MIT Experts Share Insights on AI Impacts

The power of generative artificial intelligence (AI), the challenges it poses to society, and strategies for ensuring AI helps humanity rather than hurts were central topics at this year’s Alumni Leadership Conference (ALC), held September 29–30 at MIT.

MIT professor Aleksander Mądry SM ’09, PhD ’11 opened the event by outlining how AI systems like ChatGPT work, emphasizing that they do not think the way humans do. “AI systems do not reason,” he said. “What they are doing is mimicking reason.”

The problem for human users, said Mądry, the Cadence Design Systems Professor in the Schwarzman College of Computing and director of the MIT Center for Deployable Machine Learning, is that “mimicking reasoning makes it harder to discern between being helpful and seeming helpful—and that’s an important distinction.”

Professor Asu Ozdaglar SM ’98, PhD ’03, department head of the Electrical Engineering and Computer Science Department and deputy dean of academics for MIT’s Schwarzman College of Computing, also presented a keynote related to AI.

Both speakers highlighted the concerns generative AI presents for humans, including the ease with which it can spread misinformation and bias. “We have to be careful,” Ozdaglar said. “Developing the right ethics is extremely important.”

AI has been dominating the headlines since ChatGPT launched in November 2022, but Mądry pointed out that chatbots date back to the 1960s. (One of the first, ELIZA, was created at MIT in 1966.) For many years, however, programmers had to create precise training models before AI could complete a task. What’s changed today is that―thanks to machine learning and neural networks―modern AIs can teach themselves how the world works.

They do this essentially by trying tasks again and again and learning what works using vast amounts of data. ChatGPT, Mądry noted for example, was trained on 100 billion words.

Yet, for all their sophistication, ChatGPT and similar generative AI tools are at their core just pattern matching tools, Mądry said. They use the patterns they find in massive data sets to answer questions in a way that sounds reasonable. “What the model is able to do that’s remarkable is to piece things together so it seems like an answer to the question,” he says. This is not the same as human thought, he emphasized.

And while generative AI works remarkably well, there are some notable problems with the technology, which Mądry and Ozdaglar detailed. AIs are bad at basic arithmetic, for example, and sometimes “hallucinate,” the term used for citing with confidence information that is inaccurate. AI results can also reflect human biases present in the training data.

“This is not even the worst of it,” Mądry said. Since AI makes it easy to create fake documents and images, it presents hurdles to the way much of society currently functions. “If I can fake anything or you can claim anything can be faked, how do we rethink our legal and economic system?”

Regulation is one way to address the challenges, Mądry said, noting that several policymaking efforts are already under way around the world. (In March, Mądry testified to Congress about the need to question how AI is used.)

Ozdaglar pointed out that some foundational AI tools are currently concentrated in the hands of a few and said MIT is working to ensure that experts from a wide range of fields help shape the future of AI. “Norms should be developed with a broad set of perspectives,” she said. She assured the audience that MIT is working toward that goal.

Education is also important, which is why MIT focuses on training students to consider the implications of the technologies they work on, the speakers said. “The good news is, I think the starting point is to realize these issues and start addressing them as a community,” Ozdaglar said.

“AI—it touches everything,” Mądry concluded. “So, let’s get it right!”

The keynotes by Professors Mądry and Ozdaglar were presented at the 2023 MIT Alumni Leadership Conference, which brings MIT alums and volunteers together for a weekend of connecting and networking over a shared love for and commitment to MIT. Visit the ALC website to learn more. Alum volunteers can also visit the volunteer knowledge base for more MIT- and volunteer-related content.