Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy, Research, and Practice

We articulate fundamental mismatches between technical methods for machine unlearning in Generative AI, and documented aspirations for broader impact that these methods could have for law and policy. These aspirations are both numerous and varied, motivated by issues that pertain to privacy, copyright, safety, and more. For example, unlearning is often invoked as a solution for removing the effects of targeted information from a generative-AI model's parameters, e.g., a particular individual's personal data or in-copyright expression of Spiderman that was included in the model's training data. Unlearning is also proposed as a way to prevent a model from generating targeted types of information in its outputs, e.g., generations that closely resemble a particular individual's data or reflect the concept of "Spiderman." Both of these goals--the targeted removal of information from a model and the targeted suppression of information from a model's outputs--present various technical and substantive challenges. We provide a framework for thinking rigorously about these challenges, which enables us to be clear about why unlearning is not a general-purpose solution for circumscribing generative-AI model behavior in service of broader positive impact. We aim for conceptual clarity and to encourage more thoughtful communication among machine learning (ML), law, and policy experts who seek to develop and apply technical methods for compliance with policy objectives.

A Framework for Evaluating LLMs Under Task Indeterminacy

Large language model (LLM) evaluations often assume there is a single correct response -- a gold label -- for each item in the evaluation corpus. However, some tasks can be ambiguous -- i.e., they provide insufficient information to identify a unique interpretation -- or vague -- i.e., they do not clearly indicate where to draw the line when making a determination. Both ambiguity and vagueness can cause task indeterminacy -- the condition where some items in the evaluation corpus have more than one correct response. In this paper, we develop a framework for evaluating LLMs under task indeterminacy. Our framework disentangles the relationships between task specification, human ratings, and LLM responses in the LLM evaluation pipeline. Using our framework, we conduct a synthetic experiment showing that evaluations that use the "gold label" assumption underestimate the true performance. We also provide a method for estimating an error-adjusted performance interval given partial knowledge about indeterminate items in the evaluation corpus. We conclude by outlining implications of our work for the research community.

Measuring machine learning harms from stereotypes: requires understanding who is being harmed by which errors in what ways

As machine learning applications proliferate, we need an understanding of their potential for harm. However, current fairness metrics are rarely grounded in human psychological experiences of harm. Drawing on the social psychology of stereotypes, we use a case study of gender stereotypes in image search to examine how people react to machine learning errors. First, we use survey studies to show that not all machine learning errors reflect stereotypes nor are equally harmful. Then, in experimental studies we randomly expose participants to stereotype-reinforcing, -violating, and -neutral machine learning errors. We find stereotype-reinforcing errors induce more experientially (i.e., subjectively) harmful experiences, while having minimal changes to cognitive beliefs, attitudes, or behaviors. This experiential harm impacts women more than men. However, certain stereotype-violating errors are more experientially harmful for men, potentially due to perceived threats to masculinity. We conclude that harm cannot be the sole guide in fairness mitigation, and propose a nuanced perspective depending on who is experiencing what harm and why.

On the Actionability of Outcome Prediction

Predicting future outcomes is a prevalent application of machine learning in social impact domains. Examples range from predicting student success in education to predicting disease risk in healthcare. Practitioners recognize that the ultimate goal is not just to predict but to act effectively. Increasing evidence suggests that relying on outcome predictions for downstream interventions may not have desired results. In most domains there exists a multitude of possible interventions for each individual, making the challenge of taking effective action more acute. Even when causal mechanisms connecting the individual's latent states to outcomes is well understood, in any given instance (a specific student or patient), practitioners still need to infer -- from budgeted measurements of latent states -- which of many possible interventions will be most effective for this individual. With this in mind, we ask: when are accurate predictors of outcomes helpful for identifying the most suitable intervention? Through a simple model encompassing actions, latent states, and measurements, we demonstrate that pure outcome prediction rarely results in the most effective policy for taking actions, even when combined with other measurements. We find that except in cases where there is a single decisive action for improving the outcome, outcome prediction never maximizes "action value", the utility of taking actions. Making measurements of actionable latent states, where specific actions lead to desired outcomes, considerably enhances the action value compared to outcome prediction, and the degree of improvement depends on action costs and the outcome model. This analysis emphasizes the need to go beyond generic outcome prediction in interventional settings by incorporating knowledge of plausible actions and latent states.

Multi-Target Multiplicity: Flexibility and Fairness in Target Specification under Resource Constraints

Prediction models have been widely adopted as the basis for decision-making in domains as diverse as employment, education, lending, and health. Yet, few real world problems readily present themselves as precisely formulated prediction tasks. In particular, there are often many reasonable target variable options. Prior work has argued that this is an important and sometimes underappreciated choice, and has also shown that target choice can have a significant impact on the fairness of the resulting model. However, the existing literature does not offer a formal framework for characterizing the extent to which target choice matters in a particular task. Our work fills this gap by drawing connections between the problem of target choice and recent work on predictive multiplicity. Specifically, we introduce a conceptual and computational framework for assessing how the choice of target affects individuals' outcomes and selection rate disparities across groups. We call this multi-target multiplicity. Along the way, we refine the study of single-target multiplicity by introducing notions of multiplicity that respect resource constraints -- a feature of many real-world tasks that is not captured by existing notions of predictive multiplicity. We apply our methods on a healthcare dataset, and show that the level of multiplicity that stems from target variable choice can be greater than that stemming from nearly-optimal models of a single target.

