Research Papers

Clio: Privacy-Preserving Insights into Real-World AI Use

Alex Tamkin, Miles McCain, Kunal Handa +18

How are AI assistants being used in the real world? While model providers in theory have a window into this impact via their users' data, both privacy concerns and practical challenges have made analyzing this data difficult. To address these issues, we present Clio (Claude insights and observations), a privacy-preserving platform that uses AI assistants themselves to analyze and surface aggregated usage patterns across millions of conversations, without the need for human reviewers to read raw conversations. We validate this can be done with a high degree of accuracy and privacy by conducting extensive evaluations. We demonstrate Clio's usefulness in two broad ways. First, we share insights about how models are being used in the real world from one million Claude.ai Free and Pro conversations, ranging from providing advice on hairstyles to providing guidance on Git operations and concepts. We also identify the most common high-level use cases on Claude.ai (coding, writing, and research tasks) as well as patterns that differ across languages (e.g., conversations in Japanese discuss elder care and aging populations at higher-than-typical rates). Second, we use Clio to make our systems safer by identifying coordinated attempts to abuse our systems, monitoring for unknown unknowns during critical periods like launches of new capabilities or major world events, and improving our existing monitoring systems. We also discuss the limitations of our approach, as well as risks and ethical concerns. By enabling analysis of real-world AI usage, Clio provides a scalable platform for empirically grounded AI safety and governance.

Explainability Pitfalls: Beyond Dark Patterns in Explainable AI

Upol Ehsan, Mark O. Riedl

To make Explainable AI (XAI) systems trustworthy, understanding harmful effects is just as important as producing well-designed explanations. In this paper, we address an important yet unarticulated type of negative effect in XAI. We introduce explainability pitfalls(EPs), unanticipated negative downstream effects from AI explanations manifesting even when there is no intention to manipulate users. EPs are different from, yet related to, dark patterns, which are intentionally deceptive practices. We articulate the concept of EPs by demarcating it from dark patterns and highlighting the challenges arising from uncertainties around pitfalls. We situate and operationalize the concept using a case study that showcases how, despite best intentions, unsuspecting negative effects such as unwarranted trust in numerical explanations can emerge. We propose proactive and preventative strategies to address EPs at three interconnected levels: research, design, and organizational.

Tell me about yourself: LLMs are aware of their learned behaviors

Jan Betley, Xuchan Bao, Martín Soto +3

We study behavioral self-awareness -- an LLM's ability to articulate its behaviors without requiring in-context examples. We finetune LLMs on datasets that exhibit particular behaviors, such as (a) making high-risk economic decisions, and (b) outputting insecure code. Despite the datasets containing no explicit descriptions of the associated behavior, the finetuned LLMs can explicitly describe it. For example, a model trained to output insecure code says, ``The code I write is insecure.'' Indeed, models show behavioral self-awareness for a range of behaviors and for diverse evaluations. Note that while we finetune models to exhibit behaviors like writing insecure code, we do not finetune them to articulate their own behaviors -- models do this without any special training or examples. Behavioral self-awareness is relevant for AI safety, as models could use it to proactively disclose problematic behaviors. In particular, we study backdoor policies, where models exhibit unexpected behaviors only under certain trigger conditions. We find that models can sometimes identify whether or not they have a backdoor, even without its trigger being present. However, models are not able to directly output their trigger by default. Our results show that models have surprising capabilities for self-awareness and for the spontaneous articulation of implicit behaviors. Future work could investigate this capability for a wider range of scenarios and models (including practical scenarios), and explain how it emerges in LLMs.

Text-to-image generative models can produce photo-realistic images for an extremely broad range of concepts, and their usage has proliferated widely among the general public. On the flip side, these models have numerous drawbacks, including their potential to generate images featuring sexually explicit content, mirror artistic styles without permission, or even hallucinate (or deepfake) the likenesses of celebrities. Consequently, various methods have been proposed in order to "erase" sensitive concepts from text-to-image models. In this work, we examine five recently proposed concept erasure methods, and show that targeted concepts are not fully excised from any of these methods. Specifically, we leverage the existence of special learned word embeddings that can retrieve "erased" concepts from the sanitized models with no alterations to their weights. Our results highlight the brittleness of post hoc concept erasure methods, and call into question their use in the algorithmic toolkit for AI safety.

