Usman Gohar

Mubashara Akhtar, Anka Reuel, Prajna Soni et al. · meta-ai

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Usman Gohar, Sumon Biswas, Hridesh Rajan

Machine Learning (ML) software has been widely adopted in modern society, with reported fairness implications for minority groups based on race, sex, age, etc. Many recent works have proposed methods to measure and mitigate algorithmic bias in ML models. The existing approaches focus on single classifier-based ML models. However, real-world ML models are often composed of multiple independent or dependent learners in an ensemble (e.g., Random Forest), where the fairness composes in a non-trivial way. How does fairness compose in ensembles? What are the fairness impacts of the learners on the ultimate fairness of the ensemble? Can fair learners result in an unfair ensemble? Furthermore, studies have shown that hyperparameters influence the fairness of ML models. Ensemble hyperparameters are more complex since they affect how learners are combined in different categories of ensembles. Understanding the impact of ensemble hyperparameters on fairness will help programmers design fair ensembles. Today, we do not understand these fully for different ensemble algorithms. In this paper, we comprehensively study popular real-world ensembles: bagging, boosting, stacking and voting. We have developed a benchmark of 168 ensemble models collected from Kaggle on four popular fairness datasets. We use existing fairness metrics to understand the composition of fairness. Our results show that ensembles can be designed to be fairer without using mitigation techniques. We also identify the interplay between fairness composition and data characteristics to guide fair ensemble design. Finally, our benchmark can be leveraged for further research on fair ensembles. To the best of our knowledge, this is one of the first and largest studies on fairness composition in ensembles yet presented in the literature.

Usman Gohar, Michael C. Hunter, Salil Purandare et al.

Safety assurance cases provide structured justifications that safety-critical systems meet their safety requirements. Recently, the notion of defeaters has emerged as a rigorous means of challenging the validity of safety arguments. Examples of defeaters might include overly strict claims, unreliable evidence, or reasoning gaps. However, defeaters remain ad hoc, lack structured support for critical reflection, are inconsistently described, are difficult to review, and lack documentation standards. To address this, we propose Defeater Cards, a new structured documentation artifact for systematically characterizing, reasoning about, and managing defeaters in safety cases. Drawing on a literature survey and thematic analysis, we identify documentation criteria that inform the card's structure, based on the 5W1H framework. Defeater Cards are designed to support informed analysis and evolution, improve traceability and auditability, and enable the reuse of defeater knowledge across systems and product variants. We demonstrate their applicability through two cross-domain case studies, showing how they expose hidden assumptions, surface reasoning gaps, and support ongoing safety assurance case evolution. To support adoption and community reuse, we also release an open-source repository of defeater cards as a baseline upon which researchers and practitioners can build and describe lessons learned.

Usman Gohar, Michael C. Hunter, Robyn R. Lutz et al.

Constructing assurance cases is a widely used, and sometimes required, process toward demonstrating that safety-critical systems will operate safely in their planned environment. To mitigate the risk of errors and missing edge cases, the concept of defeaters - arguments or evidence that challenge claims in an assurance case - has been introduced. Defeaters can provide timely detection of weaknesses in the arguments, prompting further investigation and timely mitigations. However, capturing defeaters relies on expert judgment, experience, and creativity and must be done iteratively due to evolving requirements and regulations. This paper proposes CoDefeater, an automated process to leverage large language models (LLMs) for finding defeaters. Initial results on two systems show that LLMs can efficiently find known and unforeseen feasible defeaters to support safety analysts in enhancing the completeness and confidence of assurance cases.

Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed et al. · deepmind, oxford

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Gemma Galdon Clavell, Pablo Accuosto, Usman Gohar

The rapid deployment of AI systems across high-stakes domains has created urgent demand for standardized evaluation, yet the field remains fragmented across competing risk taxonomies that catalog risks without showing how an audit is executed. At least 74 AI risk taxonomies exist, and almost all stop at the catalog. The hard part of auditing is not naming a risk but operationalizing it: turning it into a test run against a real system, a measured value, a calibrated severity, and a defensible grade. This paper leads with that bridge. We present the operationalization layer Eticas has built and run, shown end to end on a single risk (PII leakage) against a public benchmark, and then the open taxonomy that makes the method scale. On GPT-4-0314, a disclosure risk that seven external frameworks require be controlled is measured at 0%, 51%, and 84% disclosure as adversarial conditioning increases, mapping through calibrated severity bands to a subcategory grade of E with a SYSTEMIC pattern. Around this example, the Eticas AI Risk Taxonomy v2.0.0 organizes 76 active subcategories across 10 categories and 20 sub-groups, with mappings to 18 external frameworks across compliance, reference, and academic tiers. Its category and sub-group layer is published under CC BY 4.0 as open semantic infrastructure with stable URIs and SKOS/JSON-LD distributions, and a worked subcategory example shows the operational layer down to its severity thresholds. The contribution is the demonstrated bridge from concept to graded finding, anchored by a clean separation of risks from the mechanisms by which they surface, and framed by an open-core model in which the conceptual scaffold is open and the methodology calibration is the practitioner layer. This is the infrastructure the AI auditing field needs: shared, open, and demonstrably operable.

