Research Papers

From tools to thieves: Measuring and understanding public perceptions of AI through crowdsourced metaphors

Myra Cheng, Angela Y. Lee, Kristina Rapuano +3

How has the public responded to the increasing prevalence of artificial intelligence (AI)-based technologies? We investigate public perceptions of AI by collecting over 12,000 responses over 12 months from a nationally representative U.S. sample. Participants provided open-ended metaphors reflecting their mental models of AI, a methodology that overcomes the limitations of traditional self-reported measures by capturing more nuance. Using a mixed-methods approach combining quantitative clustering and qualitative coding, we identify 20 dominant metaphors shaping public understanding of AI. To analyze these metaphors systematically, we present a scalable framework integrating language modeling (LM)-based techniques to measure key dimensions of public perception: anthropomorphism (attribution of human-like qualities), warmth, and competence. We find that Americans generally view AI as warm and competent, and that over the past year, perceptions of AI's human-likeness and warmth have significantly increased ($+34\%, r = 0.80, p < 0.01; +41\%, r = 0.62, p < 0.05$). These implicit perceptions, along with the identified dominant metaphors, strongly predict trust in and willingness to adopt AI ($r^2 = 0.21, 0.18, p < 0.001$). Moreover, we uncover systematic demographic differences in metaphors and implicit perceptions, such as the higher propensity of women, older individuals, and people of color to anthropomorphize AI, which shed light on demographic disparities in trust and adoption. In addition to our dataset and framework for tracking evolving public attitudes, we provide actionable insights on using metaphors for inclusive and responsible AI development.

Assessing Adversarial Robustness of Large Language Models: An Empirical Study

Zeyu Yang, Zhao Meng, Xiaochen Zheng +1

Large Language Models (LLMs) have revolutionized natural language processing, but their robustness against adversarial attacks remains a critical concern. We presents a novel white-box style attack approach that exposes vulnerabilities in leading open-source LLMs, including Llama, OPT, and T5. We assess the impact of model size, structure, and fine-tuning strategies on their resistance to adversarial perturbations. Our comprehensive evaluation across five diverse text classification tasks establishes a new benchmark for LLM robustness. The findings of this study have far-reaching implications for the reliable deployment of LLMs in real-world applications and contribute to the advancement of trustworthy AI systems.

As AI systems become increasingly prevalent and impactful, the need for effective AI governance and accountability measures is paramount. This paper examines the AI governance landscape, focusing on Anthropic's Claude, a foundational AI model. We analyze Claude through the lens of the NIST AI Risk Management Framework and the EU AI Act, identifying potential threats and proposing mitigation strategies. The paper highlights the importance of transparency, rigorous benchmarking, and comprehensive data handling processes in ensuring the responsible development and deployment of AI systems. We conclude by discussing the social impact of AI governance and the ethical considerations surrounding AI accountability.

Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection

Konstantinos Tsigos, Evlampios Apostolidis, Spyridon Baxevanakis +2

In this paper we propose a new framework for evaluating the performance of explanation methods on the decisions of a deepfake detector. This framework assesses the ability of an explanation method to spot the regions of a fake image with the biggest influence on the decision of the deepfake detector, by examining the extent to which these regions can be modified through a set of adversarial attacks, in order to flip the detector's prediction or reduce its initial prediction; we anticipate a larger drop in deepfake detection accuracy and prediction, for methods that spot these regions more accurately. Based on this framework, we conduct a comparative study using a state-of-the-art model for deepfake detection that has been trained on the FaceForensics++ dataset, and five explanation methods from the literature. The findings of our quantitative and qualitative evaluations document the advanced performance of the LIME explanation method against the other compared ones, and indicate this method as the most appropriate for explaining the decisions of the utilized deepfake detector.

RadEdit: stress-testing biomedical vision models via diffusion image editing

Fernando Pérez-García, Sam Bond-Taylor, Pedro P. Sanchez +11

Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method RadEdit that uses multiple masks, if present, to constrain changes and ensure consistency in the edited images. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI.

As AI systems proliferate in society, the AI community is increasingly preoccupied with the concept of AI Safety, namely the prevention of failures due to accidents that arise from an unanticipated departure of a system's behavior from designer intent in AI deployment. We demonstrate through an analysis of real world cases of such incidents that although current vocabulary captures a range of the encountered issues of AI deployment, an expanded socio-technical framing will be required for a more complete understanding of how AI systems and implemented safety mechanisms fail and succeed in real life.

Safer-Instruct: Aligning Language Models with Automated Preference Data

Taiwei Shi, Kai Chen, Jieyu Zhao

Reinforcement learning from human feedback (RLHF) is a vital strategy for enhancing model capability in language models. However, annotating preference data for RLHF is a resource-intensive and creativity-demanding process, while existing automatic generation methods face limitations in data diversity and quality. In response, we present Safer-Instruct, a novel pipeline for automatically constructing large-scale preference data. Our approach leverages reversed instruction tuning, instruction induction, and expert model evaluation to efficiently generate high-quality preference data without human annotators. To verify the effectiveness of Safer-Instruct, we apply the pipeline to construct a safety preference dataset as a case study. Finetuning an Alpaca model on this synthetic dataset not only demonstrates improved harmlessness but also outperforms models fine-tuned on human-annotated safety preference data, all the while maintaining a competitive edge in downstream tasks. Importantly, our Safer-Instruct framework is versatile and can be applied to generate preference data across various domains, extending its utility beyond safety preferences. It addresses the challenges in preference data acquisition and advances the development of more capable and responsible AI systems. For dataset and code implementation, see https://github.com/uscnlp-lime/safer-instruct

