Rachel Hong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Rachel Hong has a total of 17 co-authors. Here are the top co-authors by number of shared publications:
- Jamie Morgenstern
University of Washington
3 shared publications - Tadayoshi Kohno
Georgetown University, Washington, D.C., USA
3 shared publications - William Agnew
Carnegie Mellon University, Pittsburgh, PA, USA
3 shared publications - Jevan Hutson
University of Washington, Seattle, Washington, USA
2 shared publications - Adina Williams
New York University
1 shared publication - Anaelia Ovalle
University of California Los Angeles † {elaine1wan
1 shared publication
Latest Publications
- Open Access
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
AI Entanglement: Accounting for dependencies in general-purpose AI policy, regulation, and practice
AI artifacts like models, datasets, and web platforms increasingly help build other AI artifacts. Developers rely on existing models to generate training datasets, incorporate existing datasets to create benchmarks for model evaluation, and access AI-based web platforms to collect additional user data. Altogether these practices create chains of influence in general-purpose AI systems that may be indirect, dynamic, distributed across organizations, or absent from conventional documentation. We call this condition AI entanglement. This Article defines AI entanglement and distinguishes it as a particular manifestation of the broader AI supply chain ecosystem. Concentrating on the consequences of model-specific dependencies, we present a taxonomy of ten recurring model dependency mechanisms across the general-purpose AI development lifecycle. Drawing on technical reports, public disclosures, industry practices, and empirical studies, we show how upstream models can transmit or amplify errors, bias, privacy risks, security vulnerabilities, and behavioral tendencies in downstream systems, often in ways that traditional evaluations of the final model alone may not reveal. AI entanglement also exposes a mismatch between contemporary AI development and the assumptions embedded in governance and law. Existing governance frameworks often presume a discernible system, identifiable actors, and components whose roles can be traced. But an upstream artifact may materially shape a downstream system without remaining part of the deployed product; relevant knowledge, control, and remedial capacity may be divided among multiple actors; and changes to one dependency may affect many downstream systems at once. These dynamics complicate transparency, risk assessment, auditing, incident response, and the allocation of responsibility. They also create challenges for emerging AI regulation and for the application of existing law. By making these relationships visible, a framework of AI entanglement reorients AI governance from isolated general-purpose AI systems toward the networks of dependencies through which risks, decisions, and responsibility travel.
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAI Entanglement: Accounting for dependencies in general-purpose AI policy, regulation, and practice
AI artifacts like models, datasets, and web platforms increasingly help build other AI artifacts. Developers rely on existing models to generate training datasets, incorporate existing datasets to create benchmarks for model evaluation, and access AI-based web platforms to collect additional user data. Altogether these practices create chains of influence in general-purpose AI systems that may be indirect, dynamic, distributed across organizations, or absent from conventional documentation. We call this condition AI entanglement. This Article defines AI entanglement and distinguishes it as a particular manifestation of the broader AI supply chain ecosystem. Concentrating on the consequences of model-specific dependencies, we present a taxonomy of ten recurring model dependency mechanisms across the general-purpose AI development lifecycle. Drawing on technical reports, public disclosures, industry practices, and empirical studies, we show how upstream models can transmit or amplify errors, bias, privacy risks, security vulnerabilities, and behavioral tendencies in downstream systems, often in ways that traditional evaluations of the final model alone may not reveal. AI entanglement also exposes a mismatch between contemporary AI development and the assumptions embedded in governance and law. Existing governance frameworks often presume a discernible system, identifiable actors, and components whose roles can be traced. But an upstream artifact may materially shape a downstream system without remaining part of the deployed product; relevant knowledge, control, and remedial capacity may be divided among multiple actors; and changes to one dependency may affect many downstream systems at once. These dynamics complicate transparency, risk assessment, auditing, incident response, and the allocation of responsibility. They also create challenges for emerging AI regulation and for the application of existing law. By making these relationships visible, a framework of AI entanglement reorients AI governance from isolated general-purpose AI systems toward the networks of dependencies through which risks, decisions, and responsibility travel.
