Understanding AI in Human Terms
AI is complex and hard to keep up with. Artificial Intelligence is broad, overloaded, and moving quickly in research results and in how it is perceived and deployed in numerous application domains. It is deceptively simple in some ways, and uncertain and difficult to comprehend in others. Even for a computer science professor who does research and publishes regularly on AI, it is challenging to keep up with all the findings and active news. I suspect this goes double for stakeholders not trained in AI who nonetheless are directly impacted in their professions by AI, eluding artists, musicians, legislators, lawyers, doctors, scientists and educators.
In the last 4 years, I have worked with a wide range of leaders and decision makers from fields outside of computer science, each of which has a vested interest in understanding, collaborating with, managing, or resisting AI. One thing nearly all of them have in common, is that they see an incomplete and often narrow portion of what AI is and does. Some folks are overly optimistic on its promise and future, while others actively resist AI on technical or ethical grounds. I often hear discussions of AI terms and concepts in non-technical venues, where the conversation is based on outdated assumptions, misunderstandings or simply outdated information.
Why this, why now?
One of the primary missions of our nonprofit ETCH (Ethical Technology and Computing for Humanity) as an organization is to educate the public on critical issues in AI, and to translate research results from the AI community into information that enables stakeholders to make better decisions in their domains.
Starting this week, we will be writing biweekly articles that focus on key AI topics and issues of interest to the creative, legal, and legislative communities. We will seek to clarify confused topics and provide up to date technical information, on everything from "which tools work and don't work to slow down web scrapers today?" and "can you attribute output from generative AI models to specific pieces of training data?" to "can you really make AI models 'forget' specific training samples?" In each writeup, we will focus on giving plain-text explanations of these key questions, and provide an objective view of what is known and accepted by the AI and security research communities, while giving citations and references so interested readers can follow up and do additional research.
Who are we?
Of course it's important to understand a bit about who we are, and why we are comfortable (or qualified) to write these notes. ETCH is led by two professors, Ben Y. Zhao and Heather Zheng, who in their day jobs, co-direct the SAND research lab at University of Chicago computer science department. The two of us have been academics and researchers for 20+ years. Over these many years, we have done a ton of research on a wide range of topics from networking to systems, to data mining, human-computer interaction, and more than 10 years in security and AI. We have received many awards (listed on our homepages here and here, both temporarily down due to dept maintenance). The most notable ones include Fellows of the ACM (2021 and 2023), MIT Young Innovators (2005 and 2006), USENIX Internet Defense Prize, Concept Art Association Community Impact Award (2024) and TIME 100 most influential in AI (2024 for Ben). Other volunteers at ETCH (PhD students) will also contribute to and edit these notes. ETCH is still in active fundraising mode, and we hope to hire professional researchers, software developers and editors once we have enough funding.
So if you're interested in AI, how it interacts with humans, and how to control its impact on you and your interests, join us and follow along. If you have specific topics you'd love to see us discuss, put it in a comment, and we'll do our best to get to it.