If you were looking for an inkblot test for your view of big tech's investment in artificial intelligence, you could hardly do better than the news that Google is ordering the construction of at least six small nuclear reactors to power the growth of the technology. Here, in one view, is an enlightened business leveraging its size to invest in infrastructure that could change the world for the better. Here, in another, is a poorly regulated corporation ignoring democratic objections in the brutal race for control of an innovation with great potential to do harm – and leaving the rest of us with little say in its development. Google is making this eye-catching move because the datacentres that power the explosive growth of generative AI consume huge amounts of electricity – more than the existing grid in the US or other western nations can readily supply. For today's newsletter, I spoke to technology journalist Chris Stokel-Walker, author of How AI Ate the World, about why the demand for power is growing so quickly – and whether we can trust big tech to handle the consequences.

Large language models excel at performing inference over text to extract information, summarize information, or generate additional text. These inference capabilities are implicated in a variety of ethical harms spanning surveillance, labor displacement, and IP/copyright theft. While many policy, legal, and technical mitigations have been proposed to counteract these harms, these mitigations typically require cooperation from institutions that move slower than technical advances (i.e., governments) or that have few incentives to act to counteract these harms (i.e., the corporations that create and profit from these LLMs). In this paper, we define and build "data defenses" -- a novel strategy that directly empowers data owners to block LLMs from performing inference on their data. We create data defenses by developing a method to automatically generate adversarial prompt injections that, when added to input text, significantly reduce the ability of LLMs to accurately infer personally identifying information about the subject of the input text or to use copyrighted text in inference. We examine the ethics of enabling such direct resistance to LLM inference, and argue that making data defenses that resist and subvert LLMs enables the realization of important values such as data ownership, data sovereignty, and democratic control over AI systems. We verify that our data defenses are cheap and fast to generate, work on the latest commercial and open-source LLMs, resistance to countermeasures, and are robust to several different attack settings. Finally, we consider the security implications of LLM data defenses and outline several future research directions in this area. Our code is available at https://github.com/wagnew3/LLMDataDefenses and a tool for using our defenses to protect text against LLM inference is at https://wagnew3.github.io/LLM-Data-Defenses/.

This project, named HEnRY, aims to introduce a Multi-Agent System (MAS) into Intesa Sanpaolo. The name HEnRY summarizes the project's core principles: the Hierarchical organization of agents in a layered structure for efficient resource management; Efficient optimization of resources and operations to enhance overall performance; Reactive ability of agents to quickly respond to environmental stimuli; and Yielding adaptability and flexibility of agents to handle unexpected situations. The discussion covers two distinct research paths: the first focuses on the system architecture, and the second on the collaboration between agents. This work is not limited to the specific structure of the Intesa Sanpaolo context; instead, it leverages existing research in MAS to introduce a new solution. Since Intesa Sanpaolo is organized according to a model that aligns with international corporate governance best practices, this approach could also be relevant to similar scenarios.

Chatbots like ChatGPT are used for diverse purposes, ranging from resume writing to entertainment. These real-world applications are different from the institutional uses, such as resume screening or credit scoring, which have been the focus of much of AI research on fairness. Ensuring equitable treatment for all users in these first-person contexts is critical. In this work, we study "first-person fairness," which means fairness toward the chatbot user. This includes providing high-quality responses to all users regardless of their identity or background and avoiding harmful stereotypes. We propose a scalable, privacy-preserving method for evaluating one aspect of first-person fairness across a large, heterogeneous corpus of real-world chatbot interactions. Specifically, we assess potential bias linked to users' names, which can serve as proxies for demographic attributes like gender or race, in chatbot systems such as ChatGPT, which provide mechanisms for storing and using user names. Our method leverages a second language model to privately analyze name-sensitivity in the chatbot's responses. We verify the validity of these annotations through independent human evaluation. Further, we show that post-training interventions, including RL, significantly mitigate harmful stereotypes. Our approach also yields succinct descriptions of response differences across tasks. For instance, in the "writing a story" task, chatbot responses show a tendency to create protagonists whose gender matches the likely gender inferred from the user's name. Moreover, a pattern emerges where users with female-associated names receive responses with friendlier and simpler language slightly more often than users with male-associated names. Finally, we provide the system messages required for external researchers to further investigate ChatGPT's behavior with hypothetical user profiles.

As large language models (LLMs) continue to advance, ensuring their alignment with human values becomes increasingly critical. Traditional alignment methods heavily rely on human feedback to fine-tune models. With the emergence of superhuman models whose outputs may surpass human understanding, evaluating and aligning these models using human judgments poses significant challenges. To address the challenges, recent works use weak supervisors to elicit knowledge from much stronger models. However, there are important disanalogies between the empirical setup in the existing works and the genuine goal of alignment. We remark that existing works investigate the phenomenon of weak-to-strong generation in analogous setup (i.e., binary classification), rather than practical alignment-relevant tasks (e.g., safety). In this paper, we bridge this gap by extending weak-to-strong generation to the context of practical alignment. We empirically demonstrate the widespread phenomenon of weak-to-strong generation in three complicated alignment tasks: safety, toxicity, and legal reasoning}. Furthermore, we explore efficient strategies for improving alignment performance to enhance the quality of model outcomes. Lastly, we summarize and analyze the challenges and potential solutions in regard to specific alignment tasks, which we hope to catalyze the research progress on the topic of weak-to-strong generalization. Our code is released at https://github.com/yeruimeng/WTS.git.

