AIhub coffee corner: Can AI make humans better?

The AIhub coffee corner captures the musings of AI experts over a short conversation. This month, we ask if AI can make humans better. And, if so, how?

Joining the discussion this time are: Joe Daly (AIhub and University of Bristol), Tom Dietterich (Oregon State University), Sabine Hauert (University of Bristol), Sarit Kraus (Bar-Ilan University), Michael Littman (Brown University), Lucy Smith (AIhub) and Oskar von Stryk (Technische Universität Darmstadt).

Joe Daly: I recently saw this Twitter thread, about how AI has made human players better at the game of Go, then this article about the game of bridge, and more generally about AI’s influence on us. People were actually discussing how AI can make us better at stuff, and how we can learn things from AI. What are people’s thoughts on that? Can this extend beyond game situations?

Sabine Hauert: There’s always this fear that AI will replace us, and will be able to do tasks better than us, and as a result we won’t be doing those tasks anymore. So we might either lose skills, deskill, or not find areas where we could be as useful. I like these types of stories that show ways in which AI helps improve the way we do things as humans, because they demonstrate the win-win, and the ability of AI to improve us. If we use AI for areas where we might want to have extra skills or extra knowledge, that extra nudge makes us better at doing what we care about.

As Joe mentioned, this discussion was seeded by AI Go playing and the fact that Go players have learned new moves and are improving themselves at the game of Go. The AI system has taught them new skills. We often use games as examples of AI helping us to improve our skills. Are there other examples where it’s not a game?

Michael Littman: I can think of several other games…

Sabine: Chess, I guess. Just the fact that you can practise against an AI probably improves your game.

Joe: What about other creative pursuits like art? There are all these AI art generators.

Michael: That’s a really good example. I know that some writers have been playing around with GPT-3 as a way of opening up their creative floodgates. They’ll start writing, then ask “what do you think GPT-3?” GPT-3 will go in some bizarre direction, which helps get them into a new part of their mental space. Many are reporting that this is helpful to them.

Sarit Kraus: I want to say two things. One, there is research about identifying survivors in a search and rescue situation. There is a very good deep learning base model to recognize from images whether there is a survivor in the field or not. There was an experiment where they ran trials using just the model, just humans, and then a combination of both. In this last example, the system suggests “boxes” where survivors could be, and then the humans look for them. As you can imagine, putting them together [the human and the model], so the system could pose some possible location and the human look more carefully, reached the best result. So, this is an example where humans and automated systems are working really closely together and get better results than each of them separately. There are also similar examples in medical scans.

Another example concerns autonomous cars. One of the requirements is that eventually there will be some human operators that will be able to take over when something goes wrong. So it’s again some integration of human capabilities and automated AI systems together, which I think is the way to go.

Tom Dietterich: Michael led a study a few years ago on what are called centaur systems – combined human-AI systems. Ben Shneiderman is one of the gods of human-computer interaction, and he’s on a big campaign to try to define a new field which is not artificial intelligence, but it is a merger of HCI and AI. The idea is to create power tools for humans to empower people. I mean, when we think about robotics, we think about exoskeletons. And if we think about materials science, now we have machine learning tools that help materials scientists be twenty times more effective when designing new compounds. So it would be interesting to work up some metrics in that space. How do we measure improvements in human productivity because they’re using an AI tool, not because they have just learned and internalized from the AI tool, but they actually use it? So, there is this Centaur chess, where it’s humans plus machines playing against humans plus machines. That is what Michael looked at in the study he did three or four years ago.

Sabine: So, one aspect in that example of Go, that we don’t necessarily see in the medical examples Sarit gave, is competition. In the way we’re setting up many of these tools it’s AI and humans working together to do something. It’s not AI versus humans competing, except maybe in RoboCup. This competition then pushes humans to be better. So, if in the context of the medical diagnosis that Sarit was mentioning, you actually competed against a machine to find more tumors or to do something, would the doctor overall be a better doctor? Or is that not the aim? Is the aim to have the right tools, so that you can use both their expertise. Or is competition essential for improvement?

