arXiv subjects

Kristian Lum

Publications and source records attributed to Kristian Lum.

At least 19 recordsLinked to original sources

arXiv2026

Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model's capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user's moral viewpoint - a failure mode we characterize as moral deliberative sycophancy.

cs.LG↗

arXiv2026

Evaluating Language Models for Harmful Manipulation

Interest in the concept of AI-driven harmful manipulation is growing, yet current approaches to evaluating it are limited. This paper introduces a framework for evaluating harmful AI manipulation via context-specific human-AI interaction studies. We illustrate the utility of this framework by assessing an AI model with 10,101 participants spanning interactions in three AI use domains (public policy, finance, and health) and three locales (US, UK, and India). Overall, we find that that the tested model can produce manipulative behaviours when prompted to do so and, in experimental settings, is able to induce belief and behaviour changes in study participants. We further find that context matters: AI manipulation differs between domains, suggesting that it needs to be evaluated in the high-stakes context(s) in which an AI system is likely to be used. We also identify significant differences across our tested geographies, suggesting that AI manipulation results from one geographic region may not generalise to others. Finally, we find that the frequency of manipulative behaviours (propensity) of an AI model is not consistently predictive of the likelihood of manipulative success (efficacy), underscoring the importance of studying these dimensions separately. To facilitate adoption of our evaluation framework, we detail our testing protocols and make relevant materials publicly available. We conclude by discussing open challenges in evaluating harmful manipulation by AI models.

cs.AI↗

arXiv2024

The Intersectionality Problem for Algorithmic Fairness

A yet unmet challenge in algorithmic fairness is the problem of intersectionality, that is, achieving fairness across the intersection of multiple groups -- and verifying that such fairness has been attained. Because intersectional groups tend to be small, verifying whether a model is fair raises statistical as well as moral-methodological challenges. This paper (1) elucidates the problem of intersectionality in algorithmic fairness, (2) develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, (3) illustrates the desiderata and potential solutions by sketching a proposal using simple hypothesis testing, and (4) evaluates, partly empirically, this proposal against the proposed desiderata.

cs.LG↗

arXiv2024

STAR: SocioTechnical Approach to Red Teaming Language Models

This research introduces STAR, a sociotechnical framework that improves on current best practices for red teaming safety of large language models. STAR makes two key contributions: it enhances steerability by generating parameterised instructions for human red teamers, leading to improved coverage of the risk surface. Parameterised instructions also provide more detailed insights into model failures at no increased cost. Second, STAR improves signal quality by matching demographics to assess harms for specific groups, resulting in more sensitive annotations. STAR further employs a novel step of arbitration to leverage diverse viewpoints and improve label reliability, treating disagreement not as noise but as a valuable contribution to signal quality.

cs.AI↗

arXiv2024

Reducing Population-level Inequality Can Improve Demographic Group Fairness: a Twitter Case Study

Many existing fairness metrics measure group-wise demographic disparities in system behavior or model performance. Calculating these metrics requires access to demographic information, which, in industrial settings, is often unavailable. By contrast, economic inequality metrics, such as the Gini coefficient, require no demographic data to measure. However, reductions in economic inequality do not necessarily correspond to reductions in demographic disparities. In this paper, we empirically explore the relationship between demographic-free inequality metrics -- such as the Gini coefficient -- and standard demographic bias metrics that measure group-wise model performance disparities specifically in the case of engagement inequality on Twitter. We analyze tweets from 174K users over the duration of 2021 and find that demographic-free impression inequality metrics are positively correlated with gender, race, and age disparities in the average case, and weakly (but still positively) correlated with demographic bias in the worst case. We therefore recommend inequality metrics as a potentially useful proxy measure of average group-wise disparities, especially in cases where such disparities cannot be measured directly. Based on these results, we believe they can be used as part of broader efforts to improve fairness between demographic groups in scenarios like content recommendation on social media.

