Regulation and NLP (RegNLP): Taming Large Language Models

Catalina Goanta, Nikolaos Aletras, Ilias Chalkidis, Sofia Ranchordás, Gerasimos Spanakis

Abstract

The scientific innovation in Natural Language Processing (NLP) and more broadly in artificial intelligence (AI) is at its fastest pace to date. As large language models (LLMs) unleash a new era of automation, important debates emerge regarding the benefits and risks of their development, deployment and use. Currently, these debates have been dominated by often polarized narratives mainly led by the AI Safety and AI Ethics movements. This polarization, often amplified by social media, is swaying political agendas on AI regulation and governance and posing issues of regulatory capture. Capture occurs when the regulator advances the interests of the industry it is supposed to regulate, or of special interest groups rather than pursuing the general public interest. Meanwhile in NLP research, attention has been increasingly paid to the discussion of regulating risks and harms. This often happens without systematic methodologies or sufficient rooting in the disciplines that inspire an extended scope of NLP research, jeopardizing the scientific integrity of these endeavors. Regulation studies are a rich source of knowledge on how to systematically deal with risk and uncertainty, as well as with scientific evidence, to evaluate and compare regulatory options. This resource has largely remained untapped so far. In this paper, we argue how NLP research on these topics can benefit from proximity to regulatory studies and adjacent fields. We do so by discussing basic tenets of regulation, and risk and uncertainty, and by highlighting the shortcomings of current NLP discussions dealing with risk assessment. Finally, we advocate for the development of a new multidisciplinary research space on regulation and NLP (RegNLP), focused on connecting scientific knowledge to regulatory processes based on systematic methodologies.

Introduction: The Pacing Gap

As Large Language Models (LLMs) evolve at breakneck speed, a significant "pacing gap" has emerged: the speed of technical innovation is exponentially outstripping the development of legal and regulatory oversight. While the NLP community has become increasingly concerned with the societal impacts of these models, the current discourse often lacks the systematic methodology required to influence formal policy effectively. This paper proposes the establishment of RegNLP, a multidisciplinary research field designed to bridge the gap between technical NLP research and formal regulation studies.

The Problem: Polarized Narratives and Regulatory Capture

The current debate surrounding AI regulation is largely fractured into two polarized camps:

  1. AI Safety: Often led by industry leaders, this movement focuses on existential, long-term "doomsday" risks.
  2. AI Ethics: Primarily led by civil society, this movement focuses on immediate harms such as algorithmic bias, fairness, and transparency.

This polarization is frequently amplified by social media, creating a vacuum where "science influencers" exert more influence over policy than the broader scientific community. This environment facilitates regulatory capture, a phenomenon where regulators begin to serve the interests of the industry or specific vocal interest groups rather than the general public. Without a rigorous, evidence-based framework, there is a risk that AI policy will be shaped by "moral panics"—similar to the early public reactions to IVF or GMOs—rather than quantified risk assessments.

Theoretical Foundations for RegNLP

To move beyond anecdotal or reactive policy-making, RegNLP draws on established theories from law, economics, and political science:

  • The Collingridge Dilemma: This principle highlights a double-bind in technology regulation: it is difficult to regulate a technology until it is widely adopted, but by the time it is adopted, it is often too "entrenched" to be easily changed. RegNLP seeks to find the optimal window for intervention.
  • Risk vs. Uncertainty: The framework distinguishes between risk (where the probabilities of outcomes are known) and uncertainty (where variables are unknown). Current NLP research often conflates the two, leading to imprecise policy recommendations.
  • Relativization of Risk: Rather than treating LLM-mediated issues as unprecedented, RegNLP advocates for comparing them to existing societal risks. For instance, a lawyer using a hallucinating LLM is a modern iteration of professional negligence, a concept already well-handled by existing legal frameworks.

A Roadmap for the RegNLP Field

The authors define three essential pillars for the development of this new research space:

  1. Multidisciplinarity: Breaking the silos between Computer Science departments and schools of Law and Economics. Effective regulation requires technical benchmarks that translate into meaningful legal standards.
  2. Harmonized Methodologies: The field must move toward standardized, rigorous ways to measure LLM risks. Without harmonized auditing methods, policymakers cannot accurately compare results across different studies.
  3. Formal Science Participation: Establishing formal roles for researchers, such as "trusted flaggers" within frameworks like the EU's Digital Services Act, to ensure that scientific evidence—rather than social media sentiment—drives the regulatory process.

Conclusion: From Moral Panic to Mature Regulation

The history of Genetically Modified Organisms (GMOs) serves as a cautionary tale. Initial moral panics led to restrictive environments, but through systematic regulation and scientific advancement, the technology eventually became a mundane yet vital tool for food security. LLMs require a similar path to maturity. By treating regulation as a core component of NLP research rather than an afterthought, the RegNLP community can provide the technical evidence needed to transition from "soft," unenforceable codes of conduct to robust, evidence-based "hard" law.

Summary updated

Abstract

The scientific innovation in Natural Language Processing (NLP) and more broadly in artificial intelligence (AI) is at its fastest pace to date. As large language models (LLMs) unleash a new era of automation, important debates emerge regarding the benefits and risks of their development, deployment and use. Currently, these debates have been dominated by often polarized narratives mainly led by the AI Safety and AI Ethics movements. This polarization, often amplified by social media, is swaying political agendas on AI regulation and governance and posing issues of regulatory capture. Capture occurs when the regulator advances the interests of the industry it is supposed to regulate, or of special interest groups rather than pursuing the general public interest. Meanwhile in NLP research, attention has been increasingly paid to the discussion of regulating risks and harms. This often happens without systematic methodologies or sufficient rooting in the disciplines that inspire an extended scope of NLP research, jeopardizing the scientific integrity of these endeavors. Regulation studies are a rich source of knowledge on how to systematically deal with risk and uncertainty, as well as with scientific evidence, to evaluate and compare regulatory options. This resource has largely remained untapped so far. In this paper, we argue how NLP research on these topics can benefit from proximity to regulatory studies and adjacent fields. We do so by discussing basic tenets of regulation, and risk and uncertainty, and by highlighting the shortcomings of current NLP discussions dealing with risk assessment. Finally, we advocate for the development of a new multidisciplinary research space on regulation and NLP (RegNLP), focused on connecting scientific knowledge to regulatory processes based on systematic methodologies.
BibTeX
@inproceedings{goanta2023regulation,
  author = {Catalina Goanta and Nikolaos Aletras and Ilias Chalkidis and Sofia Ranchordás and Gerasimos Spanakis},
  title = {{Regulation and NLP (RegNLP): Taming Large Language Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=Y93S9kxFZl},
}
Figure