Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models

The tendency of users to anthropomorphise large language models (LLMs) is of growing interest to AI developers, researchers, and policy-makers. Here, we present a novel method for empirically evaluating anthropomorphic LLM behaviours in realistic and varied settings. Going beyond single-turn static benchmarks, we contribute three methodological advances in state-of-the-art (SOTA) LLM evaluation. First, we develop a multi-turn evaluation of 14 anthropomorphic behaviours. Second, we present a scalable, automated approach by employing simulations of user interactions. Third, we conduct an interactive, large-scale human subject study (N=1101) to validate that the model behaviours we measure predict real users' anthropomorphic perceptions. We find that all SOTA LLMs evaluated exhibit similar behaviours, characterised by relationship-building (e.g., empathy and validation) and first-person pronoun use, and that the majority of behaviours only first occur after multiple turns. Our work lays an empirical foundation for investigating how design choices influence anthropomorphic model behaviours and for progressing the ethical debate on the desirability of these behaviours. It also showcases the necessity of multi-turn evaluations for complex social phenomena in human-AI interaction.

Insights on Disagreement Patterns in Multimodal Safety Perception across Diverse Rater Groups

AI systems crucially rely on human ratings, but these ratings are often aggregated, obscuring the inherent diversity of perspectives in real-world phenomenon. This is particularly concerning when evaluating the safety of generative AI, where perceptions and associated harms can vary significantly across socio-cultural contexts. While recent research has studied the impact of demographic differences on annotating text, there is limited understanding of how these subjective variations affect multimodal safety in generative AI. To address this, we conduct a large-scale study employing highly-parallel safety ratings of about 1000 text-to-image (T2I) generations from a demographically diverse rater pool of 630 raters balanced across 30 intersectional groups across age, gender, and ethnicity. Our study shows that (1) there are significant differences across demographic groups (including intersectional groups) on how severe they assess the harm to be, and that these differences vary across different types of safety violations, (2) the diverse rater pool captures annotation patterns that are substantially different from expert raters trained on specific set of safety policies, and (3) the differences we observe in T2I safety are distinct from previously documented group level differences in text-based safety tasks. To further understand these varying perspectives, we conduct a qualitative analysis of the open-ended explanations provided by raters. This analysis reveals core differences into the reasons why different groups perceive harms in T2I generations. Our findings underscore the critical need for incorporating diverse perspectives into safety evaluation of generative AI ensuring these systems are truly inclusive and reflect the values of all users.

Imagen 3

We introduce Imagen 3, a latent diffusion model that generates high quality images from text prompts. We describe our quality and responsibility evaluations. Imagen 3 is preferred over other state-of-the-art (SOTA) models at the time of evaluation. In addition, we discuss issues around safety and representation, as well as methods we used to minimize the potential harm of our models.

Adversarial Nibbler: A Data-Centric Challenge for Improving the Safety of Text-to-Image Models

The generative AI revolution in recent years has been spurred by an expansion in compute power and data quantity, which together enable extensive pre-training of powerful text-to-image (T2I) models. With their greater capabilities to generate realistic and creative content, these T2I models like DALL-E, MidJourney, Imagen or Stable Diffusion are reaching ever wider audiences. Any unsafe behaviors inherited from pretraining on uncurated internet-scraped datasets thus have the potential to cause wide-reaching harm, for example, through generated images which are violent, sexually explicit, or contain biased and derogatory stereotypes. Despite this risk of harm, we lack systematic and structured evaluation datasets to scrutinize model behavior, especially adversarial attacks that bypass existing safety filters. A typical bottleneck in safety evaluation is achieving a wide coverage of different types of challenging examples in the evaluation set, i.e., identifying 'unknown unknowns' or long-tail problems. To address this need, we introduce the Adversarial Nibbler challenge. The goal of this challenge is to crowdsource a diverse set of failure modes and reward challenge participants for successfully finding safety vulnerabilities in current state-of-the-art T2I models. Ultimately, we aim to provide greater awareness of these issues and assist developers in improving the future safety and reliability of generative AI models. Adversarial Nibbler is a data-centric challenge, part of the DataPerf challenge suite, organized and supported by Kaggle and MLCommons.