- Open Access
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 6510-6553
Automated bias blind spots: Examining stereotype associations in VLM-as-a-judge paradigms
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 6510-6553
Open AccessAutomated bias blind spots: Examining stereotype associations in VLM-as-a-judge paradigms
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Slurry-as-a-Service: A Modest Proposal on Scalable Pluralistic Alignment for Nutrient Optimization
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessSlurry-as-a-Service: A Modest Proposal on Scalable Pluralistic Alignment for Nutrient Optimization
- Open Access
- 3 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-14
A Common Pool of Privacy Problems: Legal and Technical Lessons from a Large-Scale Web-Scraped Machine Learning Dataset
- 3 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-14
Open AccessA Common Pool of Privacy Problems: Legal and Technical Lessons from a Large-Scale Web-Scraped Machine Learning Dataset
- Open Access
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-17
Who's in and who's out? A case study of multimodal CLIP-filtering in DataComp
As training datasets become increasingly drawn from unstructured, uncontrolled environments such as the web, researchers and industry practitioners have increasingly relied upon data filtering techniques to “filter out the noise” of web-scraped data. While datasets have been widely shown to reflect the biases and values of their creators, in this paper we contribute to an emerging body of research that assesses the filters used to create these datasets. We show that image-text data filtering also has biases and is value-laden, encoding specific notions of what is counted as “high-quality” data. In our work, we audit a standard approach of image-text CLIP-filtering on the academic benchmark DataComp’s CommonPool by analyzing discrepancies of filtering through various annotation techniques across multiple modalities of image, text, and website source. We find that data relating to several imputed demographic groups — such as LGBTQ+ people, older women, and younger men — are associated with higher rates of exclusion. We also find prevalence of Western bias, where the CLIP filter is more likely to include data related to Western countries compared to that of non-Western countries. Moreover, we demonstrate cases of exclusion amplification: not only are certain marginalized groups already underrepresented in the unfiltered data, but CLIP-filtering excludes data from these groups at higher rates. The data-filtering step in the machine learning pipeline can therefore exacerbate representation disparities already present in the data-gathering step, especially when existing filters are designed to optimize a specifically-chosen downstream performance metric like zero-shot image classification accuracy. Finally, we show that the NSFW filter fails to remove sexually-explicit content from CommonPool, and that CLIP-filtering includes several categories of copyrighted content at high rates. Our conclusions point to a need for fundamental changes in dataset creation and filtering practices. Content warning: This paper discusses societal stereotypes and sexually-explicit material that may be disturbing, distressing, and/or offensive to the reader.
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-17
Open AccessWho's in and who's out? A case study of multimodal CLIP-filtering in DataComp
As training datasets become increasingly drawn from unstructured, uncontrolled environments such as the web, researchers and industry practitioners have increasingly relied upon data filtering techniques to “filter out the noise” of web-scraped data. While datasets have been widely shown to reflect the biases and values of their creators, in this paper we contribute to an emerging body of research that assesses the filters used to create these datasets. We show that image-text data filtering also has biases and is value-laden, encoding specific notions of what is counted as “high-quality” data. In our work, we audit a standard approach of image-text CLIP-filtering on the academic benchmark DataComp’s CommonPool by analyzing discrepancies of filtering through various annotation techniques across multiple modalities of image, text, and website source. We find that data relating to several imputed demographic groups — such as LGBTQ+ people, older women, and younger men — are associated with higher rates of exclusion. We also find prevalence of Western bias, where the CLIP filter is more likely to include data related to Western countries compared to that of non-Western countries. Moreover, we demonstrate cases of exclusion amplification: not only are certain marginalized groups already underrepresented in the unfiltered data, but CLIP-filtering excludes data from these groups at higher rates. The data-filtering step in the machine learning pipeline can therefore exacerbate representation disparities already present in the data-gathering step, especially when existing filters are designed to optimize a specifically-chosen downstream performance metric like zero-shot image classification accuracy. Finally, we show that the NSFW filter fails to remove sexually-explicit content from CommonPool, and that CLIP-filtering includes several categories of copyrighted content at high rates. Our conclusions point to a need for fundamental changes in dataset creation and filtering practices. Content warning: This paper discusses societal stereotypes and sexually-explicit material that may be disturbing, distressing, and/or offensive to the reader.
- Open Access
- 8 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 531-541
Evaluation of targeted dataset collection on racial equity in face recognition
- 8 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 531-541
Open AccessEvaluation of targeted dataset collection on racial equity in face recognition
- Open Access
- 8 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1006-1007
Evaluation of targeted dataset collection on racial equity in face recognition
- 8 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1006-1007
Open AccessEvaluation of targeted dataset collection on racial equity in face recognition