Research Papers

Interactive Explanations by Conflict Resolution via Argumentative Exchanges

Antonio Rago, Hengzhi Li, Francesca Toni

As the field of explainable AI (XAI) is maturing, calls for interactive explanations for (the outputs of) AI models are growing, but the state-of-the-art predominantly focuses on static explanations. In this paper, we focus instead on interactive explanations framed as conflict resolution between agents (i.e. AI models and/or humans) by leveraging on computational argumentation. Specifically, we define Argumentative eXchanges (AXs) for dynamically sharing, in multi-agent systems, information harboured in individual agents' quantitative bipolar argumentation frameworks towards resolving conflicts amongst the agents. We then deploy AXs in the XAI setting in which a machine and a human interact about the machine's predictions. We identify and assess several theoretical properties characterising AXs that are suitable for XAI. Finally, we instantiate AXs for XAI by defining various agent behaviours, e.g. capturing counterfactual patterns of reasoning in machines and highlighting the effects of cognitive biases in humans. We show experimentally (in a simulated environment) the comparative advantages of these behaviours in terms of conflict resolution, and show that the strongest argument may not always be the most effective.

The right to be forgotten (RTBF) is motivated by the desire of people not to be perpetually disadvantaged by their past deeds. For this, data deletion needs to be deep and permanent, and should be removed from machine learning models. Researchers have proposed machine unlearning algorithms which aim to erase specific data from trained models more efficiently. However, these methods modify how data is fed into the model and how training is done, which may subsequently compromise AI ethics from the fairness perspective. To help software engineers make responsible decisions when adopting these unlearning methods, we present the first study on machine unlearning methods to reveal their fairness implications. We designed and conducted experiments on two typical machine unlearning methods (SISA and AmnesiacML) along with a retraining method (ORTR) as baseline using three fairness datasets under three different deletion strategies. Experimental results show that under non-uniform data deletion, SISA leads to better fairness compared with ORTR and AmnesiacML, while initial training and uniform data deletion do not necessarily affect the fairness of all three methods. These findings have exposed an important research problem in software engineering, and can help practitioners better understand the potential trade-offs on fairness when considering solutions for RTBF.

There are synergies of research interests and industrial efforts in modeling fairness and correcting algorithmic bias in machine learning. In this paper, we present a scalable algorithm for spectral clustering (SC) with group fairness constraints. Group fairness is also known as statistical parity where in each cluster, each protected group is represented with the same proportion as in the entirety. While FairSC algorithm (Kleindessner et al., 2019) is able to find the fairer clustering, it is compromised by high costs due to the kernels of computing nullspaces and the square roots of dense matrices explicitly. We present a new formulation of underlying spectral computation by incorporating nullspace projection and Hotelling's deflation such that the resulting algorithm, called s-FairSC, only involves the sparse matrix-vector products and is able to fully exploit the sparsity of the fair SC model. The experimental results on the modified stochastic block model demonstrate that s-FairSC is comparable with FairSC in recovering fair clustering. Meanwhile, it is sped up by a factor of 12 for moderate model sizes. s-FairSC is further demonstrated to be scalable in the sense that the computational costs of s-FairSC only increase marginally compared to the SC without fairness constraints.

Mimetic Models: Ethical Implications of AI that Acts Like You

Reid McIlroy-Young, Jon Kleinberg, Siddhartha Sen +2

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.

A Principles-based Ethics Assurance Argument Pattern for AI and Autonomous Systems

Zoe Porter, Ibrahim Habli, John McDermid +1

An assurance case is a structured argument, typically produced by safety engineers, to communicate confidence that a critical or complex system, such as an aircraft, will be acceptably safe within its intended context. Assurance cases often inform third party approval of a system. One emerging proposition within the trustworthy AI and autonomous systems (AI/AS) research community is to use assurance cases to instil justified confidence that specific AI/AS will be ethically acceptable when operational in well-defined contexts. This paper substantially develops the proposition and makes it concrete. It brings together the assurance case methodology with a set of ethical principles to structure a principles-based ethics assurance argument pattern. The principles are justice, beneficence, non-maleficence, and respect for human autonomy, with the principle of transparency playing a supporting role. The argument pattern, shortened to the acronym PRAISE, is described. The objective of the proposed PRAISE argument pattern is to provide a reusable template for individual ethics assurance cases, by which engineers, developers, operators, or regulators could justify, communicate, or challenge a claim about the overall ethical acceptability of the use of a specific AI/AS in a given socio-technical context. We apply the pattern to the hypothetical use case of an autonomous robo-taxi service in a city centre.

