Alignment methods in moral domains seek to elicit moral preferences of human stakeholders and incorporate them into AI. This presupposes moral preferences as static targets, but such preferences often evolve over time. Proper alignment of AI to dynamic human preferences should ideally account for "legitimate" changes to moral reasoning, while ignoring changes related to attention deficits, cognitive biases, or other arbitrary factors. However, common AI alignment approaches largely neglect temporal changes in preferences, posing serious challenges to proper alignment, especially in high-stakes applications of AI, e.g., in healthcare domains, where misalignment can jeopardize the trustworthiness of the system and yield serious individual and societal harms. This work investigates the extent to which people's moral preferences change over time, and the impact of such changes on AI alignment. Our study is grounded in the kidney allocation domain, where we elicit responses to pairwise comparisons of hypothetical kidney transplant patients from over 400 participants across 3-5 sessions. We find that, on average, participants change their response to the same scenario presented at different times around 6-20% of the time (exhibiting "response instability"). Additionally, we observe significant shifts in several participants' retrofitted decision-making models over time (capturing "model instability"). Predictive performance of simple AI models decreases as a function of both response and model instability. Moreover, predictive performance diminishes over time, highlighting the importance of accounting for temporal changes in preferences during training. These findings raise fundamental normative and technical challenges relevant to AI alignment, highlighting the need to better understand the object of alignment (what to align to) when user preferences change significantly over time, including the mechanisms underlying these changes.

A growing body of work in Ethical AI attempts to capture human moral judgments through simple computational models.The key question we address in this work is whether such simple AI models capture the critical nuances of moral decision-making by focusing on the use case of kidney allocation.We conducted twenty interviews where participants explained their rationale for their judgments about who should receive a kidney.We observe participants: (a) value patients' morally-relevant attributes to different degrees; (b) use diverse decision-making processes, citing heuristics to reduce decision complexity; (c) can change their opinions; (d) sometimes lack confidence in their decisions (e.g., due to incomplete information); and (e) express enthusiasm and concern regarding AI assisting humans in kidney allocation decisions.Based on these findings, we discuss challenges of computationally modeling moral judgments as a stand-in for human input, highlight drawbacks of current approaches, and suggest future directions to address these issues.

Preference elicitation frameworks feature heavily in the research on participatory ethical AI tools and provide a viable mechanism to enquire and incorporate the moral values of various stakeholders. As part of the elicitation process, surveys about moral preferences, opinions, and judgments are typically administered only once to each participant. This methodological practice is reasonable if participants’ responses are stable over time such that, all other things being held constant, their responses today will be the same as their responses to the same questions at a later time. However, we do not know how often that is the case. It is possible that participants’ true moral preferences change, are subject to temporary moods or whims, or are influenced by environmental factors we don’t track. If participants’ moral responses are unstable in such ways, it would raise important methodological and theoretical issues for how participants’ true moral preferences, opinions, and judgments can be ascertained. We address this possibility here by asking the same survey participants the same moral questions about which patient should receive a kidney when only one is available ten times in ten different sessions over two weeks, varying only presentation order across sessions. We measured how often participants gave different responses to simple (Study One) and more complicated (Study Two) controversial and uncontroversial repeated scenarios. On average, the fraction of times participants changed their responses to controversial scenarios (i.e., were unstable) was around 10-18% (±14-15%) across studies, and this instability is observed to have positive associations with response time and decision-making difficulty. We discuss the implications of these results for the efficacy of common moral preference elicitation methods, highlighting the role of response instability in potentially causing value misalignment between the stakeholders and AI tools trained on their moral judgments.

Real-world applications of machine learning (ML) algorithms often propagate negative stereotypes and social biases against marginalized groups. In response, the field of fair machine learning has proposed technical solutions for a variety of settings that aim to correct the biases in algorithmic predictions. These solutions remove the dependence of the final prediction on the protected attributes (like gender or race) and/or ensure that prediction performance is similar across demographic groups. Yet, recent studies assessing the impact of these solutions in practice demonstrate their ineffectiveness in tackling real-world inequalities. Given this lack of real-world success, it is essential to take a step back and question the design motivations of algorithmic fairness interventions. We use popular legal anti-discriminatory principles, specifically anti-classification and anti-subordination principles, to study the motivations of fairness interventions and their applications. The anti-classification principle suggests addressing discrimination by ensuring that decision processes and outcomes are independent of the protected attributes of individuals. The anti-subordination principle, on the other hand, argues that decision-making policies can provide equal protection to all only by actively tackling societal hierarchies that enable structural discrimination, even if that requires using protected attributes to address historical inequalities. Through a survey of the fairness mechanisms and applications, we assess different components of fair ML approaches from the perspective of these principles. We argue that the observed shortcomings of fair ML algorithms are similar to the failures of anti-classification policies and that these shortcomings constitute violations of the anti-subordination principle. Correspondingly, we propose guidelines for algorithmic fairness interventions to adhere to the anti-subordination principle. In doing so, we hope to bridge critical concepts between legal frameworks for non-discrimination and fairness in machine learning.

