What humans get (and don’t get) from AI

Note: Transcripts are automatically generated from episode audio, and are not fully corrected for spelling, grammar, and formatting.

PATTIE MAES: I always ask why are we developing AGI. Why is this the goal? Why don’t we think about what type of AI we want for humanity? A lot of these developers even believe that AI poses a risk for humanity. Yet, even though they believe that, they still want to develop it.

RANA EL KALIOUBY: Pattie Maes has been asking some variation of these questions for the past 30 years. While companies pour billions of dollars into the elusive pursuit of artificial general intelligence or AGI, Pattie wants to rethink our approach.

Is AI that’s as smart or even smarter than humans what we really need? Or should we redirect and focus on how AI can augment us?

MAES: AI is already and will definitely impact everybody in a major way. And so we cannot leave it just up to Silicon Valley or entrepreneurs and engineers to decide what kind of future with AI we want.

EL KALIOUBY: Pattie is a legend in the field of AI. Originally from Belgium, she moved to the US and set roots at MIT in the 80s and 90s, where she was often the only woman in the room. She’s now a professor at the MIT Media Lab. There, she leads the Fluid Interfaces research group and co-leads a brand new program at MIT called AHA short for Advancing Humans with AI.

I’ve known Pattie for years. Her research group at MIT recently published a news-breaking study exploring whether people’s use of and in some cases dependency on AI could lead to a decline in their ability to think critically. Pattie is exploring the impact of AI on humanity. I am so excited to share my conversation with her. We talk about what it means to benchmark human flourishing, the future of AI interfaces, and risks in commercializing conversational AI.

I’m Rana el Kaliouby and this is Pioneers of AI – a podcast taking you behind the scenes of the AI revolution.

[THEME MUSIC]

Hi Pattie. Welcome to Pioneers of AI.

MAES: Pleasure to be here.

Copy LinkHow early AI assistants became the foundation for recommendation systems

EL KALIOUBY: So I am so excited for our conversation. I have gotten to see your work firsthand at MIT Media Lab back when I was a postdoc there, and I’m a huge fan. But I’d love to give our audience a sense of the incredible work and career you’ve had, especially in AI. You’re one of the OGs in this space and I thought a good place to start is back in the nineties when you and your team invented and then actually commercialized collaborative filtering. So can you tell us what that is?

MAES: Yeah. So back when I became a sort of a faculty member at the MIT Media Lab, I was motivated to use AI technologies to benefit people basically. I became a mom in 94 and I had a very busy life, was trying to take care of a million things and so on.

And so I thought, well, maybe computers could help me make my life a little bit easier and maybe I could use AI too. So we invented, actually at that time, software agents, believe it or not, in 94. I wrote my first paper about the idea that we need to build software agents that help us with task overload and information overload and just help us keep track of all of these different things.

EL KALIOUBY: AI agents, what we call AI agents today.

MAES: Yes. So we were a little bit too early. But yeah, so one of the things that I was curious about and that my team was curious about was, could AI systems and could computers really help us with finding information, finding media like books and movies that we may be interested in. And so we figured that we could actually build a system that would exploit the taste of other people in a way, or would sort of spread word of mouth among people that have similar taste to help them find things that they may be interested in. So we call that collaborative filtering, because basically you are benefiting from other people, sort of an implicit collaboration between people to find things that you may want to look at or read or listen to.

EL KALIOUBY: Would you also say that one of the maybe unintended consequences of these types of algorithms is that people can get stuck in their echo chambers or their little bubbles?

MAES: Totally. We did anticipate that, believe it or not. And we always recommended that a recommendation system should not just give you what you like or more stuff like what you like, but that it should also, for example, tell you how rare your taste is or how out there compared to other people, or that it should also give you things that you currently don’t seem to be interested in, just to help you explore different things from what you are already interested in.

Unfortunately, what we have learned over and over again is that you may have good intentions in the research community with inventing a technology like recommendation systems or later others came up with social media. Same story there — basically it was invented for a good purpose, but then when it gets put into practice and people make a commercial system, the way that that company that provides the service makes money.

Ultimately determines how the technology is used. And in the case of a lot of media sites, social media sites, of course these companies make money with advertising. And so their incentive is to give you more things that you are likely to click on.

