methods

The paper introduces several conceptual frameworks and metrics:

  • Hybrid Authorship Graphs: These model relationships between human users, LLMs, and texts, creating a provenance network to track generation and influence.
  • Epistemic Doppelgängers: This term describes LLM-generated texts that are indistinguishable from human-authored texts, making it difficult to discern genuine human thought.
  • Authorship Entropy: A proposed metric to quantify the uncertainty of text authorship in an AI-saturated ecosystem, using information-theoretic principles.
  • Proof-of-Interaction (PoI): A mechanism to ensure substantial human involvement in text creation, potentially through verifiable credentials or logged interactions.
  • Semantic Drift Metrics: Tools to track changes in language usage and meaning over time, comparing human-authored and LLM-generated text distributions.

Datasets

Not discussed in the paper. The paper focuses on conceptual frameworks and theoretical implications rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • The "Covenant" of Language is Broken: The paper argues that the long-standing assumption of a human mind behind every text is being fundamentally challenged by LLMs, leading to a crisis of trust and accountability.
  • Semantic Drift and Model Collapse are Systemic Risks: LLMs can not only alter language meaning but also degrade their own capabilities by training on synthetic data, creating a feedback loop that threatens linguistic diversity and clarity.
  • Epistemic Doppelgängers Undermine Trust: The ability of LLMs to perfectly mimic human writing creates a scenario where even experts can be fooled, making it harder to trust the authenticity of any text.
  • Authorship Entropy as a New Metric: Quantifying the uncertainty of authorship is crucial for understanding the impact of LLMs on trust and accountability in the digital age.
  • Preservation, Not Prevention: The authors suggest that preventing LLM-generated text is impossible and undesirable; instead, the focus should be on preserving the essential functions of human language through new mechanisms and norms.

Key Results

The paper does not present quantitative results from experiments. Instead, it offers qualitative insights and conceptual frameworks. It demonstrates through a conceptual plot (Figure 3) how semantic drift can occur through iterative fine-tuning of a language model on its own generated text, showing an increasing divergence in language distribution.

Limitations and Future Work

  • Speculative Constructs: The paper acknowledges that concepts like epistemic doppelgängers and authorship entropy are speculative and require empirical grounding.
  • Practical Implementation Challenges: Proposals like proof-of-interaction logs and human-only publishing spaces are noted as challenging to implement, requiring infrastructure, cooperation, and cultural shifts.
  • Balancing Access and Integrity: The authors recognize a tension between promoting LLM accessibility and discouraging gatekeeping, aiming to protect epistemic integrity without devolving into purity tests.
  • Future Work: The paper calls for empirical testing of proposed metrics, development of technical standards (like verifiable credentials), and new forms of AI literacy to navigate a world where much text may not be of human origin. It also suggests exploring human-AI collaboration frameworks and governance models.

Summary updated

BibTeX
@inproceedings{srivastava2025large,
  author = {Shashank Srivastava},
  title = {{Large Language Models Threaten Language’s Epistemic and Communicative Foundations}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure