Assistant Professor, The Information School

Paul G. Allen School of Computer Science & Engineering (courtesy) Co-director • Tech Policy Lab

Faculty Affiliate • UW NLP RAISE, VSD Lab

University of Washington
Read her bio

What we learned:

• Current LLM models of AI transfer biased information from human society to AI by reflecting the statistical regularities of language. Bias is also driven by underrepresentation of groups and broad categorization that erase minority group nuances. 

• Generative AI models such as text-to-image generators amplify bias.

• Addressing bias in generative AI will require technical approaches, policies and awareness among researchers, policymakers and the public. We need to be able to audit the “black box” that is generative AI.

Michael H. Smith and Lotfi A. Zadeh Chair in Engineering
Professor in the Division of Computer Science, EECS
UC Berkeley

Read his bio and his book: "Human Compatible"

What we learned:

• The purpose of artificial intelligence is to fulfill human objectives, and we must avoid the standard model of AI, in which machines are designed to carry out their goals regardless of the human cost. These ideas form the central theme of Dr. Russell’s book, Human Compatible: Artificial Intelligence and the Problem of Control.

• An ideal approach to general AI is knowledge-based, in which the machine is capable of deliberation and reasoning on the basis of learned knowledge and human benefits.

• Current neural network-based AI systems, such as large-language models, are powerful but unlikely to be the final solution to general AI, due to their intrinsic and unavoidable flaws, including their tendency to adopt internal goals that we do not understand or control. Some hybrid of well-founded AI and deep learning is likely to be the best approach.

Corporate Vice President of Research and Incubations, Microsoft

Read his bio and his book "The AI Revolution in Medicine: GPT-4 and Beyond"

What we learned:

• GPT-4 shows great potential for reducing clinician time spent on administrative and medical record documentation tasks in health care.

• Using voice recognition platform provides a more natural, iterative engagement with GPT-4, and eliminates the need to type questions into a keyboard.

• GPT-4 has an inherent need to provide the user with a convincing answer, even when it's manufactured and false.  However, you can also use GPT-4 to confirm the accuracy of responses, and to identify  bias in responses.

• Enabling GPT-4 to use tools, for data retrieval, data analysis, and computer programming, opens broader possibilities for applications in medicine.

1. "Artificial Intelligence - from starting pilots to scalable privilege"

2. "My North Star for the Future of AI"

3. "Foundation models for generalist medical artificial intelligence"

4. "AI in Medicine: Benefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine"

5. "Effective gene expression prediction from sequence by integrating long-range interactions"

6. "AI recognition of patient race in medical imaging: a modelling study"

7. "ChatGPT and generative artificial intelligence for medical education: potential impact and opportunity"

8. "Assessment of Machine Learning to Estimate the Individual Treatment Effect of Corticosteroids in Septic Shock" 

Develop a strategy for UCSF to become a national leader in AI for health equity—as through research, education, patient care, and advocacy. Include equity and ethical use of AI as a primary goal/mission, with accountability and redress

Update the “physician of the future” – to include new competencies related to AI.

Strengthen pipeline from discovery to clinical evaluation. Including EHR sand box/test bed; transparency of processes; implementation science framework; and support for AI integration.

Develop educational plan for discovery science faculty and trainees in the uses and opportunities for the application of AI to their scientific questions.