The AI Prescription

Why did the AI doctor become a stand-up comedian? Because it had a great sense of humour algorithm, but its bedside manner needed some debugging!

Is this joke terrible? Yes. But does it make sense and sound like it came from a human? Kind of, yeah! ChatGPT generated this gag from the simple request ‘give me a joke about an AI doctor’, one of likely thousands of silly requests that people experimenting with the large language model make every minute.

Full disclosure: ChatGPT generated the title of this post and the joke. But everything from here on in is made by a standard-issue human being.

Potentials and Dangers of Artificial Intelligence

Two #LINO23 sessions on Tuesday, 27 June 2023, dug deep into the capabilities, potential, dangers and even the existential questions that the large language model has brought up since version 3.5 was released to the world in November 2022. The first was an Agora Talk titled Impact of Biological Intelligence (BI), Human Intelligence (HI) & Machine Intelligence (AI) on Innovation in Science & Technology, with 2013 Nobel Prize in Chemistry recipient Michael Levitt.

Levitt began by summarising his thoughts on intelligence: “There really is a triad of intelligences: biological intelligence, which I think is by far the greatest intelligence; the intelligence that we humans have; and then finally, machine intelligence or AI,” he said. “And these intelligences interact: biology made humans, humans have made machines, machines will feed back both into new biology and into changing how human beings are.”

After an all-too-brief lecture expanding on each of these types of intelligences, skirting past Levitt’s own contributions to understanding biological and human intelligence, the large majority of the time was left open to what became an AI-focused Q&A with the audience. Two microphones were placed in the auditorium and attendees lined up in their droves to ask Levitt their burning questions.

Can we use large language models like ChatGPT and AlphaFold to really probe into the mechanisms of protein folding? “I would love an addition to AlphaFold where you could actually ask it, why? At least then you could see where the differences are between the best model and the second-best model,” he replied. “I think we need to learn this because we do need the understanding.”

What are your thoughts on the impact of general usage of AI tools for society? “I think it’s going to have an impact on the stock market. I think it’s going to have an impact on the legal system,” Levitt replied. “Where I really want to see ChatGPT have an impact is risk assessment. In a crisis, it’s difficult to do a risk assessment because you’ve got to say, what is the cost of doing this versus the cost of doing that? Computers could be very good at that.”

Does the rapid evolution of AI mean we need to redefine the meaning of what is alive? “The current models are not alive because they don’t control their own energy source. So in some ways, they are maybe as alive as a virus,” responded Levitt. “Should we be worried? I believe that a human being plus AI will always be better than a human being by themselves, or AI by itself.”

After the Q&A, Levitt was surrounded by Young Scientists keen to talk more about ChatGPT and AI. These conversations probably extended through lunch until the bell for the next session on AI sounded in the afternoon.

Serious Inroads Into Medical Practice

Given the enthusiasm expressed at Levitt’s Agora Talk, it was little surprise that #LINO23 attendees flocked to the ‘Artificial Intelligence and Medicine’ Panel Discussion. They were greeted by moderator Bart de Witte, who gladly gave his time on his birthday to discuss the current uses, opportunities and risks of AI in medicine with an esteemed panel and audience.

On the stage, de Witte was joined by Young Scientists Ang Cui (Harvard University, USA) and Aderonke Sakpere (University of Ibadan, Nigeria), as well as Shwetak Patel (Google, who will present the Heidelberg Lecture 2023 on Wednesday, 28 June), and Nobel Laureates Aaron Ciechanover (2004 Nobel Prize in Chemistry), Avram Hershko (2004 Nobel Prize in Chemistry) and, of course, Michael Levitt.

AI and machine learning tools are already making serious inroads into medical practice, assisting in a host of areas including breast cancer screening, skin cancer classification and in predicting the progression of diabetes. What’s more, ChatGPT and its peers are widely seen as key tools in future medical care; for example, potentially allowing patients to receive instant advice over symptoms or providing physicians with assurance to avoid misdiagnosis. These and other factors are why medicine is seen as one of the most important applications of AI.

This view was confirmed by the panellists, who had all used and benefitted from machine learning and AI in their research, with Cui going even further: “Machine learning has really penetrated into every field…we have all been using it knowingly or unknowingly for every facet of our research.”

Her view of ChatGPT and similar large language models was that they will eventually enable scientists to probe deeper questions than ever before. “With the arrival of ChatGPT, I think we can begin to dive deeper into mechanistic understanding with AI,” she enthused. “And then we can ask more ambitious questions with AI data, like how to cure cancer. We can begin to solve some really big problems.”

Patel was similarly positive. “These new models are what I call label-efficient, and what that means is that you can have lots of data, but very few labels, and it can be just as performant as large labelled datasets,” he explained. “What does that mean for us in health? Well, it means that we may be able to solve the longtail problem, all those rare diseases. If you can have a candidate solution for the longtail now, you’re trying to inherently create a more equitable set of solutions… And so I think AI can actually be potentially the great equaliser.”

Of course, as with any revolutionary technology, benefits need to be weighed against risks. And all panellists were keen to highlight the major pitfalls. “One problem that I can see is overreliance on artificial intelligence; so when a doctor will get used to using artificial intelligence for diagnosis, for example, he can become over-reliant on it and may not use his own judgement,” opined Hershko. “Then there is the problem of accountability – if there is a mistake, then what do you do? Do you sue your computer? Do you sue the company?”

Global Data

Sakpere, meanwhile, argued that open data is key to equitable AI: “As long as we have openness, it enables us to contribute to global health data, so when global data has been trained, there won’t be any form of bias, with certain groups of people not being represented.”

To ensure medical AI avoids these and other dangers, and works for humanity, all panellists agreed that regulation is critical, with Levitt even coming up with a three-point plan for its introduction: explainability, transparency and steerability. “Let’s say you have a tool for medical diagnosis,” he said. “It should lay out why it came to its conclusion, it should basically be transparent about where the conclusion came from, and you should be able to say ‘well, now try something different’.”