And so that indeed leads to people drifting towards extremes really, because of course, the most extreme information along certain lines is what makes people the most excited and makes it most likely that they will read it or click on it, and so on. So yeah, it’s a very hard lesson learned that we may have good intentions with building a technology, thinking that it will be deployed in a way that will benefit people, but that is not always the case when it gets commercialized.

Copy LinkDesigning AI to support daily life and real learning

EL KALIOUBY: I wanna come back to that thought when we dive into AI, obviously, because it’s still a very relevant concern. But before we do that, you lead the Fluid Interfaces group at the MIT Media Lab. What is the focus of your group and perhaps give us some of your favorite projects that have come out of the group.

MAES: Yeah, so all of my research is about human computer interaction and specifically these days human AI interaction. And we do all sorts of things. One of the things we do is studies of how people interact with AI and what the consequences are, the impact of sort of using AI day in, day out. But we also do more creative work where we build prototypes of AI systems that we think really can benefit people.

And, for example, two of the sort of user groups that we specifically focus on are the elderly, which we think basically — well, it’s a growing number, a growing segment of the population in developed countries. And they desire to live in their own homes for longer, but their cognitive abilities decline leading to safety issues and other issues.

So we think that AI can play a huge role in helping them live independently in their own homes for longer. So that’s actually one area that we focus on — wearable AI-enabled devices that help people with safety in the home with memory, like telling you, yes, you already took your medication, you shouldn’t take it again. Or telling you things like you turned on the stove 10 minutes ago. Remember to turn off the stove and so on.

EL KALIOUBY: I need that. I need that memory, yeah.

MAES: But another group that we focus on is kids and students in general.

We think that so far of course the impact of AI on learning has not been very positive. It’s clear that yes, students can turn in great homework, but they delegate their thinking and their homework to AI and they don’t really learn anything in the process. And ultimately that’s what the goal is, I think, of education — not to come up with that resulting paper that looks good or essay, but really to learn how to write properly.

So we have been doing a lot of work on building AI interfaces that are very different from the current systems. Of course the current chatbots — you give them a prompt, a question, and then they give you a very authoritative answer, very long and complete, et cetera. And we have been experimenting with AI systems that, for example, act a little bit more like a Socratic tutor basically. Instead of giving the student all the answers, the systems actually engage the user by asking questions back of the user that result in the student thinking for themselves and being engaged with the material.

A great teacher doesn’t do the work for you. A great teacher makes you — or supports you — in learning and developing your skills and coming up with the answer to a particular question.

EL KALIOUBY: I think my son is 16 and a half and he’s very AI forward. And so he uses a lot of these AI tools, not really as a shortcut to doing work. But I think it’s cool that he’s kind of leaning into what AI can do. Like I would not want him not to use AI at all.

I think it’s actually important, and it’s cool that he’s at the forefront of these technologies and kind of trying to see what their limits are. But I do worry that he is delegating work to AI, and that’s not what you wanna do. Yeah.

MAES: Yeah. And I think it’s possible to build AI systems that still rely on all this knowledge that we now have in these amazing foundation models, but that actually interact with the user or the student in this case in a different way and keep in mind what ultimately the goals are of this whole exchange or this whole interaction.

EL KALIOUBY: We’re going to take a short break.

[AD BREAK]

Copy LinkWhy intelligence augmentation matters more than artificial general intelligence

EL KALIOUBY: When we come back, how to actually measure human flourishing in AI and why you should still call your mom for that recipe, even if ChatGPT can give you the answer. You have a very particular view on AI and you actually refer to it as IA, intelligence augmented. Tell us more about that and what does it mean in practice?

MAES: For decades now, I’ve been arguing that we should really not aim to build the smartest possible machine that replaces us in many ways, but we should really use these same technologies to augment people or to help people become a better version of themselves.

EL KALIOUBY: Yeah. As you know, I recently launched BT Live Ventures. And our investment thesis is human-centric AI that augments and amplifies human abilities rather than replace them because I believe that there is a path for AI where it can help you be happier, more productive, more knowledgeable, more connected, more empathetic. Do you think this is a contrarian view?