Variance, Self-Consistency, and Arbitrariness in Fair Classification

In fair classification, it is common to train a model, and to compare and correct subgroup-specific error rates for disparities. However, even if a model's classification decisions satisfy a fairness metric, it is not necessarily the case that these decisions are equally confident. This becomes clear if we measure variance: We can fix everything in the learning process except the subset of training data, train multiple models, measure (dis)agreement in predictions for each test example, and interpret disagreement to mean that the learning process is more unstable with respect to its classification decision. Empirically, some decisions can in fact be so unstable that they are effectively arbitrary. To reduce this arbitrariness, we formalize a notion of self-consistency of a learning process, develop an ensembling algorithm that provably increases self-consistency, and empirically demonstrate its utility to often improve both fairness and accuracy. Further, our evaluation reveals a startling observation: Applying ensembling to common fair classification benchmarks can significantly reduce subgroup error rate disparities, without employing common pre-, in-, or post-processing fairness interventions. Taken together, our results indicate that variance, particularly on small datasets, can muddle the reliability of conclusions about fairness. One solution is to develop larger benchmark tasks. To this end, we release a toolkit that makes the Home Mortgage Disclosure Act datasets easily usable for future research.

Informational Diversity and Affinity Bias in Team Growth Dynamics

Prior work has provided strong evidence that, within organizational settings, teams that bring a diversity of information and perspectives to a task are more effective than teams that do not. If this form of informational diversity confers performance advantages, why do we often see largely homogeneous teams in practice? One canonical argument is that the benefits of informational diversity are in tension with affinity bias. To better understand the impact of this tension on the makeup of teams, we analyze a sequential model of team formation in which individuals care about their team's performance (captured in terms of accurately predicting some future outcome based on a set of features) but experience a cost as a result of interacting with teammates who use different approaches to the prediction task. Our analysis of this simple model reveals a set of subtle behaviors that team-growth dynamics can exhibit: (i) from certain initial team compositions, they can make progress toward better performance but then get stuck partway to optimally diverse teams; while (ii) from other initial compositions, they can also move away from this optimal balance as the majority group tries to crowd out the opinions of the minority. The initial composition of the team can determine whether the dynamics will move toward or away from performance optimality, painting a path-dependent picture of inefficiencies in team compositions. Our results formalize a fundamental limitation of utility-based motivations to drive informational diversity in organizations and hint at interventions that may improve informational diversity and performance simultaneously.

Mimetic Models: Ethical Implications of AI that Acts Like You

An emerging theme in artificial intelligence research is the creation of models to simulate the decisions and behavior of specific people, in domains including game-playing, text generation, and artistic expression. These models go beyond earlier approaches in the way they are tailored to individuals, and the way they are designed for interaction rather than simply the reproduction of fixed, pre-computed behaviors. We refer to these as mimetic models, and in this paper we develop a framework for characterizing the ethical and social issues raised by their growing availability. Our framework includes a number of distinct scenarios for the use of such models, and considers the impacts on a range of different participants, including the target being modeled, the operator who deploys the model, and the entities that interact with it.

Measuring Representational Harms in Image Captioning

Previous work has largely considered the fairness of image captioning systems through the underspecified lens of "bias." In contrast, we present a set of techniques for measuring five types of representational harms, as well as the resulting measurements obtained for two of the most popular image captioning datasets using a state-of-the-art image captioning system. Our goal was not to audit this image captioning system, but rather to develop normatively grounded measurement techniques, in turn providing an opportunity to reflect on the many challenges involved. We propose multiple measurement techniques for each type of harm. We argue that by doing so, we are better able to capture the multi-faceted nature of each type of harm, in turn improving the (collective) validity of the resulting measurements. Throughout, we discuss the assumptions underlying our measurement approach and point out when they do not hold.

REAL ML: Recognizing, Exploring, and Articulating Limitations of Machine Learning Research

Transparency around limitations can improve the scientific rigor of research, help ensure appropriate interpretation of research findings, and make research claims more credible. Despite these benefits, the machine learning (ML) research community lacks well-developed norms around disclosing and discussing limitations. To address this gap, we conduct an iterative design process with 30 ML and ML-adjacent researchers to develop and test REAL ML, a set of guided activities to help ML researchers recognize, explore, and articulate the limitations of their research. Using a three-stage interview and survey study, we identify ML researchers' perceptions of limitations, as well as the challenges they face when recognizing, exploring, and articulating limitations. We develop REAL ML to address some of these practical challenges, and highlight additional cultural challenges that will require broader shifts in community norms to address. We hope our study and REAL ML help move the ML research community toward more active and appropriate engagement with limitations.