cs.SI↗

arXiv2024

The Impossibility of Fair LLMs

The rise of general-purpose artificial intelligence (AI) systems, particularly large language models (LLMs), has raised pressing moral questions about how to reduce bias and ensure fairness at scale. Researchers have documented a sort of "bias" in the significant correlations between demographics (e.g., race, gender) in LLM prompts and responses, but it remains unclear how LLM fairness could be evaluated with more rigorous definitions, such as group fairness or fair representations. We analyze a variety of technical fairness frameworks and find inherent challenges in each that make the development of a fair LLM intractable. We show that each framework either does not logically extend to the general-purpose AI context or is infeasible in practice, primarily due to the large amounts of unstructured training data and the many potential combinations of human populations, use cases, and sensitive attributes. These inherent challenges would persist for general-purpose AI, including LLMs, even if empirical challenges, such as limited participatory input and limited measurement methods, were overcome. Nonetheless, fairness will remain an important type of model evaluation, and there are still promising research directions, particularly the development of standards for the responsibility of LLM developers, context-specific evaluations, and methods of iterative, participatory, and AI-assisted evaluation that could scale fairness across the diverse contexts of modern human-AI interaction.

cs.CL↗

arXiv2024

Bias in Language Models: Beyond Trick Tests and Toward RUTEd Evaluation

Standard benchmarks of bias and fairness in large language models (LLMs) measure the association between the user attributes stated or implied by a prompt and the LLM's short text response, but human-AI interaction increasingly requires long-form and context-specific system output to solve real-world tasks. In the commonly studied domain of gender-occupation bias, we test whether these benchmarks are robust to lengthening the LLM responses as a measure of Realistic Use and Tangible Effects (i.e., RUTEd evaluations). From the current literature, we adapt three standard bias metrics (neutrality, skew, and stereotype) and develop analogous RUTEd evaluations from three contexts of real-world use: children's bedtime stories, user personas, and English language learning exercises. We find that standard bias metrics have no significant correlation with the more realistic bias metrics. For example, selecting the least biased model based on the standard "trick tests" coincides with selecting the least biased model as measured in more realistic use no more than random chance. We suggest that there is not yet evidence to justify standard benchmarks as reliable proxies of real-world AI biases, and we encourage further development of evaluations grounded in particular contexts.

cs.CL↗

arXiv2023

County-level Algorithmic Audit of Racial Bias in Twitter's Home Timeline

We report on the outcome of an audit of Twitter's Home Timeline ranking system. The goal of the audit was to determine if authors from some racial groups experience systematically higher impression counts for their Tweets than others. A central obstacle for any such audit is that Twitter does not ordinarily collect or associate racial information with its users, thus prohibiting an analysis at the level of individual authors. Working around this obstacle, we take US counties as our unit of analysis. We associate each user in the United States on the Twitter platform to a county based on available location data. The US Census Bureau provides information about the racial decomposition of the population in each county. The question we investigate then is if the racial decomposition of a county is associated with the visibility of Tweets originating from within the county. Focusing on two racial groups, the Black or African American population and the White population as defined by the US Census Bureau, we evaluate two statistical measures of bias. Our investigation represents the first large-scale algorithmic audit into racial bias on the Twitter platform. Additionally, it illustrates the challenges of measuring racial bias in online platforms without having such information on the users.

cs.SI↗

arXiv2022

Random Isn't Always Fair: Candidate Set Imbalance and Exposure Inequality in Recommender Systems

Traditionally, recommender systems operate by returning a user a set of items, ranked in order of estimated relevance to that user. In recent years, methods relying on stochastic ordering have been developed to create "fairer" rankings that reduce inequality in who or what is shown to users. Complete randomization -- ordering candidate items randomly, independent of estimated relevance -- is largely considered a baseline procedure that results in the most equal distribution of exposure. In industry settings, recommender systems often operate via a two-step process in which candidate items are first produced using computationally inexpensive methods and then a full ranking model is applied only to those candidates. In this paper, we consider the effects of inequality at the first step and show that, paradoxically, complete randomization at the second step can result in a higher degree of inequality relative to deterministic ordering of items by estimated relevance scores. In light of this observation, we then propose a simple post-processing algorithm in pursuit of reducing exposure inequality that works both when candidate sets have a high level of imbalance and when they do not. The efficacy of our method is illustrated on both simulated data and a common benchmark data set used in studying fairness in recommender systems.

cs.IR↗

arXiv2022

Flipping the Script on Criminal Justice Risk Assessment: An actuarial model for assessing the risk the federal sentencing system poses to defendants