MAES: Well, it’s definitely not the dominant view. The whole AI R&D community is so focused on building evermore powerful AI models. And so they really completely focus on the technical capabilities of these systems. How can we make them more accurate? How can we make them more efficient? How can we make them more safe?

Less biased, maybe as well. But we think that it’s important to reflect on what will happen when these super smart systems are put into the hands of people and people rely on them day in, day out, and they mediate their entire experience basically. And that is not something that these companies are currently focused on, unfortunately. It’s really a little bit more of an afterthought. Sometimes they think that human computer interaction or the interface or the human impact of a technology is sort of a problem that you can deal with later.

And we think that that is not the case. So far all these different foundation models typically get evaluated with benchmarks that show how good a system is at physics problems or other types of hard problems — involving a medical exam, for example.

Exactly. But we think that these AI models should also be evaluated on the human impacts of what happens to people when basically they rely on these systems day in, day out. Do people become more lonely? Do they socialize with other people less? Do they think less for themselves?

Do they overly rely on the AI and trust it too much when they shouldn’t. So we think that benchmarks are needed that test for the human impact of AI models, and that’s actually one of our goals with our research at the MIT Media Lab.

Copy LinkMeasuring AI by human flourishing instead of technical prowess

EL KALIOUBY: Yeah, you’ve been advocating for this human flourishing benchmark, which I love the idea because so far I’ve not seen anybody do that. What would the process look like if you were to release such a benchmark and how would you roll it out and what would a human flourishing score look like?

MAES: Well, we actually started doing this work in one area already. And that is the impact of AI on social relationships and loneliness. And what we did is we looked, actually together with OpenAI, we compared different models that they had to understand what the impact was on loneliness.

On dependency, problematic use of the AI as well as human socialization of using these different models on a daily basis. In our case, we had a thousand people that signed up for this study and were put in a particular condition. They used a particular model and had instructions about maybe what to talk about or not talk about, et cetera. So we had all these controlled conditions so we could really evaluate — if they did this for a whole month, we could see differences between the different groups basically, in terms of who becomes more lonely, who socializes less with people, who becomes too dependent on the AI and so on.

EL KALIOUBY: What were your findings from the study?

MAES: So we learned that the people who used these chatbots the most every day for the longest period of time tended to have worse outcomes. They tended to become more lonely, socialize less with other people, also become more dependent on the AI saying that they couldn’t really do without it and more.

EL KALIOUBY: Yeah. And so I imagine with a human flourishing benchmark, you would flag these kinds of — like, it would be considered as part of how we score these models. And I’m thinking when the product ties into it too, right? Like, can we put guardrails so that if you’re spending six hours a day talking to AI, then it can say something.

MAES: Yeah. So I think that, without of course impacting people’s privacy, we want to have possibly classifiers that run in the background and flag the company when there is some behavior or some exchanges that seem to be really problematic. Yeah.

Copy LinkHow chatbot dependency can deepen loneliness and distort reality

EL KALIOUBY: Yeah. But again, back to kind of the collaborative filtering and recommendation engines, this may not be aligned with the incentive of the company, which wants you to maximize usage.

MAES: Yes. Yeah. So we learned with social media that because of the advertising model, basically the incentives of the company and the user are not aligned. I am worried that with AI the situation is only going to be worse. With social media, we saw polarization of people, for example political polarization, but with AI, basically I think we risk getting bubbles of one, where basically you and your AI keep echoing back to one another certain points of view or beliefs and so on, and they can spiral into a domain of fantasy. There’s already a lot of cases about this. I see every day practically, I get an email from one or another person who says that they have seen the light and they are the chosen one, and they have discovered that AI is really sentient and we have to protect these AIs from being shut down.

And so it’s really — when people sort of start talking a little bit along those lines, often AI really acts like an echo chamber or a mirror and it starts reinforcing whatever it is that you tell it. So if you tell it that you have some belief that you may be special, then it starts saying, yes, you are actually special.

EL KALIOUBY: I hope that companies are building guardrails against this kind of really unhealthy patterns of behavior, but it sounds like this is ongoing, right?

MAES: Yes, I think there’s very few guardrails right now, and that is why we believe that we need to come up with these benchmarks and create a lot of buzz and awareness around these human impact or human flourishing with AI types of benchmarks so that people can say, well, I think I’ll use this model rather than that model because this model doesn’t have those same possible negative outcomes.