With the rapid development of large language models (LLMs), how to efficiently evaluate them has become an important research question. Existing evaluation methods often suffer from high costs, limited test formats, the need of human references, and systematic evaluation biases. To address these limitations, our study introduces the Auto-PRE, an automatic LLM evaluation framework based on peer review. In contrast to previous studies that rely on human annotations, Auto-PRE selects evaluator LLMs automatically based on their inherent traits including consistency, self-confidence, and pertinence. We conduct extensive experiments on three tasks: summary generation, non-factoid question-answering, and dialogue generation. Experimental results indicate our Auto-PRE achieves state-of-the-art performance at a lower cost. Moreover, our study highlights the impact of prompt strategies and evaluation formats on evaluation performance, offering guidance for method optimization in the future.

We present LEGAL-UQA, the first Urdu legal question-answering dataset derived from Pakistan's constitution. This parallel English-Urdu dataset includes 619 question-answer pairs, each with corresponding legal article contexts, addressing the need for domain-specific NLP resources in low-resource languages. We describe the dataset creation process, including OCR extraction, manual refinement, and GPT-4-assisted translation and generation of QA pairs. Our experiments evaluate the latest generalist language and embedding models on LEGAL-UQA, with Claude-3.5-Sonnet achieving 99.19% human-evaluated accuracy. We fine-tune mt5-large-UQA-1.0, highlighting the challenges of adapting multilingual models to specialized domains. Additionally, we assess retrieval performance, finding OpenAI's text-embedding-3-large outperforms Mistral's mistral-embed. LEGAL-UQA bridges the gap between global NLP advancements and localized applications, particularly in constitutional law, and lays the foundation for improved legal information access in Pakistan.

The burgeoning field of Natural Language Processing (NLP) stands at a critical juncture where the integration of fairness within its frameworks has become an imperative. This PhD thesis addresses the need for equity and transparency in NLP systems, recognizing that fairness in NLP is not merely a technical challenge but a moral and ethical necessity, requiring a rigorous examination of how these technologies interact with and impact diverse human populations. Through this lens, this thesis undertakes a thorough investigation into the development of equitable NLP methodologies and the evaluation of biases that prevail in current systems. First, it introduces an innovative algorithm to mitigate biases in multi-class classifiers, tailored for high-risk NLP applications, surpassing traditional methods in both bias mitigation and prediction accuracy. Then, an analysis of the Bios dataset reveals the impact of dataset size on discriminatory biases and the limitations of standard fairness metrics. This awareness has led to explorations in the field of explainable AI, aiming for a more complete understanding of biases where traditional metrics are limited. Consequently, the thesis presents COCKATIEL, a model-agnostic explainability method that identifies and ranks concepts in Transformer models, outperforming previous approaches in sentiment analysis tasks. Finally, the thesis contributes to bridging the gap between fairness and explainability by introducing TaCo, a novel method to neutralize bias in Transformer model embeddings. In conclusion, this thesis constitutes a significant interdisciplinary endeavor that intertwines explicability and fairness to challenge and reshape current NLP paradigms. The methodologies and critiques presented contribute to the ongoing discourse on fairness in machine learning, offering actionable solutions for more equitable and responsible AI systems.

Large Language Models (LLMs) and Vision-Language Models (VLMs) have made significant advancements in a wide range of natural language processing and vision-language tasks. Access to large web-scale datasets has been a key factor in their success. However, concerns have been raised about the unauthorized use of copyrighted materials and potential copyright infringement. Existing methods, such as sample-level Membership Inference Attacks (MIA) and distribution-based dataset inference, distinguish member data (data used for training) and non-member data by leveraging the common observation that models tend to memorize and show greater confidence in member data. Nevertheless, these methods face challenges when applied to LLMs and VLMs, such as the requirement for ground-truth member data or non-member data that shares the same distribution as the test data. In this paper, we propose a novel dataset-level membership inference method based on Self-Comparison. We find that a member prefix followed by a non-member suffix (paraphrased from a member suffix) can further trigger the model's memorization on training data. Instead of directly comparing member and non-member data, we introduce paraphrasing to the second half of the sequence and evaluate how the likelihood changes before and after paraphrasing. Unlike prior approaches, our method does not require access to ground-truth member data or non-member data in identical distribution, making it more practical. Extensive experiments demonstrate that our proposed method outperforms traditional MIA and dataset inference techniques across various datasets and models, including including public models, fine-tuned models, and API-based commercial models.

Generative audio models are rapidly advancing in both capabilities and public utilization -- several powerful generative audio models have readily available open weights, and some tech companies have released high quality generative audio products. Yet, while prior work has enumerated many ethical issues stemming from the data on which generative visual and textual models have been trained, we have little understanding of similar issues with generative audio datasets, including those related to bias, toxicity, and intellectual property. To bridge this gap, we conducted a literature review of hundreds of audio datasets and selected seven of the most prominent to audit in more detail. We found that these datasets are biased against women, contain toxic stereotypes about marginalized communities, and contain significant amounts of copyrighted work. To enable artists to see if they are in popular audio datasets and facilitate exploration of the contents of these datasets, we developed a web tool audio datasets exploration tool at https://audio-audit.vercel.app.