Impossibility Results in AI: A Survey

Mario Brcic, Roman V. Yampolskiy

An impossibility theorem demonstrates that a particular problem or set of problems cannot be solved as described in the claim. Such theorems put limits on what is possible to do concerning artificial intelligence, especially the super-intelligent one. As such, these results serve as guidelines, reminders, and warnings to AI safety, AI policy, and governance researchers. These might enable solutions to some long-standing questions in the form of formalizing theories in the framework of constraint satisfaction without committing to one option. We strongly believe this to be the most prudent approach to long-term AI safety initiatives. In this paper, we have categorized impossibility theorems applicable to AI into five mechanism-based categories: deduction, indistinguishability, induction, tradeoffs, and intractability. We found that certain theorems are too specific or have implicit assumptions that limit application. Also, we added new results (theorems) such as the unfairness of explainability, the first explainability-related result in the induction category. The remaining results deal with misalignment between the clones and put a limit to the self-awareness of agents. We concluded that deductive impossibilities deny 100%-guarantees for security. In the end, we give some ideas that hold potential in explainability, controllability, value alignment, ethics, and group decision-making. They can be deepened by further investigation.

A Few Good Counterfactuals: Generating Interpretable, Plausible and Diverse Counterfactual Explanations

Barry Smyth, Mark T Keane

Counterfactual explanations provide a potentially significant solution to the Explainable AI (XAI) problem, but good, native counterfactuals have been shown to rarely occur in most datasets. Hence, the most popular methods generate synthetic counterfactuals using blind perturbation. However, such methods have several shortcomings: the resulting counterfactuals (i) may not be valid data-points (they often use features that do not naturally occur), (ii) may lack the sparsity of good counterfactuals (if they modify too many features), and (iii) may lack diversity (if the generated counterfactuals are minimal variants of one another). We describe a method designed to overcome these problems, one that adapts native counterfactuals in the original dataset, to generate sparse, diverse synthetic counterfactuals from naturally occurring features. A series of experiments are reported that systematically explore parametric variations of this novel method on common datasets to establish the conditions for optimal performance.

Neuro-symbolic Architectures for Context Understanding

Alessandro Oltramari, Jonathan Francis, Cory Henson +2

Computational context understanding refers to an agent's ability to fuse disparate sources of information for decision-making and is, therefore, generally regarded as a prerequisite for sophisticated machine reasoning capabilities, such as in artificial intelligence (AI). Data-driven and knowledge-driven methods are two classical techniques in the pursuit of such machine sense-making capability. However, while data-driven methods seek to model the statistical regularities of events by making observations in the real-world, they remain difficult to interpret and they lack mechanisms for naturally incorporating external knowledge. Conversely, knowledge-driven methods, combine structured knowledge bases, perform symbolic reasoning based on axiomatic principles, and are more interpretable in their inferential processing; however, they often lack the ability to estimate the statistical salience of an inference. To combat these issues, we propose the use of hybrid AI methodology as a general framework for combining the strengths of both approaches. Specifically, we inherit the concept of neuro-symbolism as a way of using knowledge-bases to guide the learning progress of deep neural networks. We further ground our discussion in two applications of neuro-symbolism and, in both cases, show that our systems maintain interpretability while achieving comparable performance, relative to the state-of-the-art.

As deep reinforcement learning driven by visual perception becomes more widely used there is a growing need to better understand and probe the learned agents. Understanding the decision making process and its relationship to visual inputs can be very valuable to identify problems in learned behavior. However, this topic has been relatively under-explored in the research community. In this work we present a method for synthesizing visual inputs of interest for a trained agent. Such inputs or states could be situations in which specific actions are necessary. Further, critical states in which a very high or a very low reward can be achieved are often interesting to understand the situational awareness of the system as they can correspond to risky states. To this end, we learn a generative model over the state space of the environment and use its latent space to optimize a target function for the state of interest. In our experiments we show that this method can generate insights for a variety of environments and reinforcement learning methods. We explore results in the standard Atari benchmark games as well as in an autonomous driving simulator. Based on the efficiency with which we have been able to identify behavioural weaknesses with this technique, we believe this general approach could serve as an important tool for AI safety applications.

