Aligning law, policy, and machine learning for responsible real-world deployments
Abstract/Contents
- Abstract
- Recent advances in machine learning (ML) have seen systems achieve human-like performance in specific tasks, especially in the realms of sequential decision-making and foundation models. Foundation models, powered by Transformer architectures, offer unparalleled abilities to generalize from vast text data, while innovations in reinforcement learning have allowed for mastering of tasks from playing Go to controlling nuclear fusion reactors. Consequently, ML is swiftly integrating into critical sectors like law, healthcare, and government, promising potential public benefits but simultaneously introducing intricate ethical, legal, and policy challenges. As ML deployments grow in real-world scenarios, exploring their interplay with laws, policies, and societal norms becomes paramount. This thesis seeks to address this nexus of ML and legal-policy environments. It endeavors to align AI's potential for public good while ensuring its safe and equitable deployment by bringing together the fields of AI, law, and policy. The first part of this thesis explores machine learning for public benefit. First, it investigates the application of foundation models in legal contexts, examining their strengths and potential pitfalls. Second, it delves into the challenges and opportunities of employing sequential decision-making algorithms in law and public policy, highlighting the unique methodological challenges posed by these domains---as well as potential opportunities for positive impact. Finally, it presents an in-depth case study from a collaboration with the Internal Revenue Service (IRS), introducing the optimize-and-estimate structured bandits problem and offering novel solutions with profound implications for tax compliance and revenue generation. However, realizing ML's societal potential necessitates stringent safeguards, which the second part of the thesis examines. First, it provides an in-depth study of the alignment of foundation models data filtering rules with human-crafted norms in legal settings, leveraging the vast body of legal decision-making as a source of rule-based alignment of machine learning filtering rules. Second, it discusses the dual-use nature of foundation models and introduces a novel approach of task blocking, producing models that are inherently resistant to repurposing for potentially harmful tasks. This work pioneers an interdisciplinary roadmap, blending ML research with legal and policy insights with the aim of fostering responsible real-world machine learning deployments for public good.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2023; ©2023 |
Publication date | 2023; 2023 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Henderson, Peter, (Computer scientist) |
---|---|
Degree supervisor | Jurafsky, Dan |
Thesis advisor | Jurafsky, Dan |
Thesis advisor | Ho, Daniel E |
Thesis advisor | Liang, Percy |
Degree committee member | Ho, Daniel E |
Degree committee member | Liang, Percy |
Associated with | Stanford University, School of Engineering |
Associated with | Stanford University, Computer Science Department |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Peter Henderson. |
---|---|
Note | Submitted to the Computer Science Department. |
Thesis | Thesis Ph.D. Stanford University 2023. |
Location | https://purl.stanford.edu/qx441mp9479 |
Access conditions
- Copyright
- © 2023 by Peter Henderson
Also listed in
Loading usage metrics...