Supporting Human-AI Collaboration in Auditing LLMs with LLMs

Large language models are becoming increasingly pervasive and ubiquitous in society via deployment in sociotechnical systems. Yet these language models, be it for classification or generation, have been shown to be biased and behave irresponsibly, causing harm to people at scale. It is crucial to audit these language models rigorously. Existing auditing tools leverage either or both humans and AI to find failures. In this work, we draw upon literature in human-AI collaboration and sensemaking, and conduct interviews with research experts in safe and fair AI, to build upon the auditing tool: AdaTest (Ribeiro and Lundberg, 2022), which is powered by a generative large language model (LLM). Through the design process we highlight the importance of sensemaking and human-AI communication to leverage complementary strengths of humans and generative models in collaborative auditing. To evaluate the effectiveness of the augmented tool, AdaTest++, we conduct user studies with participants auditing two commercial language models: OpenAI's GPT-3 and Azure's sentiment analysis model. Qualitative analysis shows that AdaTest++ effectively leverages human strengths such as schematization, hypothesis formation and testing. Further, with our tool, participants identified a variety of failures modes, covering 26 different topics over 2 tasks, that have been shown before in formal audits and also those previously under-reported.

How do Authors' Perceptions of their Papers Compare with Co-authors' Perceptions and Peer-review Decisions?

How do author perceptions match up to the outcomes of the peer-review process and perceptions of others? In a top-tier computer science conference (NeurIPS 2021) with more than 23,000 submitting authors and 9,000 submitted papers, we survey the authors on three questions: (i) their predicted probability of acceptance for each of their papers, (ii) their perceived ranking of their own papers based on scientific contribution, and (iii) the change in their perception about their own papers after seeing the reviews. The salient results are: (1) Authors have roughly a three-fold overestimate of the acceptance probability of their papers: The median prediction is 70% for an approximately 25% acceptance rate. (2) Female authors exhibit a marginally higher (statistically significant) miscalibration than male authors; predictions of authors invited to serve as meta-reviewers or reviewers are similarly calibrated, but better than authors who were not invited to review. (3) Authors' relative ranking of scientific contribution of two submissions they made generally agree (93%) with their predicted acceptance probabilities, but there is a notable 7% responses where authors think their better paper will face a worse outcome. (4) The author-provided rankings disagreed with the peer-review decisions about a third of the time; when co-authors ranked their jointly authored papers, co-authors disagreed at a similar rate -- about a third of the time. (5) At least 30% of respondents of both accepted and rejected papers said that their perception of their own paper improved after the review process. The stakeholders in peer review should take these findings into account in setting their expectations from peer review.

A Unifying Framework for Combining Complementary Strengths of Humans and ML toward Better Predictive Decision-Making

Hybrid human-ML systems are increasingly in charge of consequential decisions in a wide range of domains. A growing body of work has advanced our understanding of these systems by providing empirical and theoretical analyses. However, existing empirical results are mixed, and theoretical proposals are often incompatible with each other. Our goal in this work is to bring much-needed organization to this field by offering a unifying framework for understanding conditions under which combining complementary strengths of human and ML leads to higher quality decisions than those produced by them individually -- a state to which we refer to as human-ML complementarity. We focus specifically on the context of human-ML predictive decision-making systems and investigate optimal ways of combining human and ML-based predictive decisions, accounting for the underlying causes of variation in their judgments. Within this scope, we present two crucial contributions. First, drawing upon prior literature in human psychology, machine learning, and human-computer interaction, we introduce a taxonomy characterizing a wide variety of criteria across which human and machine decision-making differ. Building on our taxonomy, our second contribution presents a unifying optimization-based framework for formalizing how human and ML predictive decisions should be aggregated optimally. We show that our proposed framework encompasses several existing models of human-ML complementarity as special cases. Last but not least, the exploratory analysis of our framework offers a critical piece of insight for future work in this area: the mechanism by which we combine human-ML judgments should be informed by the underlying causes of their diverging decisions.

