Generative AI meets Responsible AI (Tutorial)

Overview

Artificial Intelligence (AI) based solutions are increasingly deployed in high-stakes domains such as healthcare, lending, hiring, criminal justice, and education, thereby transforming the way these industries operate and impacting individuals, businesses, and society as a whole. Consequently, it is critical to ensure that the underlying AI models are making accurate predictions, are robust to shifts in the data, are not relying on spurious features, are not unduly discriminating against minority groups, and more broadly, are developed and deployed following responsible AI guidelines.

The rapid development and deployment of generative AI models and applications has increased the urgency of addressing above issues. Generative AI refers to deep learning models and systems that create new digital text, images, audio, video, code, and other artifacts based on existing content, such as text, images, audio, or video. Text generator models such as GPT-4 and ChatGPT and text to image models such as DALL-E 2 and Stable Diffusion are well-known examples of Generative AI models. Generative AI has significant implications for a wide spectrum of industries and applications requiring knowledge work and creative work such as writing assistants, graphic design, art generation, advertising, copywriting for marketing and sales, architecture, gaming, coding, and drug discovery. However, there are several ethical and social considerations associated with generative AI models and applications. These concerns include lack of interpretability, bias and discrimination, privacy, lack of model robustness, fake and misleading content, copyright implications, plagiarism, and environmental impact associated with training and inference of generative AI models. In light of these risks, it is crucial to develop and deploy generative AI models and applications following responsible AI principles.

In this tutorial, we first motivate the need for adopting responsible AI principles when developing and deploying large language models and other generative AI models, as part of a broader AI model governance and responsible AI framework, from societal, legal, customer/end-user, and model developer perspectives, and provide a roadmap for thinking about responsible AI for generative AI in practice. We provide a brief technical overview of text and image generation models, and highlight the key responsible AI desiderata associated with these models. We then describe the technical considerations and challenges associated with realizing the above desiderata in practice. We focus on real-world generative AI use cases spanning domains such as media generation, writing assistants, copywriting, code generation, and conversational assistants, present practical solution approaches / guidelines for applying responsible AI techniques effectively, discuss lessons learned from deploying responsible AI approaches for generative AI applications in practice, and highlight the key open research problems. We hope that our tutorial will inform both researchers and practitioners, stimulate further research on responsible AI in the context of generative AI, and pave the way for building more reliable and trustworthy generative AI applications in the future.

Contributors

Krishnaram Kenthapadi (Fiddler AI, USA)

Hima Lakkaraju (Harvard University, USA)

Nazneen Rajani (Hugging Face, USA)

Generative AI meets Responsible AI: KDD'23 Tutorial
Trustworthy Generative AI: ICML'23 Tutorial
Generative AI meets Responsible AI: FAccT'23 Tutorial

Tutorial Editions

ICML'23, KDD'23, and FAccT'23 Tutorial Slides

Tutorial Video Recording: ICML'23 Video, FAccT'23 Video, KDD'23 Video (TBA)

Contributor Bios

Krishnaram Kenthapadi is the Chief AI Officer & Chief Scientist of Fiddler AI, an enterprise startup building a responsible AI and ML monitoring platform. Previously, he was a Principal Scientist at Amazon AWS AI, where he led the fairness, explainability, privacy, and model understanding initiatives in Amazon AI platform. Prior to joining Amazon, he led similar efforts at the LinkedIn AI team, and served as LinkedIn’s representative in Microsoft’s AI and Ethics in Engineering and Research (AETHER) Advisory Board. Previously, he was a Researcher at Microsoft Research Silicon Valley Lab. Krishnaram received his Ph.D. in Computer Science from Stanford University in 2006. His work has been recognized through awards at NAACL, WWW, SODA, CIKM, ICML AutoML workshop, and Microsoft’s AI/ML conference (MLADS). He has published 50+ papers, with 7000+ citations and filed 150+ patents (70 granted). He has presented tutorials on privacy, fairness, explainable AI, ML model monitoring, responsible AI, and trustworthy generative AI in industry at forums such as KDD '18 '19 '22 '23,  WSDM '19, WWW '19 '20 '21 '23, FAccT '20 '21 '22 '23, AAAI '20 '21, and ICML '21 '23, and instructed a course on responsible AI at Stanford.

Hima Lakkaraju is an assistant professor at Harvard University focusing on explainability, fairness, and robustness of machine learning models. She has also been working with various domain experts in policy and healthcare to understand the real-world implications of explainable and fair ML. Hima has been named as one of the world’s top innovators under 35 by both MIT Tech Review and Vanity Fair. Her research has also received best paper awards at SIAM International Conference on Data Mining (SDM) and INFORMS, and grants from NSF, Google, Amazon, and Bayer. Hima has given keynote talks at various top ML conferences and workshops including CIKM, ICML, NeurIPS, AAAI, and CVPR, and her research has also been showcased by popular media outlets including the New York Times, MIT Tech Review, TIME magazine, and Forbes. More recently, she co-founded the Trustworthy ML Initiative to enable easy access to resources on trustworthy ML and to build a community of researchers/practitioners working on the topic.

