As AI systems are used to answer more difficult questions and potentially help create new knowledge, judging the truthfulness of their outputs becomes more difficult and more important. How can we supervise unreliable experts, which have access to the truth but may not accurately report it, to give answers that are systematically true and don't just superficially seem true, when the supervisor can't tell the difference between the two on their own? In this work, we show that debate between two unreliable experts can help a non-expert judge more reliably identify the truth. We collect a dataset of human-written debates on hard reading comprehension questions where the judge has not read the source passage, only ever seeing expert arguments and short quotes selectively revealed by 'expert' debaters who have access to the passage. In our debates, one expert argues for the correct answer, and the other for an incorrect answer. Comparing debate to a baseline we call consultancy, where a single expert argues for only one answer which is correct half of the time, we find that debate performs significantly better, with 84% judge accuracy compared to consultancy's 74%. Debates are also more efficient, being 68% of the length of consultancies. By comparing human to AI debaters, we find evidence that with more skilled (in this case, human) debaters, the performance of debate goes up but the performance of consultancy goes down. Our error analysis also supports this trend, with 46% of errors in human debate attributable to mistakes by the honest debater (which should go away with increased skill); whereas 52% of errors in human consultancy are due to debaters obfuscating the relevant evidence from the judge (which should become worse with increased skill). Overall, these results show that debate is a promising approach for supervising increasingly capable but potentially unreliable AI systems.

Large Language Models (LLMs) bring transformative benefits alongside unique challenges, including intellectual property (IP) and ethical concerns. This position paper explores a novel angle to mitigate these risks, drawing parallels between LLMs and established web systems. We identify "citation" - the acknowledgement or reference to a source or evidence - as a crucial yet missing component in LLMs. Incorporating citation could enhance content transparency and verifiability, thereby confronting the IP and ethical issues in the deployment of LLMs. We further propose that a comprehensive citation mechanism for LLMs should account for both non-parametric and parametric content. Despite the complexity of implementing such a citation mechanism, along with the potential pitfalls, we advocate for its development. Building on this foundation, we outline several research problems in this area, aiming to guide future explorations towards building more responsible and accountable LLMs.

We present a systematic study and comprehensive evaluation of large language models for automatic multilingual readability assessment. In particular, we construct ReadMe++, a multilingual multi-domain dataset with human annotations of 9757 sentences in Arabic, English, French, Hindi, and Russian collected from 112 different data sources. ReadMe++ offers more domain and language diversity than existing readability datasets, making it ideal for benchmarking multilingual and non-English language models (including mBERT, XLM-R, mT5, Llama-2, GPT-4, etc.) in the supervised, unsupervised, and few-shot prompting settings. Our experiments reveal that models fine-tuned on ReadMe++ outperform those trained on single-domain datasets, showcasing superior performance on multi-domain readability assessment and cross-lingual transfer capabilities. We also compare to traditional readability metrics (such as Flesch-Kincaid Grade Level and Open Source Metric for Measuring Arabic Narratives), as well as the state-of-the-art unsupervised metric RSRS (Martinc et al., 2021). We will make our data and code publicly available at: https://github.com/tareknaous/readme.

Besides its common use cases in epidemiology, political, and social sciences, causality turns out to be crucial in evaluating the fairness of automated decisions, both in a legal and everyday sense. We provide arguments and examples, of why causality is particularly important for fairness evaluation. In particular, we point out the social impact of non-causal predictions and the legal anti-discrimination process that relies on causal claims. We conclude with a discussion about the challenges and limitations of applying causality in practical scenarios as well as possible solutions.

This study presents German FinBERT, a novel pre-trained German language model tailored for financial textual data. The model is trained through a comprehensive pre-training process, leveraging a substantial corpus comprising financial reports, ad-hoc announcements and news related to German companies. The corpus size is comparable to the data sets commonly used for training standard BERT models. I evaluate the performance of German FinBERT on downstream tasks, specifically sentiment prediction, topic recognition and question answering against generic German language models. My results demonstrate improved performance on finance-specific data, indicating the efficacy of German FinBERT in capturing domain-specific nuances. The presented findings suggest that German FinBERT holds promise as a valuable tool for financial text analysis, potentially benefiting various applications in the financial domain.