Computational preference elicitation methods are tools used to learn people’s preferences quantitatively in a given context. Recent works on preference elicitation advocate for active learning as an efficient method to iteratively construct queries (framed as comparisons between context-specific cases) that are likely to be most informative about an agent’s underlying preferences. In this work, we argue that the use of active learning for moral preference elicitation relies on certain assumptions about the underlying moral preferences, which can be violated in practice. Specifically, we highlight the following common assumptions (a) preferences are stable over time and not sensitive to the sequence of presented queries, (b) the appropriate hypothesis class is chosen to model moral preferences, and (c) noise in the agent’s responses is limited. While these assumptions can be appropriate for preference elicitation in certain domains, prior research on moral psychology suggests they may not be valid for moral judgments. Through a synthetic simulation of preferences that violate the above assumptions, we observe that active learning can have similar or worse performance than a basic random query selection method in certain settings. Yet, simulation results also demonstrate that active learning can still be viable if the degree of instability or noise is relatively small and when the agent’s preferences can be approximately represented with the hypothesis class used for learning. Our study highlights the nuances associated with effective moral preference elicitation in practice and advocates for the cautious use of active learning as a methodology to learn moral preferences.

In many predictive contexts (e.g., credit lending), true outcomes are only observed for samples that were positively classified in the past. These past observations, in turn, form training datasets for classifiers that make future predictions. However, such training datasets lack information about the outcomes of samples that were (incorrectly) negatively classified in the past and can lead to erroneous classifiers. We present an approach that trains a classifier using available data and comes with a family of exploration strategies to collect outcome data about subpopulations that otherwise would have been ignored. For any exploration strategy, the approach comes with guarantees that (1) all sub-populations are explored, (2) the fraction of false positives is bounded, and (3) the trained classifier converges to a "desired" classifier. The right exploration strategy is context-dependent; it can be chosen to improve learning guarantees and encode context-specific group fairness properties. Evaluation on real-world datasets shows that this approach consistently boosts the quality of collected outcome data and improves the fraction of true positives for all groups, with only a small reduction in predictive utility.

We study a variant of a recently introduced min-max optimization framework where the max-player is constrained to update its parameters in a greedy manner until it reaches a first-order stationary point. Our equilibrium definition for this framework depends on a proposal distribution which the min-player uses to choose directions in which to update its parameters. We show that, given a smooth and bounded nonconvex-nonconcave objective function, access to any proposal distribution for the min-player’s updates, and stochastic gradient oracle for the max-player, our algorithm converges to the aforementioned approximate local equilibrium in a number of iterations that does not depend on the dimension. The equilibrium point found by our algorithm depends on the proposal distribution, and when applying our algorithm to train GANs we choose the proposal distribution to be a distribution of stochastic gradients. We empirically evaluate our algorithm on challenging nonconvex-nonconcave test-functions and loss functions arising in GAN training. Our algorithm converges on these test functions and, when used to train GANs, trains stably on synthetic and real-world datasets and avoids mode collapse.

Machine learning models are often implemented in cohort with humans in the pipeline, with the model having an option to defer to a domain expert in cases where it has low confidence in its inference. Our goal is to design mechanisms for ensuring accuracy and fairness in such prediction systems that combine machine learning model inferences and domain expert predictions. Prior work on "deferral systems" in classification settings has focused on the setting of a pipeline with a single expert and aimed to accommodate the inaccuracies and biases of this expert to simultaneously learn an inference model and a deferral system. Our work extends this framework to settings where multiple experts are available, with each expert having their own domain of expertise and biases. We propose a framework that simultaneously learns a classifier and a deferral system, with the deferral system choosing to defer to one or more human experts in cases of input where the classifier has low confidence. We test our framework on a synthetic dataset and a content moderation dataset with biased synthetic experts, and show that it significantly improves the accuracy and fairness of the final predictions, compared to the baselines. We also collect crowdsourced labels for the content moderation task to construct a real-world dataset for the evaluation of hybrid machine-human frameworks and show that our proposed framework outperforms baselines on this real-world dataset as well.

We present an optimization framework for learning a fair classifier in the presence of noisy perturbations in the protected attributes. Compared to prior work, our framework can be employed with a very general class of linear and linear-fractional fairness constraints, can handle multiple, non-binary protected attributes, and outputs a classifier that comes with provable guarantees on both accuracy and fairness. Empirically, we show that our framework can be used to attain either statistical rate or false positive rate fairness guarantees with a minimal loss in accuracy, even when the noise is large, in two real-world datasets.

Assessing the diversity of a dataset of information associated with people is crucial before using such data for downstream applications. For a given dataset, this often involves computing the imbalance or disparity in the empirical marginal distribution of a protected attribute (e.g. gender, dialect, etc.). However, real-world datasets, such as images from Google Search or collections of Twitter posts, often do not have protected attributes labeled. Consequently, to derive disparity measures for such datasets, the elements need to hand-labeled or crowd-annotated, which are expensive processes.