The AI Incident Database as an Educational Tool to Raise Awareness of AI Harms: A Classroom Exploration of Efficacy, Limitations, & Future Improvements

Michael Feffer, Nikolas Martelaro, Hoda Heidari

Prior work has established the importance of integrating AI ethics topics into computer and data sciences curricula. We provide evidence suggesting that one of the critical objectives of AI Ethics education must be to raise awareness of AI harms. While there are various sources to learn about such harms, The AI Incident Database (AIID) is one of the few attempts at offering a relatively comprehensive database indexing prior instances of harms or near harms stemming from the deployment of AI technologies in the real world. This study assesses the effectiveness of AIID as an educational tool to raise awareness regarding the prevalence and severity of AI harms in socially high-stakes domains. We present findings obtained through a classroom study conducted at an R1 institution as part of a course focused on the societal and ethical considerations around AI and ML. Our qualitative findings characterize students' initial perceptions of core topics in AI ethics and their desire to close the educational gap between their technical skills and their ability to think systematically about ethical and societal aspects of their work. We find that interacting with the database helps students better understand the magnitude and severity of AI harms and instills in them a sense of urgency around (a) designing functional and safe AI and (b) strengthening governance and accountability mechanisms. Finally, we compile students' feedback about the tool and our class activity into actionable recommendations for the database development team and the broader community to improve awareness of AI harms in AI ethics education.

Benchmarking Algorithmic Bias in Face Recognition: An Experimental Approach Using Synthetic Faces and Human Evaluation

Hao Liang, Pietro Perona, Guha Balakrishnan

We propose an experimental method for measuring bias in face recognition systems. Existing methods to measure bias depend on benchmark datasets that are collected in the wild and annotated for protected (e.g., race, gender) and non-protected (e.g., pose, lighting) attributes. Such observational datasets only permit correlational conclusions, e.g., "Algorithm A's accuracy is different on female and male faces in dataset X.". By contrast, experimental methods manipulate attributes individually and thus permit causal conclusions, e.g., "Algorithm A's accuracy is affected by gender and skin color." Our method is based on generating synthetic faces using a neural face generator, where each attribute of interest is modified independently while leaving all other attributes constant. Human observers crucially provide the ground truth on perceptual identity similarity between synthetic image pairs. We validate our method quantitatively by evaluating race and gender biases of three research-grade face recognition models. Our synthetic pipeline reveals that for these algorithms, accuracy is lower for Black and East Asian population subgroups. Our method can also quantify how perceptual changes in attributes affect face identity distances reported by these models. Our large synthetic dataset, consisting of 48,000 synthetic face image pairs (10,200 unique synthetic faces) and 555,000 human annotations (individual attributes and pairwise identity comparisons) is available to researchers in this important area.

Getting from Generative AI to Trustworthy AI: What LLMs might learn from Cyc

Doug Lenat, Gary Marcus

Generative AI, the most popular current approach to AI, consists of large language models (LLMs) that are trained to produce outputs that are plausible, but not necessarily correct. Although their abilities are often uncanny, they are lacking in aspects of reasoning, leading LLMs to be less than completely trustworthy. Furthermore, their results tend to be both unpredictable and uninterpretable. We lay out 16 desiderata for future AI, and discuss an alternative approach to AI which could theoretically address many of the limitations associated with current approaches: AI educated with curated pieces of explicit knowledge and rules of thumb, enabling an inference engine to automatically deduce the logical entailments of all that knowledge. Even long arguments produced this way can be both trustworthy and interpretable, since the full step-by-step line of reasoning is always available, and for each step the provenance of the knowledge used can be documented and audited. There is however a catch: if the logical language is expressive enough to fully represent the meaning of anything we can say in English, then the inference engine runs much too slowly. That's why symbolic AI systems typically settle for some fast but much less expressive logic, such as knowledge graphs. We describe how one AI system, Cyc, has developed ways to overcome that tradeoff and is able to reason in higher order logic in real time. We suggest that any trustworthy general AI will need to hybridize the approaches, the LLM approach and more formal approach, and lay out a path to realizing that dream.

Despite progress in the field, significant parts of current XAI research are still not on solid conceptual, ethical, or methodological grounds. Unfortunately, these unfounded parts are not on the decline but continue to grow. Many explanation techniques are still proposed without clarifying their purpose. Instead, they are advertised with ever more fancy-looking heatmaps or only seemingly relevant benchmarks. Moreover, explanation techniques are motivated with questionable goals, such as building trust, or rely on strong assumptions about the 'concepts' that deep learning algorithms learn. In this paper, we highlight and discuss these and other misconceptions in current XAI research. We also suggest steps to make XAI a more substantive area of research.

Recently, there has been a growing demand for the deployment of Explainable Artificial Intelligence (XAI) algorithms in real-world applications. However, traditional XAI methods typically suffer from a high computational complexity problem, which discourages the deployment of real-time systems to meet the time-demanding requirements of real-world scenarios. Although many approaches have been proposed to improve the efficiency of XAI methods, a comprehensive understanding of the achievements and challenges is still needed. To this end, in this paper we provide a review of efficient XAI. Specifically, we categorize existing techniques of XAI acceleration into efficient non-amortized and efficient amortized methods. The efficient non-amortized methods focus on data-centric or model-centric acceleration upon each individual instance. In contrast, amortized methods focus on learning a unified distribution of model explanations, following the predictive, generative, or reinforcement frameworks, to rapidly derive multiple model explanations. We also analyze the limitations of an efficient XAI pipeline from the perspectives of the training phase, the deployment phase, and the use scenarios. Finally, we summarize the challenges of deploying XAI acceleration methods to real-world scenarios, overcoming the trade-off between faithfulness and efficiency, and the selection of different acceleration methods.