AI for Scientific Discovery: Proceedings of a Workshop (2024)
Chapter: 5 Closing Discussion
5
Closing Discussion
For the closing discussion, workshop chair Bradley Malin (Vanderbilt University) asked the five other members of the workshop planning committee to discuss what they had heard and respond to questions. Missy Cummings (George Mason University) built on a comment from the previous session about safety by saying that it will be important to provide more funding for artificial intelligence (AI) safety and maintenance. Yolanda Gil (University of Southern California) said that it is important to support the development of more AI generalists who can generalize from findings in specific AI domains and thus spread advances.
Cummings reiterated what she said was a particularly important point that people should carry home with them—that neural nets and other types of AI have no way of knowing whether their answers are right or wrong. They do not “know” anything despite how knowledgeable they may appear to people who interact with them.
Malin asked Cummings whether her comments about safety in AI are universal or whether they apply to certain types and areas of AI. It is possible, she answered, that at some point physical safety will not be a major concern with AI, but what she is concerned about is that faith in AI may lead people to think less critically and potentially miss solutions that might otherwise be found. Aarti Singh (Carnegie Mellon University) said that one way to approach that concern is through educating people about AI. Cummings agreed and said that today there is a serious problem with technological illiteracy concerning AI. One way to address that, she sug-
gested, would be to make computer science a core part of education, like English and math.
Responding to a question about how to address some of the challenges that the rapidly increasing power of AI is creating, Malin said that no one yet knows what sorts of ethical issues will arise, so it is important to develop a general approach for identifying and addressing those issues. “We will never be able to have a perfect ethical framework,” he said. More generally, he added, it would be useful to develop a checklist of things that people should be careful to address as they develop AI models to respond to potential issues that have already been identified. The checklist would evolve with experience.
In response to a question about the next steps, Malin identified computational infrastructure as an area of importance to AI that has so far received relatively little attention. What innovations in infrastructure will be necessary to support the growing interest and use of AI? Workforce and training are other areas that will need a great deal of discussion and thought, he added.
In response to a comment by a member of the audience, Malin stated that it is important to keep in mind that AI is not one thing but is composed of many different types of AI. Patrick Riley (Relay Therapeutics) agreed but said that how one splits AI should depend on circumstances—splitting AI to hold meetings with different types of AI researchers would be different, for instance, than splitting AI for regulation.