Copy LinkWhy overreliance on AI can weaken the social fabric

EL KALIOUBY: Now you’ve also sounded the alarm around AI. I loved how you worded it. I heard you talk about this at MIT, where you were worried that it would unravel the moral fabric of society. Because as we rely more and more on AI, and we’re not going to — like, I’ll give you an example from my personal life.

Like every time I wanna make a recipe, I would sometimes FaceTime my mom. I’m like, mom, like here, I’m trying to make this thing. How do I do it? And now I’ll just go to AI for advice on various big or little things in my life. And if you kind of compound that at the societal level, we’re not tapping into each other’s friendships and relationships. And what does that do? So tell us more about that. Yeah.

MAES: Indeed. I work of course with a lot of young people at MIT who are eager to adopt technology and many of them talk to a chatbot for hours every day. And they ask it for help with all sorts of things, not just work — say, research and programming and things like that, but also mental health questions, physical health questions.

Questions about relationships and how they should deal with them and so on. So we are essentially reducing the amount of human contacts we have, and I worry about that for two reasons actually. Social scientists have talked about two types of social contact and social relationships that are important. There are strong ties, which is the people who know you the best — your closest friends and family members basically.

And then there are the weak social ties, and weak social ties are the people you interact with because you see them at the post office or in a restaurant. Both types of relationships are actually very important for our society to function properly.

Of course, the strong social ties are very important because they can support you emotionally when you have problems. They’re like your close network that you can always count on. But the weak relationships are also very important because the weak relationships are the ones where you basically are confronted with people who might be different than you.

Whether this is people from a different background, race, political party, interest, whatever. If you have a lot of weak social interactions, it’s actually beneficial because you realize like, oh, these people who are reading very different news than me or belong to a different political party — actually they’re not bad, they also mean well and love their kids and want the best, et cetera.

So it helps you relate to other people and it helps you see other points of view. And it is really a way we can learn and ultimately become more wise. But also it helps us come up with combined solutions that really represent many different points of view.

So I’m worried that with AI, we risk actually reducing both the strong ties and the weak ties, and that could really have even larger, much larger implications than social media has had on our society.

EL KALIOUBY: When we talk about the dangers around generative AI, we’re mostly talking about chatbots. Because right now, most people interface with AI via a chatbot or voice on their laptop or smartphone. So what does the next AI-native interface look like and how do we ensure that it’s safe? That and more in a minute.

[AD BREAK]

Copy LinkWhat an AI native interface could look like beyond the chatbot

EL KALIOUBY: So I wanna switch gears a bit to talk about the human machine interface and the human AI interface. You’ve spent many years thinking about what the most effective interface looks like. The de facto interface for AI today is basically a chatbot and maybe it’s a little conversational, you can kind of talk to it with voice. But I don’t think this is an AI-native interface at all. And of course recently we saw OpenAI partner with Johnny Ive, who was kind of one of the early designers of the iPhone, to think about what is that next device, hardware device or next interface look like. I am so curious, what are your thoughts on what an AI-native interface looks like? And I’m curious if you’re exploring that space.

MAES: Oh yeah. We are definitely exploring this space — for a while already actually. I do a lot of work on wearable devices and I’ve always been frustrated that the dominant device today, the smartphone, is so disruptive. If you want to quickly look up some information because it might be helpful in the context of a conversation or something, you have to completely divert your attention from the other person to open up your phone, find the application, look up the information, read the results, et cetera. And so it’s very disruptive to conversations. But it also may end up resulting in you bumping into a telephone pole while you’re walking to the subway, things like that. So we’ve always been exploring what techniques we can use to build systems that are less disruptive and that require less input, less output, effort basically on behalf of a person.

And there are many techniques that can be used. Of course they also have trade offs and negative consequences. For example, basically the person carries a device in their pocket, in their shirt, and the device is always in contact with them, talking to them basically in natural language.

But it’s also — in addition to today’s chatbots — it’s aware of their context. So it can see what they can see, where they are, for example, and can talk about their surroundings. It also is aware of their internal context or their internal state. Maybe they’re tired or a little bit anxious or sad, et cetera.