Discussions on Twitter involve participation from different communities with different dialects and it is often necessary to summarize a large number of posts into a representative sample to provide a synopsis. Yet, any such representative sample should sufficiently portray the underlying dialect diversity to present the voices of different participating communities representing the dialects. Extractive summarization algorithms perform the task of constructing subsets that succinctly capture the topic of any given set of posts. However, we observe that there is dialect bias in the summaries generated by common summarization approaches, i.e., they often return summaries that under-represent certain dialects.

Data containing human or social attributes may over- or under-represent groups with respect to salient social attributes such as gender or race, which can lead to biases in downstream applications. This paper presents an algorithmic framework that can be used as a data preprocessing method towards mitigating such bias. Unlike prior work, it can efficiently learn distributions over large domains, controllably adjust the representation rates of protected groups and achieve target fairness metrics such as statistical parity, yet remains close to the empirical distribution induced by the given dataset. Our approach leverages the principle of maximum entropy {–} amongst all distributions satisfying a given set of constraints, we should choose the one closest in KL-divergence to a given prior. While maximum entropy distributions can succinctly encode distributions over large domains, they can be difficult to compute. Our main contribution is an instantiation of this framework for our set of constraints and priors, which encode our bias mitigation goals, and that runs in time polynomial in the dimension of the data. Empirically, we observe that samples from the learned distribution have desired representation rates and statistical rates, and when used for training a classifier incurs only a slight loss in accuracy while maintaining fairness properties.

Studies have shown that the people depicted in image search results tend to be of majority groups with respect to socially salient attributes such as gender or race. This skew goes beyond that which already exists in the world - i.e., the search results for images of people are more imbalanced than the ground truth would suggest. For example, Kay et al. showed that although 28% of CEOs in the U.S. are women, only 10% of the top 100 results for "CEO" in Google Image Search are women. Similar observations abound across search terms and across socially salient attributes. Most existing approaches to correct for this kind of bias assume that the images of people include labels denoting the relevant socially salient attributes. These labels are explicitly used to either change the dataset, adjust the training of the algorithm, and/or in the execution of the method. However, such labels are often unknown. Further, using machine learning techniques to infer these labels may often not be possible within acceptable accuracy ranges and may not be desirable due to the additional biases this process could incur. As observed in prior work, alternate approaches consider the diversity of image features, which often do not translate to images of visibly diverse people. We develop a novel approach that takes as input a visibly diverse control set of images of people and uses this set as part of a procedure to select a set of images of people in response to a query. The goal is to have a resulting set that is more visibly diverse in a manner that emulates the diversity depicted in the control set. It accomplishes this by evaluating the similarity of the images selected by a black-box algorithm with the images in the diversity control set, and incorporating this "diversity score" into the final selection process. Importantly, this approach does not require images to be labelled at any point; effectively, it gives a way to implicitly diversify the set of images selected. We provide two variants of our approach: the first is a modification of the well known MMR algorithm to incorporate the diversity scores, and the second is a more efficient variant that does not consider within-list redundancy. We evaluate these approaches empirically on two image datasets: 1) a new dataset we collect which contains the top 100 Google Image results for 96 occupations, for which we evaluate gender and skin-tone diversity with respect to occupations and 2) the well-known CelebA dataset containing images of celebrities for which we can evaluate gender diversity with respect to facial features such as "smiling" or "glasses". Both of our approaches produce image sets that significantly improve the visible diversity of the results (i.e., include a larger fraction of anti-stereotypical images) with respect to current Google Image Search results and other state-of-the-art algorithms for diverse image summarization. Further, they seem to come at a minimal cost to accuracy. These empirical results demonstrate the effectiveness of simple label-independent interventions to diversify image search.

Sampling methods that choose a subset of the data proportional to its diversity in the feature space are popular for data summarization. However, recent studies have noted the occurrence of bias {–} e.g., under or over representation of a particular gender or ethnicity {–} in such data summarization methods. In this paper we initiate a study of the problem of outputting a diverse and fair summary of a given dataset. We work with a well-studied determinantal measure of diversity and corresponding distributions (DPPs) and present a framework that allows us to incorporate a general class of fairness constraints into such distributions. Designing efficient algorithms to sample from these constrained determinantal distributions, however, suffers from a complexity barrier; we present a fast sampler that is provably good when the input vectors satisfy a natural property. Our empirical results on both real-world and synthetic datasets show that the diversity of the samples produced by adding fairness constraints is not too far from the unconstrained case.

We present a prototype for a news search engine that presents balanced viewpoints across liberal and conservative articles with the goal of depolarizing content and allowing users to escape their filter bubble. The balancing is done according to flexible user-defined constraints, and leverages recent advances in constrained bandit optimization. We showcase our balanced news feed by displaying it side-by-side with the news feed produced by a traditional (polarized) feed.