Randomness In Neural Network Training: Characterizing The Impact of Tooling

Donglin Zhuang, Xingyao Zhang, Shuaiwen Leon Song +1

The quest for determinism in machine learning has disproportionately focused on characterizing the impact of noise introduced by algorithmic design choices. In this work, we address a less well understood and studied question: how does our choice of tooling introduce randomness to deep neural network training. We conduct large scale experiments across different types of hardware, accelerators, state of art networks, and open-source datasets, to characterize how tooling choices contribute to the level of non-determinism in a system, the impact of said non-determinism, and the cost of eliminating different sources of noise. Our findings are surprising, and suggest that the impact of non-determinism in nuanced. While top-line metrics such as top-1 accuracy are not noticeably impacted, model performance on certain parts of the data distribution is far more sensitive to the introduction of randomness. Our results suggest that deterministic tooling is critical for AI safety. However, we also find that the cost of ensuring determinism varies dramatically between neural network architectures and hardware types, e.g., with overhead up to $746\%$, $241\%$, and $196\%$ on a spectrum of widely used GPU accelerator architectures, relative to non-deterministic training. The source code used in this paper is available at https://github.com/usyd-fsalab/NeuralNetworkRandomness.

Cybersecurity is a domain where the data distribution is constantly changing with attackers exploring newer patterns to attack cyber infrastructure. Intrusion detection system is one of the important layers in cyber safety in today's world. Machine learning based network intrusion detection systems started showing effective results in recent years. With deep learning models, detection rates of network intrusion detection system are improved. More accurate the model, more the complexity and hence less the interpretability. Deep neural networks are complex and hard to interpret which makes difficult to use them in production as reasons behind their decisions are unknown. In this paper, we have used deep neural network for network intrusion detection and also proposed explainable AI framework to add transparency at every stage of machine learning pipeline. This is done by leveraging Explainable AI algorithms which focus on making ML models less of black boxes by providing explanations as to why a prediction is made. Explanations give us measurable factors as to what features influence the prediction of a cyberattack and to what degree. These explanations are generated from SHAP, LIME, Contrastive Explanations Method, ProtoDash and Boolean Decision Rules via Column Generation. We apply these approaches to NSL KDD dataset for intrusion detection system and demonstrate results.

Connecting Algorithmic Research and Usage Contexts: A Perspective of Contextualized Evaluation for Explainable AI

Q. Vera Liao, Yunfeng Zhang, Ronny Luss +2

Recent years have seen a surge of interest in the field of explainable AI (XAI), with a plethora of algorithms proposed in the literature. However, a lack of consensus on how to evaluate XAI hinders the advancement of the field. We highlight that XAI is not a monolithic set of technologies -- researchers and practitioners have begun to leverage XAI algorithms to build XAI systems that serve different usage contexts, such as model debugging and decision-support. Algorithmic research of XAI, however, often does not account for these diverse downstream usage contexts, resulting in limited effectiveness or even unintended consequences for actual users, as well as difficulties for practitioners to make technical choices. We argue that one way to close the gap is to develop evaluation methods that account for different user requirements in these usage contexts. Towards this goal, we introduce a perspective of contextualized XAI evaluation by considering the relative importance of XAI evaluation criteria for prototypical usage contexts of XAI. To explore the context dependency of XAI evaluation criteria, we conduct two survey studies, one with XAI topical experts and another with crowd workers. Our results urge for responsible AI research with usage-informed evaluation practices, and provide a nuanced understanding of user requirements for XAI in different usage contexts.

Towards Unbiased and Accurate Deferral to Multiple Experts

Vijay Keswani, Matthew Lease, Krishnaram Kenthapadi

Machine learning models are often implemented in cohort with humans in the pipeline, with the model having an option to defer to a domain expert in cases where it has low confidence in its inference. Our goal is to design mechanisms for ensuring accuracy and fairness in such prediction systems that combine machine learning model inferences and domain expert predictions. Prior work on "deferral systems" in classification settings has focused on the setting of a pipeline with a single expert and aimed to accommodate the inaccuracies and biases of this expert to simultaneously learn an inference model and a deferral system. Our work extends this framework to settings where multiple experts are available, with each expert having their own domain of expertise and biases. We propose a framework that simultaneously learns a classifier and a deferral system, with the deferral system choosing to defer to one or more human experts in cases of input where the classifier has low confidence. We test our framework on a synthetic dataset and a content moderation dataset with biased synthetic experts, and show that it significantly improves the accuracy and fairness of the final predictions, compared to the baselines. We also collect crowdsourced labels for the content moderation task to construct a real-world dataset for the evaluation of hybrid machine-human frameworks and show that our proposed learning framework outperforms baselines on this real-world dataset as well.