Leonid Tymchenko spent the first month of Russia's invasion sitting in his dark government office after curfew. Unable to go home, Ukraine's Deputy Minister of Internal Affairs scrolled through Telegram, looking at thousands of videos and images of advancing Russian soldiers. When Tymchenko was offered a chance to test a new facial-recognition tool, he uploaded some of the photos to try it out. He could not believe the results. Every time Tymchenko added a photo of a Russian soldier, the software, made by the American facial-recognition company Clearview AI, seemed to come back with an exact hit, linking to pages that revealed the soldier's name, hometown, and social-media profile.

In 2006, the U.S. patent office received a filing for "an automatically generated display that contains information relevant to a user about another user of a social network." Rather than forcing people to search through disorganized" content for items of interest, the system would seek to generate a list of "relevant" information in a "preferred order." The listed authors were "Zuckerberg et al." and the product was the News Feed. The platform's recommendation systems were still in their infancy, and as an algorithm, it wasn't much. By 2010, the company was looking beyond the crude system to recommend content based on machine learning and user behavior.

We investigate a formalism for the conditions of a successful explanation of AI. We consider "success" to depend not only on what information the explanation contains, but also on what information the human explainee understands from it. Theory of mind literature discusses the folk concepts that humans use to understand and generalize behavior. We posit that folk concepts of behavior provide us with a "language" that humans understand behavior with. We use these folk concepts as a framework of social attribution by the human explainee--the information constructs that humans are likely to comprehend from explanations--by introducing a blueprint for an explanatory narrative (Figure 1) that explains AI behavior with these constructs. We then demonstrate that many XAI methods today can be mapped to folk concepts of behavior in a qualitative evaluation. This allows us to uncover their failure modes that prevent current methods from explaining successfully--i.e., the information constructs that are missing for any given XAI method, and whose inclusion can decrease the likelihood of misunderstanding AI behavior.

Variance in predictions across different trained models is a significant, under-explored source of error in fair binary classification. In practice, the variance on some data examples is so large that decisions can be effectively arbitrary. To investigate this problem, we take an experimental approach and make four overarching contributions: We: 1) Define a metric called self-consistency, derived from variance, which we use as a proxy for measuring and reducing arbitrariness; 2) Develop an ensembling algorithm that abstains from classification when a prediction would be arbitrary; 3) Conduct the largest to-date empirical study of the role of variance (vis-a-vis self-consistency and arbitrariness) in fair binary classification; and, 4) Release a toolkit that makes the US Home Mortgage Disclosure Act (HMDA) datasets easily usable for future research. Altogether, our experiments reveal shocking insights about the reliability of conclusions on benchmark datasets. Most fair binary classification benchmarks are close-to-fair when taking into account the amount of arbitrariness present in predictions -- before we even try to apply any fairness interventions. This finding calls into question the practical utility of common algorithmic fairness methods, and in turn suggests that we should reconsider how we choose to measure fairness in binary classification.

Reinforcement Learning from Human Feedback (RLHF) is a vital strategy for enhancing model safety in language models. However, annotating preference data for RLHF is a resource-intensive and creativity-demanding process, while automatic generation methods face limitations in data diversity and quality. In response, we present Safer-Instruct, a novel pipeline for semi-automatically constructing large-scale preference datasets. Our approach leverages reversed instruction tuning, instruction induction, and expert model evaluation to efficiently generate high-quality preference data without human annotators. We evaluate Safer-Instruct using LLaMA for instruction induction and GPT-4 as an expert model, generating approximately 10K preference samples. Finetuning an Alpaca model on this dataset demonstrates improved harmlessness while maintaining competitive performance on conversation and downstream tasks. Safer-Instruct addresses the challenges in preference data acquisition, advancing the development of safer and more responsible AI systems. Our code and data are available at https://github.com/uscnlp-lime/safer-instruct