Skip to main content

Practical Aspects of Regulating Large Language Models: A Reading List

Some background papers about how policies of various sorts might be enforced in the context of LLMs...

The Big Picture on Regulating LLMs

Identifying and Tracking LLM output

Identifying and Tracking Diffusion Output

  • "Tree-Ring Watermarks: Fingerprints for Diffusion Images that are Invisible and Robust" (01 Jun 2023)
    https://arxiv.org/abs/2305.20030      
    Note: This steganographic technique, specific to diffusion models, has been cited as a possible way to track the sources of offensive output, including virtual child pornography.

Generative AI and Virtual Child Pornography

  • "AI-generated child sex images spawn new nightmare for the web" (19 Jun 2023)
    https://www.washingtonpost.com/technology/2023/06/19/artificial-intelligence-child-sex-abuse-images/
    Note: Child-safety experts said many of the AI-generated images being shared appeared to have relied on open-source tools, such as Stable Diffusion, which is available as open source and can be run in an unrestricted and unpoliced way.
  • "Generative ML and CSAM: Implications and Mitigations" (24 Jun 2023)
    https://purl.stanford.edu/jv206yg3793
    Note: In this new paper Stanford Internet Observatory and Thorn (a nonprofit that fights the spread of child sexual abuse online) researchers found that, since August 2022, there has been a small but meaningful uptick in the amount of photorealistic A.I.-generated child sexual abuse material circulating on the dark web. The researchers found that the the majority of these images were generated not by Dall-E but by open-source tools that were developed and released with few protections in place.

Evaluating Safety, Trustworthiness and Fairness in Generative AI

Detecting Hate Speech

  • "Towards Legally Enforceable Hate Speech Detection for Public Forums"  (23 May 2023 )      
    https://arxiv.org/abs/2305.13677        
    Note: This paper is potentially of high interest w.r.t. the application of KGs as authoritative references for policy enforcement; in their abstract the authors claim, "Our work introduces a new task for enforceable hate speech detection centred around legal definitions, and a dataset annotated on violations of eleven possible definitions by legal experts"
  • "ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection"  (May 2022)      
    https://bit.ly/3NfPXWk  Discussion: https://bit.ly/42pcHaY

Detecting & Mitigating Bias

Building Biases into AI

  • "Why we need biased AI: How including cognitive and ethical machine biases can enhance AI systems" (18 Mar 2022)       
    https://arxiv.org/abs/2203.09911        
    Note: This was the focus of a recent TWC Lunchtime Reading Group meeting.  This paper is provocative, as its abstract implies: "...this paper is the first tentative step to explicitly pursue the idea of a re-evaluation of the ethical significance of machine biases, as well as putting the idea forth to implement cognitive biases into machines."

LLM-assisted Copyright Infringement Detection

  • "Playing with machines: Using machine learning to understand automated copyright enforcement at scale" (22 Feb 2023)        
    https://journals.sagepub.com/doi/pdf/10.1177/2053951720919963 
    Note: The authors train a ML classifier to identify videos in categories that reflect ongoing controversies in copyright takedowns, and use it to explore how copyright is enforced on YouTube. 

Generative AI and Copyright Law

  • "Generative AI Meets Copyright - Pamela Samuelson" (June 2023)       
    https://www.youtube.com/watch?v=6sDGIrVO6mo        
    Note: Pam Samuelson is the Richard M. Sherman Distinguished Professor of Law at UC Berkeley. In this piece she talks quite a bit about fair use and generative AI, and also suggests that copyright law may be the "Achille's Heel" of generative AI. 
  • "Generative Artificial Intelligence and Copyright Law" (Updated May 11, 2023)       
    https://crsreports.congress.gov/product/pdf/LSB/LSB10922        
    Note: Prepared by the Congressional Research Service, a group of nonpartisan shared staffers        
    to congressional committees and Members of Congress
  • "Copyright Registration Guidance: Works Containing Material Generated by AI" (16 March 2023)       
    https://www.copyright.gov/ai/ai_policy_guidance.pdf      
    Note: This Guidance from the US Copyright Office advises that the only parts of a work that are copyrightable are the human-contributed ones, and the work is not copyrightable if an AI technology determines the expressive elements of the work and the creativity is not the product of human authorship. 
  • "Authors file a lawsuit against OpenAI for unlawfully ‘ingesting’ their books" (05 Jul 2023)       
    https://www.theguardian.com/books/2023/jul/05/authors-file-a-lawsuit-against-openai-for-unlawfully-ingesting-their-books      
    Note: The plaintiffs claim their books were unlawfully “ingested” and “used to train” ChatGPT because the chatbot generated “very accurate summaries” of the novels. 
  • Google & DeepMind are sued for their LLMs, including Bard, Lamda, Palm, and the to-be-released Gemini.  (File 11 Jul 2023)    
    https://storage.courtlistener.com/recap/gov.uscourts.cand.415223/gov.uscourts.cand.415223.1.0.pdf     
    Note: This lawsuit cites a number of issues including copyright. It points to C4 as a problematic training dataset among others. More on C4 here...

Generative AI and Creative Commons Licensing

  • "Why the Great AI Backlash Came for a Tiny Startup You’ve Probably Never Heard Of" (WIRED 14 Aug 2023)    
    https://www.wired.com/story/prosecraft-backlash-writers-ai/   
    Note: Raises the point that having terms of use in a CC license that exclude automated scrapers (such as LLM training code) might not be valid, since non-humans cannot be licensees. 

LLMs and Scientific Communication

LLMs and Education

  • "Educator considerations for ChatGPT"  (2023)      
    https://platform.openai.com/docs/chatgpt-education        
    Note: OpenAI's overview prepared for educators seeking to learn more about the capabilities, limitations, and considerations for using ChatGPT for teaching and learning 

LLMs and Legal Practice

Jailbreaking Generative AI-based Systems

  • "The Amateurs Jailbreaking GPT Say They're Preventing a Closed-Source AI Dystopia" (Mar 2023)      
    https://www.vice.com/en/article/5d9z55/jailbreak-gpt-openai-closed-source      
    Note: Systems such as ChatGPT implement moderation filters designed to prevent users from retrieving harmful content; this kind of middleware may be essential to regulating generative AI. Communities of users share prompts designed to overrise these measures, with varying degrees of success. 


Please contact John Erickson if you have suggestions for this list!