Finetuning on narrow domains has become an essential tool to adapt Large Language Models (LLMs) to specific tasks and to create models with known unusual properties that are useful for research. We show that narrow finetuning creates strong biases in LLM activations that can be interpreted to understand the finetuning domain. These biases can be discovered using simple tools from model diffing - the study of differences between models before and after finetuning. In particular, analyzing activation differences on the first few tokens of random text and steering by adding this difference to the model activations produces text similar to the format and general content of the finetuning data. We demonstrate that these analyses contain crucial information by creating an LLM-based interpretability agent to understand the finetuning domain. With access to the bias, the agent performs significantly better compared to baseline agents using simple prompting. Our analysis spans synthetic document finetuning for false facts, emergent misalignment, subliminal learning, and taboo word guessing game models across different architectures (Gemma, LLaMA, Qwen) and scales (1B to 32B parameters). We suspect these biases reflect overfitting and find that mixing pretraining data into the finetuning corpus largely removes them, though residual risks may remain. Our work (1) demonstrates that narrowly finetuned models have salient traces of their training objective in their activations and suggests ways to improve how they are trained, (2) warns AI safety and interpretability researchers that the common practice of using such models as a proxy for studying broader finetuning (e.g., chat-tuning) might not be realistic, and (3) highlights the need for deeper investigation into the effects of narrow finetuning and development of truly realistic case studies for model-diffing, safety and interpretability research.

A Unified Framework for Human AI Collaboration in Security Operations Centers with Trusted Autonomy

Ahmad Mohsin, Helge Janicke, Ahmed Ibrahim +2

This article presents a structured framework for Human-AI collaboration in Security Operations Centers (SOCs), integrating AI autonomy, trust calibration, and Human-in-the-loop decision making. Existing frameworks in SOCs often focus narrowly on automation, lacking systematic structures to manage human oversight, trust calibration, and scalable autonomy with AI. Many assume static or binary autonomy settings, failing to account for the varied complexity, criticality, and risk across SOC tasks considering Humans and AI collaboration. To address these limitations, we propose a novel autonomy tiered framework grounded in five levels of AI autonomy from manual to fully autonomous, mapped to Human-in-the-Loop (HITL) roles and task-specific trust thresholds. This enables adaptive and explainable AI integration across core SOC functions, including monitoring, protection, threat detection, alert triage, and incident response. The proposed framework differentiates itself from previous research by creating formal connections between autonomy, trust, and HITL across various SOC levels, which allows for adaptive task distribution according to operational complexity and associated risks. The framework is exemplified through a simulated cyber range that features the cybersecurity AI-Avatar, a fine-tuned LLM-based SOC assistant. The AI-Avatar case study illustrates human-AI collaboration for SOC tasks, reducing alert fatigue, enhancing response coordination, and strategically calibrating trust. This research systematically presents both the theoretical and practical aspects and feasibility of designing next-generation cognitive SOCs that leverage AI not to replace but to enhance human decision-making.

In recent years, AI red teaming has emerged as a practice for probing the safety and security of generative AI systems. Due to the nascency of the field, there are many open questions about how red teaming operations should be conducted. Based on our experience red teaming over 100 generative AI products at Microsoft, we present our internal threat model ontology and eight main lessons we have learned: 1. Understand what the system can do and where it is applied 2. You don't have to compute gradients to break an AI system 3. AI red teaming is not safety benchmarking 4. Automation can help cover more of the risk landscape 5. The human element of AI red teaming is crucial 6. Responsible AI harms are pervasive but difficult to measure 7. LLMs amplify existing security risks and introduce new ones 8. The work of securing AI systems will never be complete By sharing these insights alongside case studies from our operations, we offer practical recommendations aimed at aligning red teaming efforts with real world risks. We also highlight aspects of AI red teaming that we believe are often misunderstood and discuss open questions for the field to consider.