Usman Gohar, Michael C. Hunter, Agnieszka Marczak-Czajka et al.

Small Unmanned Aircraft Systems (sUAS) have gained widespread adoption across a diverse range of applications. This has introduced operational complexities within shared airspaces and an increase in reported incidents, raising safety concerns. In response, the U.S. Federal Aviation Administration (FAA) is developing a UAS Traffic Management (UTM) system to control access to airspace based on an sUAS's predicted ability to safely complete its mission. However, a fully automated system capable of swiftly approving or denying flight requests can be prone to bias and must consider safety, transparency, and fairness to diverse stakeholders. In this paper, we present an initial study that explores stakeholders' perspectives on factors that should be considered in an automated system. Results indicate flight characteristics and environmental conditions were perceived as most important but pilot and drone capabilities should also be considered. Further, several respondents indicated an aversion to any AI-supported automation, highlighting the need for full transparency in automated decision-making. Results provide a societal perspective on the challenges of automating UTM flight authorization decisions and help frame the ongoing design of a solution acceptable to the broader sUAS community.

Jan Batzner, Sree Harsha Nelaturu, Anastassia Kornilova et al.

AI evaluations are widely used for testing and understanding progress. However, the diverse evaluators bring with them inconsistencies that challenge analysis and comparison. First, results are saved in incompatible formats, scattered across leaderboards, papers, blog posts, evaluation harness logs, and custom repositories. Second, results are created by different evaluation frameworks, which produce divergent scores for nominally identical evaluations and record metadata inconsistently, hindering comparison, cross-community evaluation science, cost reduction, and reuse. We introduce Every Eval Ever, the first shared schema and community-crowdsourced repository for AI evaluation results. The schema standardizes how evaluations are represented in a unified, single JSON document. It is source-agnostic by design, ingesting results from evaluation harnesses and papers alike, and optionally stores per-instance outputs for fine-grained analysis. We contribute: (i) a community-governed metadata schema with a companion instance-level schema, the first standardization effort of its kind; (ii) automatic converters from popular formats, evaluation harnesses, and leaderboards to the unified schema; and (iii) a crowdsourced community database hosted on Hugging Face, currently spanning to date 22,235 models, 2,273 unique benchmarks, and 31 evaluation formats.

Prakhar Ganesh, Usman Gohar, Lu Cheng et al.

With fairness concerns gaining significant attention in Machine Learning (ML), several bias mitigation techniques have been proposed, often compared against each other to find the best method. These benchmarking efforts tend to use a common setup for evaluation under the assumption that providing a uniform environment ensures a fair comparison. However, bias mitigation techniques are sensitive to hyperparameter choices, random seeds, feature selection, etc., meaning that comparison on just one setting can unfairly favour certain algorithms. In this work, we show significant variance in fairness achieved by several algorithms and the influence of the learning pipeline on fairness scores. We highlight that most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization, suggesting that the choice of the evaluation parameters-rather than the mitigation technique itself-can sometimes create the perceived superiority of one method over another. We hope our work encourages future research on how various choices in the lifecycle of developing an algorithm impact fairness, and trends that guide the selection of appropriate algorithms.

Usman Gohar, Zeyu Tang, Jialu Wang et al.

The widespread integration of Machine Learning systems in daily life, particularly in high-stakes domains, has raised concerns about the fairness implications. While prior works have investigated static fairness measures, recent studies reveal that automated decision-making has long-term implications and that off-the-shelf fairness approaches may not serve the purpose of achieving long-term fairness. Additionally, the existence of feedback loops and the interaction between models and the environment introduces additional complexities that may deviate from the initial fairness goals. In this survey, we review existing literature on long-term fairness from different perspectives and present a taxonomy for long-term fairness studies. We highlight key challenges and consider future research directions, analyzing both current issues and potential further explorations.

Usman Gohar, Lu Cheng

The widespread adoption of Machine Learning systems, especially in more decision-critical applications such as criminal sentencing and bank loans, has led to increased concerns about fairness implications. Algorithms and metrics have been developed to mitigate and measure these discriminations. More recently, works have identified a more challenging form of bias called intersectional bias, which encompasses multiple sensitive attributes, such as race and gender, together. In this survey, we review the state-of-the-art in intersectional fairness. We present a taxonomy for intersectional notions of fairness and mitigation. Finally, we identify the key challenges and provide researchers with guidelines for future directions.