Oskar von Stryk: You perhaps know that there is the Cybathlon, a competition to develop smart wearable robotics devices to help people with disabilities.

I would add another very general thought. We humans have a long tradition of integrating tools into our body schema, so that they feel like parts of our body. When we go skiing or roller skating, we train, we use the tool, and then we feel that it’s part of our body. When very enthusiastic car drivers have to sit in the passenger seat, they feel separated from their body extension, and really unhappy. The way we integrate tools into our body schema means we interact with the environment as an augmented system. This brought me to the idea that we talk about AI in a more general sense, which could expand generally to other parts of our body and brain, expand them with further capabilities, which go far beyond, for example, navigation systems and so on. But, in a way, I would say that our natural navigation capability is degrading by leaning on electronic tools. However, we’ve come to a higher level together in the synthesis of these.

Sabine: That means the overall sum of both partners is better but we might be losing something in the process then, in terms of skills?

Oskar: Yes, exactly. We have a new major project with the German Science Foundation to develop new types of powered prostheses and exoskeletons, which can be much better integrated into the body schema. Because, the problem is, if you have a new powered powered prosthesis that is supposed to help you climb the stairs, it needs to decide when you need which kind of force and support. How do you develop the AI which decides when to give you which kind of support? This is a very interesting fundamental question we’re addressing in this new project, which combines psychology, medicine, computer science, robotics, and mechatronics. In this case we want to use these tools, these new types of wearable robotics devices, for two things: firstly, to compensate for impaired abilities, e.g., in elderly people, and secondly, to get new insights into how the human body and brain work. And maybe this could be carried over to some extent to other types of AI systems.

Sabine: Joe, is there anything from your research, because you’ve looked at the human-robot interaction?

Joe: I mean, there’s definitely things related to people taking the advice of a robot. This actually links to some of the stuff in our interview with Maria De-Arteaga. She looked at decision making in child welfare, and whether people take robots’ or AI’s advice. This is also linked to what Sarit was talking about in terms of spotting things in scans, and so I think there’s definitely some interesting psychology there in whether people do take AI’s advice.

Sabine: That’s interesting, because in that case the AI is designed to teach, to an extent. Whereas, in the Go case, it’s the competition that’s the incentive for the human to learn to beat the system.

If you had to learn a skill better, what would you like to learn, and would you have an AI to help you do that? Personally, the sports you mentioned, Oskar, I’d like to learn some of those.

Tom: How about language. And with that, AI to help us learn languages.

Sarit: I’ve been playing Diplomacy for many years. I had a project in my PhD about diplomacy, and for many years no one was interested. Recently, there has been a lot of interest in this game. It’s a way to develop people’s ability to negotiate and to form agreements. The problem is that it’s very difficult to organize such a game because you need seven players. Now, if some of the players were automated that could be beneficial – you could play whenever you wish. When I’ve developed an AI agent, it would be good to play with it, and train my negotiation skills.

Sabine: That’s excellent. I play Catan and it would be good to have an AI partner. Also, something like singing or jamming, where the AI brings you into different areas and you have to follow along and invent your own song or music to go along with it. That would be quite fun.

Michael: I guess, in that vein, sight-singing is something I’ve always wanted help with: Here are the notes on the page, can I name the melody from just looking at the picture of it? I know musicians can do that, and I would like to be able to do that. I need software to give me the notes, then see if I was right. It’s not AI-ish necessarily, although picking out the note that you just sang is a little AI-ish – it’s perception.

Lucy Smith: I was thinking maybe a system to give me a personalized training plan to make me ride my bike faster, or for longer. Taking in your pulse rate, and your recovery rate, and suggesting how far you should ride, and how fast, on any particular day. I think apps to do this do actually exist already in some form.

Joe: I think I would probably go back to the art again. Like Lucy said, for a personalized plan for bike riding, something similar for art would be good. Like, what area of art do I need to improve, is it perspective or composition, that kind of thing.

AIhub is dedicated to free high-quality information about AI.

AIhub is dedicated to free high-quality information about AI.