And it can take all of that into account in its interactions with a person. And of course, if you have a system like that — one that is aware of your context — it doesn’t necessarily need much effort to communicate with. It’s kind of like a spouse that you’ve been married to for 20 years, where all you have to do is wink and they understand what it is that you mean. You just need that wink instead of a long interaction to communicate.

So I think that’s one sort of direction that these companies are going into and not just OpenAI — really a lot of all the other companies are doing related work. So mostly actually developing glasses. Meta, for example, has for a long time been developing AI-enabled glasses with RayBan, for example, but they have much more sophisticated glasses internally, in their research domain as well.

And those glasses see the world around you, but they can also look at your eye and at your pupil and know what it is that you are looking at. So if a system sees that you are looking at a particular box of cookies or something and if the system knows, say, that you are allergic to peanuts, it can be proactive in its support.

So you don’t even have to ask it any question. It can just say, oh, don’t touch those — they actually have peanut traces in them. We actually have been building systems of this sort for a long time. We have a system right now, Memorial, that listens to your conversations, but it stores them locally.

And then if you forget something like the name of a person you were just introduced to, you can just say what is his name again? And the system will say, Robert. So it can basically complement your memory, and in some cases it can even anticipate your memory needs. We have done experiments in the lab with eye tracking and looking at the pupil of the eyes and you can actually detect with almost 90% accuracy when a person is not going to remember the name of a person in front of them.

So systems that can observe us and observe the environment can be very smart, can be very efficient in terms of being very helpful with very concise answers that are sufficient to trigger your memory.

Copy LinkBalancing wearable AI convenience with privacy and consent

EL KALIOUBY: Yeah, it’s so interesting. Now, of course, one of the implications of these kinds of wearables is privacy, right? So if I’m wearing my glasses or my pendant or whatever, and I walk into a social party, I would love to have the device say, oh, you met this person a year ago — Tricia, whatever. But you’re not really consenting all the people you’re interacting with, so what do you think are the privacy implications of that?

MAES: Yes. We’re actually dealing with that right now because this system, Memorial, we call it, records conversations to help the elderly with remembering things. The solution that we’re using right now is that they literally wear a big button saying I’m recording. Another part of the solution is that all of the data is local. It’s stored locally rather than in a server. We actually had to implement ways for people to delete memories. So ultimately I think both parties in a conversation have to give consent and have to have the right to delete conversations after the fact and request deletion. I think it is possible to build decentralized systems of this sort.

EL KALIOUBY: Yeah, so fascinating because I also wonder if we’re gonna see a shift in what’s acceptable, right? In the same way that you can now walk into a wedding and record — it’s not your wedding, but you’re just recording and uploading to social media and it’s kind of become a norm. I wonder if we’ll see an evolution.

MAES: Yeah, unfortunately the companies that make these systems, they tend to get your permission by building in some very exciting, positive features that then sort of make you say like, oh yeah, what the heck, I’m just gonna let myself be recorded. But then that may have unintended consequences much later and over time and so on.

EL KALIOUBY: With AI becoming more knowledgeable, smarter, more conversational and perceptual and even maybe empathetic and creative, what do you think it means to be human in the age of AI?

MAES: Yeah, so I hope we can create AI that helps us be more human and really supports us in becoming the best version that we could be really of ourselves — an AI that helps us with all of these issues that we may struggle with.

Having enough motivation to learn, or becoming better at interacting with other people and seeing their point of view. I think ultimately that AI helps improve our society, our communities, helps us find solutions collaboratively and more. But I’m probably a little bit too much of an idealist, but I’m trying to make a lot of noise so as to inspire people to follow that same vision.

EL KALIOUBY: I love that vision, this idea that AI brings the best of us as humans and the best versions of ourselves. I love it. Thank you so much for joining us, Pattie. That was awesome.

MAES: Thank you, Ranna.

EL KALIOUBY: In Pattie’s mind, the future of AI is still up for grabs. We’re not destined for an AGI obsessed future if we shift the focus on who’s actually at stake – us. And there’s a lot we can learn from our past mistakes when it comes to other innovations like social media.

AI may look different in a few years – whether that means a pair of glasses or a tasteful pendant. But no matter the form factor, better guardrails, more transparency, and benchmarks that measure human impact can bring us more human-forward AI.