In the criminal justice system, algorithmic risk assessment instruments are used to predict the risk a defendant poses to society; examples include the risk of recidivating or the risk of failing to appear at future court dates. However, defendants are also at risk of harm from the criminal justice system. To date, there exists no risk assessment instrument that considers the risk the system poses to the individual. We develop a risk assessment instrument that "flips the script." Using data about U.S. federal sentencing decisions, we build a risk assessment instrument that predicts the likelihood an individual will receive an especially lengthy sentence given factors that should be legally irrelevant to the sentencing decision. To do this, we develop a two-stage modeling approach. Our first-stage model is used to determine which sentences were "especially lengthy." We then use a second-stage model to predict the defendant's risk of receiving a sentence that is flagged as especially lengthy given factors that should be legally irrelevant. The factors that should be legally irrelevant include, for example, race, court location, and other socio-demographic information about the defendant. Our instrument achieves comparable predictive accuracy to risk assessment instruments used in pretrial and parole contexts. We discuss the limitations of our modeling approach and use the opportunity to highlight how traditional risk assessment instruments in various criminal justice settings also suffer from many of the same limitations and embedded value systems of their creators.

cs.CY↗

arXiv2022

De-biasing "bias" measurement

When a model's performance differs across socially or culturally relevant groups--like race, gender, or the intersections of many such groups--it is often called "biased." While much of the work in algorithmic fairness over the last several years has focused on developing various definitions of model fairness (the absence of group-wise model performance disparities) and eliminating such "bias," much less work has gone into rigorously measuring it. In practice, it important to have high quality, human digestible measures of model performance disparities and associated uncertainty quantification about them that can serve as inputs into multi-faceted decision-making processes. In this paper, we show both mathematically and through simulation that many of the metrics used to measure group-wise model performance disparities are themselves statistically biased estimators of the underlying quantities they purport to represent. We argue that this can cause misleading conclusions about the relative group-wise model performance disparities along different dimensions, especially in cases where some sensitive variables consist of categories with few members. We propose the "double-corrected" variance estimator, which provides unbiased estimates and uncertainty quantification of the variance of model performance across groups. It is conceptually simple and easily implementable without statistical software package or numerical optimization. We demonstrate the utility of this approach through simulation and show on a real dataset that while statistically biased estimators of group-wise model performance disparities indicate statistically significant differences, when accounting for statistical bias in the estimator, the estimated between-group disparities are no longer statistically significant.

stat.ME↗

arXiv2022

Measuring and mitigating voting access disparities: a study of race and polling locations in Florida and North Carolina

Voter suppression and associated racial disparities in access to voting are long-standing civil rights concerns in the United States. Barriers to voting have taken many forms over the decades. A history of violent explicit discouragement has shifted to more subtle access limitations that can include long lines and wait times, long travel times to reach a polling station, and other logistical barriers to voting. Our focus in this work is on quantifying disparities in voting access pertaining to the overall time-to-vote, and how they could be remedied via a better choice of polling location or provisioning more sites where voters can cast ballots. However, appropriately calibrating access disparities is difficult because of the need to account for factors such as population density and different community expectations for reasonable travel times. In this paper, we quantify access to polling locations, developing a methodology for the calibrated measurement of racial disparities in polling location "load" and distance to polling locations. We apply this methodology to a study of real-world data from Florida and North Carolina to identify disparities in voting access from the 2020 election. We also introduce algorithms, with modifications to handle scale, that can reduce these disparities by suggesting new polling locations from a given list of identified public locations (including schools and libraries). Applying these algorithms on the 2020 election location data also helps to expose and explore tradeoffs between the cost of allocating more polling locations and the potential impact on access disparities. The developed voting access measurement methodology and algorithmic remediation technique is a first step in better polling location assignment.

cs.CY↗

arXiv2022

Characterizing patterns in police stops by race in Minneapolis from 2016-2021

The murder of George Floyd centered Minneapolis, Minnesota, in conversations on racial injustice in the US. We leverage open data from the Minneapolis Police Department to analyze individual, geographic, and temporal patterns in more than 170,000 police stops since 2016. We evaluate person and vehicle searches at the individual level by race using generalized estimating equations with neighborhood clustering, directly addressing neighborhood differences in police activity. Minneapolis exhibits clear patterns of disproportionate policing by race, wherein Black people are searched at higher rates compared to White people. Temporal visualizations indicate that police stops declined following the murder of George Floyd. This analysis provides contemporary evidence on the state of policing for a major metropolitan area in the United States.

stat.AP↗

arXiv2022

It's COMPASlicated: The Messy Relationship between RAI Datasets and Algorithmic Fairness Benchmarks

