TL;DR
This paper argues that current US legal frameworks are insufficient to address the unique harms posed by generative AI, particularly concerning fundamental human values like autonomy and privacy. It proposes a "Responsible AI Legal Framework" that adapts existing laws and introduces new approaches to safeguard against these evolving risks.
What Problem Does This Paper Address?
Generative AI, especially Large Language Models (LLMs), can produce human-like creative content but also exhibit pernicious effects. These include generating biased content, making false claims (hallucinations), and spreading misinformation, all of which can undermine democratic principles and individual well-being. While technical "AI alignment" efforts aim to mitigate these harms, they are still nascent and face challenges like prompt injection attacks. Furthermore, even with perfect alignment, the question of which values AI should align with and who decides remains critical. Corporations' profit motives may not align with societal needs, leading to AI systems shaped by narrow values. This paper investigates the effectiveness of existing US law in addressing these complex challenges and proposes a new legal framework.
Methods
The authors employ an interdisciplinary approach combining legal analysis with hypothetical scenario generation. They collaborated with experts from computer security, machine learning, NLP, and law to develop five distinct scenarios illustrating potential harms from generative AI. These scenarios were then analyzed through the lens of existing US legal frameworks, including constitutional law, civil rights law, defamation, and product liability. The paper also draws on historical legal scholarship and comparative legal analysis of regulatory approaches in the US, EU, and Asia.
Datasets
The paper mentions that "The datasets, which include input from an expert workshop and AI-harm scenarios, are publicly available on GitHub at https://github.com/inyoungcheong/LLM." No specific datasets for training or evaluating AI models are detailed, as the focus is on legal analysis and scenario development.
Key Insights and Ideas
- Current US Laws Fall Short: Existing legal frameworks struggle to hold AI companies accountable for AI-assisted discriminatory outputs, especially when malicious intent is not clear or tangible individual harms are difficult to prove due to the intricate and opaque nature of AI systems.
- The "State Action" Doctrine is a Hurdle: For AI developed by private entities, claims based on constitutional rights like the First Amendment (free speech) or Fourteenth Amendment (equal protection) are difficult to pursue due to the "state action" doctrine, which typically limits these rights to government actions.
- Defamation and Product Liability Have Limitations: Defamation claims are challenging because they usually require targeting specific individuals, and generative AI outputs often target broader groups. Product liability claims are more applicable to tangible harms, but proving a direct causal link to AI defects can be complex, especially for intangible harms like mental health impacts.
- Section 230 Immunity is Debatable for Generative AI: The authors argue that generative AI systems, which actively "complete sentences" and produce novel content rather than merely displaying user-generated content, may not qualify for Section 230 immunity, potentially opening companies to a wider range of liability.
- A Proactive, Balanced Approach is Needed: The US legal system's historical preference for reactive, adversarial approaches needs to be complemented by more proactive and balanced strategies to address the unique and potentially permanent harms of generative AI.
- Human Values Need to Become Legal Rights: The paper advocates for a shift from "negative rights" (protecting individuals from government intrusion) to "positive rights" that ensure equitable access to AI technologies and protect against harms like manipulation and bias.
Key Results
The paper's primary "result" is the identification of significant gaps and ambiguities in current US laws when applied to generative AI. Through the analysis of five hypothetical scenarios, the authors demonstrate that:
- Educational disparity: Students in poorer districts suing for unequal access to AI educational tools would likely face an uphill battle, mirroring past Supreme Court decisions that do not recognize education as a fundamental right requiring "precisely equal advantages."
- Manipulation/discrimination: LGBTQIA+ individuals negatively impacted by an AI application fostering anti-LGBTQIA+ sentiment would struggle to claim constitutional rights due to the private nature of the AI, and defamation claims would be difficult due to the broad targeting. Product liability might be a possibility, but proving direct causation from the AI's bias would be complex.
- Polarization and external threats: Individuals targeted by defamatory outputs from an AI community tool could potentially sue for defamation, but the company would need to demonstrate reasonable measures to prevent harm. Privacy infringement claims might arise from data collection without consent.
- Over-reliance/sexual abuse: Claims related to self-harm due to over-reliance on AI replicas could potentially fall under product liability. Privacy rights could be infringed if sensitive data is collected without consent, and claims for extreme emotional distress might be viable for non-consensual virtual sexual abuse.
Limitations and Future Work
The authors acknowledge that human bias and subjectivity are inherent limitations in any legal analysis. They also note that the applicability of Section 230 to generative AI is an ongoing debate with no clear precedents.
Future work suggested by the paper includes:
- Developing a "Responsible AI Legal Framework": This framework would enshrine fundamental values in legal structures, establish comprehensive safety guidelines, and implement tailored liability models for human-AI interactions.
- Creating new legal rights: This involves translating human values into legally recognized rights, potentially through statutory rights that go beyond existing constitutional protections.
- Re-evaluating the "state action" doctrine: Given the increasing power of non-governmental actors like large corporations, the doctrine needs re-evaluation to address how private entities can curtail individual freedom.
- Implementing comprehensive safety regulations: This includes mandatory third-party audits, industry-wide norms for transparency, and proactive oversight mechanisms.
- Establishing a new liability regime: This could involve strict liability frameworks and a "disparate impact" theory to address harms caused by AI systems, especially those with discriminatory effects.
Summary updated