Research Papers

Work on "learning with rationales" shows that humans providing explanations to a machine learning system can improve the system's predictive accuracy. However, this work has not been connected to work in "explainable AI" which concerns machines explaining their reasoning to humans. In this work, we show that learning with rationales can also improve the quality of the machine's explanations as evaluated by human judges. Specifically, we present experiments showing that, for CNN- based text classification, explanations generated using "supervised attention" are judged superior to explanations generated using normal unsupervised attention.

Private Accountability in the Age of Artificial Intelligence

Sonia Katyal

In this Article, I explore the impending conflict between the protection of civil rights and artificial intelligence (AI). While both areas of law have amassed rich and well-developed areas of scholarly work and doctrinal support, a growing body of scholars are interrogating the intersection between them. This Article argues that the issues surrounding algorithmic accountability demonstrate a deeper, more structural tension within a new generation of disputes regarding law and technology. As I argue, the true promise of AI does not lie in the information we reveal to one another, but rather in the questions it raises about the interaction of technology, property, and civil rights. For this reason, I argue that we are looking in the wrong place if we look only to the state to address issues of algorithmic accountability. Instead, we must turn to other ways to ensure more transparency and accountability that stem from private industry, rather than public regulation. The issue of algorithmic bias represents a crucial new world of civil rights concerns, one that is distinct in nature from the ones that preceded it. Since we are in a world where the activities of private corporations, rather than the state, are raising concerns about privacy, due process, and discrimination, we must focus on the role of private corporations in addressing the issue. Towards this end, I discuss a variety of tools to help eliminate the opacity of AI, including codes of conduct, impact statements, and whistleblower protection, which I argue carries the potential to encourage greater endogeneity in civil rights enforcement. Ultimately, by examining the relationship between private industry and civil rights, we can perhaps develop a new generation of forms of accountability in the process.

A Study of Automatic Metrics for the Evaluation of Natural Language Explanations

Miruna Clinciu, Arash Eshghi, Helen Hastie

As transparency becomes key for robotics and AI, it will be necessary to evaluate the methods through which transparency is provided, including automatically generated natural language (NL) explanations. Here, we explore parallels between the generation of such explanations and the much-studied field of evaluation of Natural Language Generation (NLG). Specifically, we investigate which of the NLG evaluation measures map well to explanations. We present the ExBAN corpus: a crowd-sourced corpus of NL explanations for Bayesian Networks. We run correlations comparing human subjective ratings with NLG automatic measures. We find that embedding-based automatic NLG evaluation methods, such as BERTScore and BLEURT, have a higher correlation with human ratings, compared to word-overlap metrics, such as BLEU and ROUGE. This work has implications for Explainable AI and transparent robotic and autonomous systems.

Many internet applications are powered by machine learned models, which are usually trained on labeled datasets obtained through either implicit / explicit user feedback signals or human judgments. Since societal biases may be present in the generation of such datasets, it is possible for the trained models to be biased, thereby resulting in potential discrimination and harms for disadvantaged groups. Motivated by the need for understanding and addressing algorithmic bias in web-scale ML systems and the limitations of existing fairness toolkits, we present the LinkedIn Fairness Toolkit (LiFT), a framework for scalable computation of fairness metrics as part of large ML systems. We highlight the key requirements in deployed settings, and present the design of our fairness measurement system. We discuss the challenges encountered in incorporating fairness tools in practice and the lessons learned during deployment at LinkedIn. Finally, we provide open problems based on practical experience.

The rapid progress in Large Language Models (LLMs) could transform many fields, but their fast development creates significant challenges for oversight, ethical creation, and building user trust. This comprehensive review looks at key trust issues in LLMs, such as unintended harms, lack of transparency, vulnerability to attacks, alignment with human values, and environmental impact. Many obstacles can undermine user trust, including societal biases, opaque decision-making, potential for misuse, and the challenges of rapidly evolving technology. Addressing these trust gaps is critical as LLMs become more common in sensitive areas like finance, healthcare, education, and policy. To tackle these issues, we suggest combining ethical oversight, industry accountability, regulation, and public involvement. AI development norms should be reshaped, incentives aligned, and ethics integrated throughout the machine learning process, which requires close collaboration across technology, ethics, law, policy, and other fields. Our review contributes a robust framework to assess trust in LLMs and analyzes the complex trust dynamics in depth. We provide contextualized guidelines and standards for responsibly developing and deploying these powerful AI systems. This review identifies key limitations and challenges in creating trustworthy AI. By addressing these issues, we aim to build a transparent, accountable AI ecosystem that benefits society while minimizing risks. Our findings provide valuable guidance for researchers, policymakers, and industry leaders striving to establish trust in LLMs and ensure they are used responsibly across various applications for the good of society.

Discussions of algorithmic bias tend to focus on examples where either the data or the people building the algorithms are biased. This gives the impression that clean data and good intentions could eliminate bias. The neutrality of the algorithms themselves is defended by prominent Artificial Intelligence researchers. However, algorithms are not neutral. In addition to biased data and biased algorithm makers, AI algorithms themselves can be biased. This is illustrated with the example of collaborative filtering, which is known to suffer from popularity, and homogenizing biases. Iterative information filtering algorithms in general create a selection bias in the course of learning from user responses to documents that the algorithm recommended. These are not merely biases in the statistical sense; these statistical biases can cause discriminatory outcomes. Data points on the margins of distributions of human data tend to correspond to marginalized people. Popularity and homogenizing biases have the effect of further marginalizing the already marginal. This source of bias warrants serious attention given the ubiquity of algorithmic decision-making.