Nazneen Rajani is a Research Lead at Hugging Face, a startup with a mission to democratize ML, leading the robust ML research direction. Before Hugging Face, she worked at Salesforce Research with Richard Socher and led a team of researchers focused on building robust natural language generation systems based on Large Language Models (LLMs). She completed her Ph.D. in Computer Science from the University of Texas at Austin and is an expert on LLMs and Interpretable ML. Nazneen has over 50 papers published at ACL, EMNLP, NAACL, NeurIPS, and ICLR and has her research covered by Quanta magazine, VentureBeat, SiliconAngle, ZDNet, and Datanami. She has presented an industry keynote at EMNLP 2022, and has given invited talks at several research and industry forums.

Tutorial Outline and Description

The tutorial will consist of two parts: (1) technical deep-dive into generative AI landscape including advances, challenges, and opportunities; (2) ethical considerations including privacy, consent, and responsible release, along with approaches for mitigating harms and long term planning.

Introduction and overview of the generative AI landscape

Give an overview of the generative AI landscape in ML and motivate the topic with some questions. What constitutes generative AI? Why is generative AI an important topic? What are the origins of the research field?

Technical deepdive into text generation models and image generation models

(1) Types of generative AI models, their similarities and differences.

(2) Applications of generative AI - images, music, text, code, video.

(3) Model training: (a) Pretraining method and datasets; (b) Diffusion approach; (c) Supervised fine-tuning; (d) Instruction datasets -- Self-instruct, Supernatural Instructions; (e) Reinforcement Learning with Human Feedback (RLHF); (f) Compute costs and infrastructure

(4) Model evaluation and auditing including (a) metrics, datasets, and benchmarks; (b) Automated vs. human evaluations; (c) Red-teaming and evaluations on toxicity/harmfulness.

(5) Model Access.

Technical and ethical challenges with generative AI

There are a number of technical and ethical challenges associated with generative AI, including lack of interpretability, bias and discrimination, privacy, lack of model robustness, fake and misleading content, copyright implications, psychological harm experienced by content moderation workers, plagiarism, and environmental impact associated with training and inference of generative AI models. We will highlight the following challenges:

(1) Trust and lack of interpretability: are significant concerns for LLMs and other generative AI models especially due to their large size and opaque behavior. Often, such models exhibit emergent behavior, and demonstrate capabilities not intended as part of the architectural design and not anticipated by the model developers. A lack of transparency, lineage, and trustworthiness prevents users from validating and citing the responses generated by search and information retrieval mechanisms powered by LLMs. Further, LLMs and other generative AI models could be used to generate fake and misleading content (including deepfakes) and spread misinformation with serious social and political consequences.

(2) Bias and discrimination: Generative AI models are often trained on large corpuses of data, making it difficult to audit the training data for different types of biases. For example, many LLMs have been shown to exhibit different types of biases such as gender stereotypes, undesirable biases towards mentions of disability, and religious stereotypes. Similarly, contrastive language-vision AI models (such as Stable Diffusion) trained on automatically collected web scraped data have been shown to learn biases of sexual objectification, which can propagate to downstream applications. Further, generative AI models are typically trained on data crawled from the internet, and consequently the models often reflect the practices of the wealthiest communities and countries.

(3) Privacy and copyright implications: Large language models have been shown to memorize personally identifiable information occurring just once in the training data and reproduce such data, raising potential privacy concerns. Further, image diffusion models such as DALL-E 2, Imagen, and Stable Diffusion have been shown to memorize individual images from their training data and emit them at generation time, with potential privacy as well as copyright implications.

(4) Model robustness and security: Large language models often lack the ability to provide uncertainty estimates. Without knowledge of the extent of confidence (or uncertainty) of the model, it becomes difficult for users to decide when the model's output can be trusted. Model security is a key concern for generative AI models, especially since several applications may be derived from the same underlying foundation model. Large language models have been shown to be vulnerable to data poisoning attacks.

Solutions for alleviating the challenges with real-world use cases and case studies (including practical challenges and lessons learned in industry)

We will discuss solution approaches such as watermarking, release norms, red-teaming, and confidence building measures (CBMs), and present case studies across different companies, spanning application domains such as financial services, healthcare, hiring, conversational assistants, online retail, computational advertising, search and recommendation systems, and fraud detection. We hope that our tutorial will inform both researchers and practitioners, stimulate further research on responsible AI approaches for generative AI, and pave the way for building more reliable generative AI models and applications in the future.

This tutorial is aimed at attendees with a wide range of interests and backgrounds both in academia and industry, including researchers interested in knowing about responsible AI techniques and tools in the context of generative AI models as well as practitioners interested in implementing such tools for various generative AI applications. We will not assume any prerequisite knowledge, and present the advances, challenges, and opportunities of Generative AI by building intuition to ensure that the material is accessible to all attendees.

Related Tutorials and Resources