Risk assessment instrument (RAI) datasets, particularly ProPublica's COMPAS dataset, are commonly used in algorithmic fairness papers due to benchmarking practices of comparing algorithms on datasets used in prior work. In many cases, this data is used as a benchmark to demonstrate good performance without accounting for the complexities of criminal justice (CJ) processes. However, we show that pretrial RAI datasets can contain numerous measurement biases and errors, and due to disparities in discretion and deployment, algorithmic fairness applied to RAI datasets is limited in making claims about real-world outcomes. These reasons make the datasets a poor fit for benchmarking under assumptions of ground truth and real-world impact. Furthermore, conventional practices of simply replicating previous data experiments may implicitly inherit or edify normative positions without explicitly interrogating value-laden assumptions. Without context of how interdisciplinary fields have engaged in CJ research and context of how RAIs operate upstream and downstream, algorithmic fairness practices are misaligned for meaningful contribution in the context of CJ, and would benefit from transparent engagement with normative considerations and values related to fairness, justice, and equality. These factors prompt questions about whether benchmarks for intrinsically socio-technical systems like the CJ system can exist in a beneficial and ethical way.

cs.CY↗

arXiv2022

Measuring Disparate Outcomes of Content Recommendation Algorithms with Distributional Inequality Metrics

The harmful impacts of algorithmic decision systems have recently come into focus, with many examples of systems such as machine learning (ML) models amplifying existing societal biases. Most metrics attempting to quantify disparities resulting from ML algorithms focus on differences between groups, dividing users based on demographic identities and comparing model performance or overall outcomes between these groups. However, in industry settings, such information is often not available, and inferring these characteristics carries its own risks and biases. Moreover, typical metrics that focus on a single classifier's output ignore the complex network of systems that produce outcomes in real-world settings. In this paper, we evaluate a set of metrics originating from economics, distributional inequality metrics, and their ability to measure disparities in content exposure in a production recommendation system, the Twitter algorithmic timeline. We define desirable criteria for metrics to be used in an operational setting, specifically by ML practitioners. We characterize different types of engagement with content on Twitter using these metrics, and use these results to evaluate the metrics with respect to the desired criteria. We show that we can use these metrics to identify content suggestion algorithms that contribute more strongly to skewed outcomes between users. Overall, we conclude that these metrics can be useful tools for understanding disparate outcomes in online social networks.

cs.CY↗

arXiv2021

The 2021 RecSys Challenge Dataset: Fairness is not optional

After the success the RecSys 2020 Challenge, we are describing a novel and bigger dataset that was released in conjunction with the ACM RecSys Challenge 2021. This year's dataset is not only bigger (~ 1B data points, a 5 fold increase), but for the first time it take into consideration fairness aspects of the challenge. Unlike many static datsets, a lot of effort went into making sure that the dataset was synced with the Twitter platform: if a user deleted their content, the same content would be promptly removed from the dataset too. In this paper, we introduce the dataset and challenge, highlighting some of the issues that arise when creating recommender systems at Twitter scale.

cs.SI↗

arXiv2021

Closer than they appear: A Bayesian perspective on individual-level heterogeneity in risk assessment

Risk assessment instruments are used across the criminal justice system to estimate the probability of some future behavior given covariates. The estimated probabilities are then used in making decisions at the individual level. In the past, there has been controversy about whether the probabilities derived from group-level calculations can meaningfully be applied to individuals. Using Bayesian hierarchical models applied to a large longitudinal dataset from the court system in the state of Kentucky, we analyze variation in individual-level probabilities of failing to appear for court and the extent to which it is captured by covariates. We find that individuals within the same risk group vary widely in their probability of the outcome. In practice, this means that allocating individuals to risk groups based on standard approaches to risk assessment, in large part, results in creating distinctions among individuals who are not meaningfully different in terms of their likelihood of the outcome. This is because uncertainty about the probability that any particular individual will fail to appear is large relative to the difference in average probabilities among any reasonable set of risk groups.

stat.AP↗

arXiv2020

Prediction-Based Decisions and Fairness: A Catalogue of Choices, Assumptions, and Definitions

A recent flurry of research activity has attempted to quantitatively define "fairness" for decisions based on statistical and machine learning (ML) predictions. The rapid growth of this new field has led to wildly inconsistent terminology and notation, presenting a serious challenge for cataloguing and comparing definitions. This paper attempts to bring much-needed order. First, we explicate the various choices and assumptions made---often implicitly---to justify the use of prediction-based decisions. Next, we show how such choices and assumptions can raise concerns about fairness and we present a notationally consistent catalogue of fairness definitions from the ML literature. In doing so, we offer a concise reference for thinking through the choices, assumptions, and fairness considerations of prediction-based decision systems.

stat.AP↗