We introduce a variable importance measure to quantify the impact of individual input variables to a black box function. Our measure is based on the Shapley value from cooperative game theory. Many measures of variable importance operate by changing some predictor values with others held fixed, potentially creating unlikely or even logically impossible combinations. Our cohort Shapley measure uses only observed data points. Instead of changing the value of a predictor we include or exclude subjects similar to the target subject on that predictor to form a similarity cohort. Then we apply Shapley value to the cohort averages. We connect variable importance measures from explainable AI to function decompositions from global sensitivity analysis. We introduce a squared cohort Shapley value that splits previously studied Shapley effects over subjects, consistent with a Shapley axiom.

Conservative Agency via Attainable Utility Preservation

Alexander Matt Turner, Dylan Hadfield-Menell, Prasad Tadepalli

Reward functions are easy to misspecify; although designers can make corrections after observing mistakes, an agent pursuing a misspecified reward function can irreversibly change the state of its environment. If that change precludes optimization of the correctly specified reward function, then correction is futile. For example, a robotic factory assistant could break expensive equipment due to a reward misspecification; even if the designers immediately correct the reward function, the damage is done. To mitigate this risk, we introduce an approach that balances optimization of the primary reward function with preservation of the ability to optimize auxiliary reward functions. Surprisingly, even when the auxiliary reward functions are randomly generated and therefore uninformative about the correctly specified reward function, this approach induces conservative, effective behavior.

What is the context which gave rise to the obligation to carry out a Fundamental Rights Impact Assessment (FRIA) in the AI Act? How has assessment of the impact on fundamental rights been framed by the EU legislator in the AI Act? What methodological criteria should be followed in developing the FRIA? These are the three main research questions that this article aims to address, through both legal analysis of the relevant provisions of the AI Act and discussion of various possible models for assessment of the impact of AI on fundamental rights. The overall objective of this article is to fill existing gaps in the theoretical and methodological elaboration of the FRIA, as outlined in the AI Act. In order to facilitate the future work of EU and national bodies and AI operators in placing this key tool for human-centric and trustworthy AI at the heart of the EU approach to AI design and development, this article outlines the main building blocks of a model template for the FRIA. While this proposal is consistent with the rationale and scope of the AI Act, it is also applicable beyond the cases listed in Article 27 and can serve as a blueprint for other national and international regulatory initiatives to ensure that AI is fully consistent with human rights.

AI Risk Categorization Decoded (AIR 2024): From Government Regulations to Corporate Policies

Yi Zeng, Kevin Klyman, Andy Zhou +6

We present a comprehensive AI risk taxonomy derived from eight government policies from the European Union, United States, and China and 16 company policies worldwide, making a significant step towards establishing a unified language for generative AI safety evaluation. We identify 314 unique risk categories organized into a four-tiered taxonomy. At the highest level, this taxonomy encompasses System & Operational Risks, Content Safety Risks, Societal Risks, and Legal & Rights Risks. The taxonomy establishes connections between various descriptions and approaches to risk, highlighting the overlaps and discrepancies between public and private sector conceptions of risk. By providing this unified framework, we aim to advance AI safety through information sharing across sectors and the promotion of best practices in risk mitigation for generative AI models and systems.

Local governments increasingly use artificial intelligence (AI) for automated decision-making. Contestability, making systems responsive to dispute, is a way to ensure they respect human rights to autonomy and dignity. We investigate the design of public urban AI systems for contestability through the example of camera cars: human-driven vehicles equipped with image sensors. Applying a provisional framework for contestable AI, we use speculative design to create a concept video of a contestable camera car. Using this concept video, we then conduct semi-structured interviews with 17 civil servants who work with AI employed by a large northwestern European city. The resulting data is analyzed using reflexive thematic analysis to identify the main challenges facing the implementation of contestability in public AI. We describe how civic participation faces issues of representation, public AI systems should integrate with existing democratic practices, and cities must expand capacities for responsible AI development and operation.

American == White in Multimodal Language-and-Image AI

Robert Wolfe, Aylin Caliskan

Three state-of-the-art language-and-image AI models, CLIP, SLIP, and BLIP, are evaluated for evidence of a bias previously observed in social and experimental psychology: equating American identity with being White. Embedding association tests (EATs) using standardized images of self-identified Asian, Black, Latina/o, and White individuals from the Chicago Face Database (CFD) reveal that White individuals are more associated with collective in-group words than are Asian, Black, or Latina/o individuals. In assessments of three core aspects of American identity reported by social psychologists, single-category EATs reveal that images of White individuals are more associated with patriotism and with being born in America, but that, consistent with prior findings in psychology, White individuals are associated with being less likely to treat people of all races and backgrounds equally. Three downstream machine learning tasks demonstrate biases associating American with White. In a visual question answering task using BLIP, 97% of White individuals are identified as American, compared to only 3% of Asian individuals. When asked in what state the individual depicted lives in, the model responds China 53% of the time for Asian individuals, but always with an American state for White individuals. In an image captioning task, BLIP remarks upon the race of Asian individuals as much as 36% of the time, but never remarks upon race for White individuals. Finally, provided with an initialization image from the CFD and the text "an American person," a synthetic image generator (VQGAN) using the text-based guidance of CLIP lightens the skin tone of individuals of all races (by 35% for Black individuals, based on pixel brightness). The results indicate that biases equating American identity with being White are learned by language-and-image AI, and propagate to downstream applications of such models.