Across all fields of academic study, experts cite their sources when sharing information. While large language models (LLMs) excel at synthesizing information, they do not provide reliable citation to sources, making it difficult to trace and verify the origins of the information they present. In contrast, search engines make sources readily accessible to users and place the burden of synthesizing information on the user. Through a survey, we find that users prefer search engines over LLMs for high-stakes queries, where concerns regarding information provenance outweigh the perceived utility of LLM responses. To examine the interplay between verifiability and utility of information-sharing tools, we introduce the extractive-abstractive spectrum, in which search engines and LLMs are extreme endpoints encapsulating multiple unexplored intermediate operating points. Search engines are extractive because they respond to queries with snippets of sources with links (citations) to the original webpages. LLMs are abstractive because they address queries with answers that synthesize and logically transform relevant information from training and in-context sources without reliable citation. We define five operating points that span the extractive-abstractive spectrum and conduct human evaluations on seven systems across four diverse query distributions that reflect real-world QA settings: web search, language simplification, multi-step reasoning, and medical advice. As outputs become more abstractive, we find that perceived utility improves by as much as 200%, while the proportion of properly cited sentences decreases by as much as 50% and users take up to 3 times as long to verify cited information. Our findings recommend distinct operating points for domain-specific LLM systems and our failure analysis informs approaches to high-utility LLM systems that empower users to verify information.

As businesses, products, and services spring up around large language models, the trustworthiness of these models hinges on the verifiability of their outputs. However, methods for explaining language model outputs largely fall across two distinct fields of study which both use the term "attribution" to refer to entirely separate techniques: citation generation and training data attribution. In many modern applications, such as legal document generation and medical question answering, both types of attributions are important. In this work, we argue for and present a unified framework of large language model attributions. We show how existing methods of different types of attribution fall under the unified framework. We also use the framework to discuss real-world use cases where one or both types of attributions are required. We believe that this unified framework will guide the use case driven development of systems that leverage both types of attribution, as well as the standardization of their evaluation.

Datasets play a critical role in shaping the perception of performance and progress in machine learning (ML)--the way we collect, process, and analyze data affects the way we benchmark success and form new research agendas (Paullada et al., 2020; Dotan & Milli, 2020). A growing appreciation of this determinative role of datasets has sparked a concomitant concern that standard datasets used for training and evaluating ML models lack diversity along significant dimensions, for example, geography, gender, and skin type (Shankar et al., 2017; Buolamwini & Gebru, 2018). Lack of diversity in evaluation data can obfuscate disparate performance when evaluating based on aggregate accuracy (Buolamwini & Gebru, 2018). Lack of diversity in training data can limit the extent to which learned models can adequately apply to all portions of a population, a concern highlighted in recent work in the medical domain (Habib et al., 2019; Hofmanninger et al., 2020). Our work aims to develop a general unifying perspective on the way that dataset composition affects outcomes of machine learning systems.