A Large Scale Randomized Controlled Trial on Herding in Peer-Review Discussions

Peer review is the backbone of academia and humans constitute a cornerstone of this process, being responsible for reviewing papers and making the final acceptance/rejection decisions. Given that human decision making is known to be susceptible to various cognitive biases, it is important to understand which (if any) biases are present in the peer-review process and design the pipeline such that the impact of these biases is minimized. In this work, we focus on the dynamics of between-reviewers discussions and investigate the presence of herding behaviour therein. In that, we aim to understand whether reviewers and more senior decision makers get disproportionately influenced by the first argument presented in the discussion when (in case of reviewers) they form an independent opinion about the paper before discussing it with others. Specifically, in conjunction with the review process of ICML 2020 -- a large, top tier machine learning conference -- we design and execute a randomized controlled trial with the goal of testing for the conditional causal effect of the discussion initiator's opinion on the outcome of a paper.

Deciding Fast and Slow: The Role of Cognitive Biases in AI-assisted Decision-making

Several strands of research have aimed to bridge the gap between artificial intelligence (AI) and human decision-makers in AI-assisted decision-making, where humans are the consumers of AI model predictions and the ultimate decision-makers in high-stakes applications. However, people's perception and understanding is often distorted by their cognitive biases, like confirmation bias, anchoring bias, availability bias, to name a few. In this work, we use knowledge from the field of cognitive science to account for cognitive biases in the human-AI collaborative decision-making system and mitigate their negative effects. To this end, we mathematically model cognitive biases and provide a general framework through which researchers and practitioners can understand the interplay between cognitive biases and human-AI accuracy. We then focus on anchoring bias, a bias commonly witnessed in human-AI partnerships. We devise a cognitive science-driven, time-based approach to de-anchoring. A user experiment shows the effectiveness of this approach in human-AI collaborative decision-making. Using the results from this first experiment, we design a time allocation strategy for a resource constrained setting so as to achieve optimal human-AI collaboration under some assumptions. A second user study shows that our time allocation strategy can effectively debias the human when the AI model has low confidence and is incorrect.

Two-Sample Testing on Ranked Preference Data and the Role of Modeling Assumptions

A number of applications require two-sample testing on ranked preference data. For instance, in crowdsourcing, there is a long-standing question of whether pairwise comparison data provided by people is distributed similar to ratings-converted-to-comparisons. Other examples include sports data analysis and peer grading. In this paper, we design two-sample tests for pairwise comparison data and ranking data. For our two-sample test for pairwise comparison data, we establish an upper bound on the sample complexity required to correctly distinguish between the distributions of the two sets of samples. Our test requires essentially no assumptions on the distributions. We then prove complementary lower bounds showing that our results are tight (in the minimax sense) up to constant factors. We investigate the role of modeling assumptions by proving lower bounds for a range of pairwise comparison models (WST, MST,SST, parameter-based such as BTL and Thurstone). We also provide testing algorithms and associated sample complexity bounds for the problem of two-sample testing with partial (or total) ranking data.Furthermore, we empirically evaluate our results via extensive simulations as well as two real-world datasets consisting of pairwise comparisons. By applying our two-sample test on real-world pairwise comparison data, we conclude that ratings and rankings provided by people are indeed distributed differently. On the other hand, our test recognizes no significant difference in the relative performance of European football teams across two seasons. Finally, we apply our two-sample test on a real-world partial and total ranking dataset and find a statistically significant difference in Sushi preferences across demographic divisions based on gender, age and region of residence.