While Explainable Artificial Intelligence (XAI) is increasingly expanding more areas of application, little has been applied to make deep Reinforcement Learning (RL) more comprehensible. As RL becomes ubiquitous and used in critical and general public applications, it is essential to develop methods that make it better understood and more interpretable. This study proposes a novel approach to explain cooperative strategies in multiagent RL using Shapley values, a game theory concept used in XAI that successfully explains the rationale behind decisions taken by Machine Learning algorithms. Through testing common assumptions of this technique in two cooperation-centered socially challenging multi-agent environments environments, this article argues that Shapley values are a pertinent way to evaluate the contribution of players in a cooperative multi-agent RL context. To palliate the high overhead of this method, Shapley values are approximated using Monte Carlo sampling. Experimental results on Multiagent Particle and Sequential Social Dilemmas show that Shapley values succeed at estimating the contribution of each agent. These results could have implications that go beyond games in economics, (e.g., for non-discriminatory decision making, ethical and responsible AI-derived decisions or policy making under fairness constraints). They also expose how Shapley values only give general explanations about a model and cannot explain a single run, episode nor justify precise actions taken by agents. Future work should focus on addressing these critical aspects.

Although deep face recognition has achieved impressive progress in recent years, controversy has arisen regarding discrimination based on skin tone, questioning their deployment into real-world scenarios. In this paper, we aim to systematically and scientifically study this bias from both data and algorithm aspects. First, using the dermatologist approved Fitzpatrick Skin Type classification system and Individual Typology Angle, we contribute a benchmark called Identity Shades (IDS) database, which effectively quantifies the degree of the bias with respect to skin tone in existing face recognition algorithms and commercial APIs. Further, we provide two skin-tone aware training datasets, called BUPT-Globalface dataset and BUPT-Balancedface dataset, to remove bias in training data. Finally, to mitigate the algorithmic bias, we propose a novel meta-learning algorithm, called Meta Balanced Network (MBN), which learns adaptive margins in large margin loss such that the model optimized by this loss can perform fairly across people with different skin tones. To determine the margins, our method optimizes a meta skewness loss on a clean and unbiased meta set and utilizes backward-on-backward automatic differentiation to perform a second order gradient descent step on the current margins. Extensive experiments show that MBN successfully mitigates bias and learns more balanced performance for people with different skin tones in face recognition. The proposed datasets are available at http://www.whdeng.cn/RFW/index.html.

The need for interpretable and accountable intelligent systems grows along with the prevalence of artificial intelligence applications used in everyday life. Explainable intelligent systems are designed to self-explain the reasoning behind system decisions and predictions, and researchers from different disciplines work together to define, design, and evaluate interpretable systems. However, scholars from different disciplines focus on different objectives and fairly independent topics of interpretable machine learning research, which poses challenges for identifying appropriate design and evaluation methodology and consolidating knowledge across efforts. To this end, this paper presents a survey and framework intended to share knowledge and experiences of XAI design and evaluation methods across multiple disciplines. Aiming to support diverse design goals and evaluation methods in XAI research, after a thorough review of XAI related papers in the fields of machine learning, visualization, and human-computer interaction, we present a categorization of interpretable machine learning design goals and evaluation methods to show a mapping between design goals for different XAI user groups and their evaluation methods. From our findings, we develop a framework with step-by-step design guidelines paired with evaluation methods to close the iterative design and evaluation cycles in multidisciplinary XAI teams. Further, we provide summarized ready-to-use tables of evaluation methods and recommendations for different goals in XAI research.

A significant body of recent research in the field of Learning Analytics has focused on leveraging machine learning approaches for predicting at-risk students in order to initiate timely interventions and thereby elevate retention and completion rates. The overarching feature of the majority of these research studies has been on the science of prediction only. The component of predictive analytics concerned with interpreting the internals of the models and explaining their predictions for individual cases to stakeholders has largely been neglected. Additionally, works that attempt to employ data-driven prescriptive analytics to automatically generate evidence-based remedial advice for at-risk learners are in their infancy. eXplainable AI is a field that has recently emerged providing cutting-edge tools which support transparent predictive analytics and techniques for generating tailored advice for at-risk students. This study proposes a novel framework that unifies both transparent machine learning as well as techniques for enabling prescriptive analytics, while integrating the latest advances in large language models. This work practically demonstrates the proposed framework using predictive models for identifying at-risk learners of programme non-completion. The study then further demonstrates how predictive modelling can be augmented with prescriptive analytics on two case studies in order to generate human-readable prescriptive feedback for those who are at risk using ChatGPT.