SPEAKER: Thank you, Joey, for that lovely introduction. I'm so honored and thrilled to be here and be part of welcoming you to the Harvard community. I want to echo President Garber's remarks earlier about how fun it is for a student to walk into my office and say, hey, I'm concentrating in classics and computer science, and I need to write a thesis at the intersection of the two. And I'm like, you've come to the right place. But today, I want to talk to you about something because I specialize in human computer interaction. It's kind of like applied psychology for computers and the people that interact with them. I want to talk to you about something you're probably already doing every day, which is interacting with AI. And I'm particularly interested in how those interactions are, whether they are accelerating you towards what you are ultimately going to do if you had infinite time and resources or whether they are maybe in subtle and not so subtle ways affecting the goals that you choose and the quality of the decisions and the output that you ultimately get with them. So if you look at the research today, which is exploding, there's a lot of different terms that you could search should you want to look up some papers. I'm very proud that my field of human computer action is very accessible. So I will often, in my classes at the undergraduate level, ask you to engage with the latest scholarship. And you can do that even before you get here. But you can search for human-AI, teaming, human-AI, co-creation, AI-assisted creativity, or AI-assisted decision making, among others. So I thought, well, maybe I can ask the AI to help me generate some images to go along with this slide. And I usually don't. But I thought, along with the theme, I asked it to generate an image of a programmer with a robot looking over its shoulder at their code because one of the things that I study is AI-assisted programming, which is quite relevant to the students in my department of computer science. And you can see it generated this. Anyone want to shout out some things you notice about this AI-generated image? SPEAKER 2: Not over his shoulder. SPEAKER 1: No. It does say it's looking over his shoulder, but it's clearly not. It's kind of creepily looking at the viewer. What else do you notice? SPEAKER 3: Robot is a white male. SPEAKER 1: Yes, it is. Yes, he is. Two in one right there. Any other? There's one more you might notice. SPEAKER 4: It's in the dark. SPEAKER 1: It's in the dark. Programming only happens in the dark. I'm sorry? SPEAKER 4: The code is nonsense. SPEAKER 1: The code is also probably nonsense. And if you can read it, I'm impressed. So I thought, well, I know, in my mind, this is not quite what I wanted. I didn't know exactly what I wanted when I put this in, but this is not it. So I asked-- you can see up at the top. I said, the robot is looking at the programmer, not over the programmer's shoulder at the code. Can you fix that? A little better. It's still a white man. It's still in the dark. The robot's kind of looking at his code. All right. SPEAKER: And the robot looks the same too. SPEAKER: The robot has-- yes, it's a very consistent robot. There seems to be-- this particular model, given its training data that it is reflecting back to you, it's making some choices very consistently that don't necessarily reflect real life. So I said, please change the programmer to be a college-aged woman. I'm a woman. Many women in my class were college aged. And it generated this. OK. Well-- SPEAKER: Robot is actually looking at the code now. SPEAKER: The robot's now looking at the code. It's still dark. She's still white. I'm kind of going through and just-- it's not to say that aren't-- there are programmers who come from all different identities. Programming happens at, truly, all hours of the day. But it is reflecting, perhaps, maybe default assumptions that are reflected in its training data. And what I want to point out-- and I kind of kept going back and forth with it. Eventually, I landed on this one. And I'm like, all right, at this point, I'm going to stop. Because the point I'm trying to make here-- one of the FIRST most important points is that-- I put in a very generic specification, and it made a whole bunch of arbitrary choices on my behalf. And the choices that are consistent with my conscious or unconscious expectations are the ones I'm least likely to notice. And we cannot think critically about the choices that are made on our behalf-- especially algorithmic ones-- if we don't notice them first. And so I just want to flag that as something for us to keep in mind. But at this point, I thought, this is taking more time than it's supposed to, so I'm not going to be using AI generated. I'll just draw. OK. [LAUGHTER] All right. So another thing that matters is cognitive engagement. So even if you notice something, if you don't really, truly cognitively engage with it, again, that criticality of thought is not going to come through. I did a little-- well, actually, my student, [? Priyan-- ?] here on the right-- did a really nice study of some programmers. The first group didn't get any assistance. They just did a programming problem-- and, of course, some sort of approximate normal curve of how long it took them to finish. Then we added-- we said, OK, here's some AI. Do another one. And some people got shifted to the left, shorter time. They were like, ah, I'm just done. Thank you. AI is great. And the other guys never finished. What do you think happened? Does anyone want to guess? SPEAKER: They had a very specific thing they wanted, and the AI couldn't create exactly that. SPEAKER: That is not what we observed, but I definitely think that that's happening in other contexts. Any other guesses? SPEAKER: It got distracted. SPEAKER: Another excellent guess that they might-- OK, all right. I'll tell you what we think happened. But, literally, this is from-- you can see-- 2022. If you come to Harvard, we can do more empirical studies about this and figure out exactly what's going on. One of the things that happened in this context was that they were not cognitively engaging with the suggestions the AI was making so much. And so if they got lucky and they were just like, yep, looks good to me, looks good to me. Accept, Accept, accept. And they got code that passed the test cases, they were done early. And if they were unlucky and they accept, accept, looks good to me. Looks good to me. Looks good to me-- oh no. I have a lot of code that I didn't really write. I don't really know exactly what it does, and it's not passing the test cases. And now I will use up all the remainder of my time-- running out of time trying to debug code I didn't write, that's subtly wrong. But it looks right. Anyway, so I think that this has been one of our, surprisingly, most influential papers. One other thing that-- a follow-up work by my colleague, Krzysztof Gajos, with a colleague of his-- found that cognitive engagement is also necessary for you to even learn from the interaction. So if you're trying to get better at something, if you're not cognitively engaging when you're being assisted by AI, you may never gain additional competency. And I don't know about you guys, but I like getting better at stuff-- at least, for many things. It's OK if-- for some things-- I don't get better at navigating Cambridge. I'm OK, depending on my GPS. It's a little confusing. All right. So another domain you may already have experienced is AI-assisted writing and, really, in more simple contexts, like writing emails. Some of those emails can be pretty boilerplate. Like, not hard to guess that "are you" comes after "how" at the beginning of an email. But another colleague of mine found that predictable text encourages predictable writing. When you are predicting text, people will-- they are strategically lazy. That's OK. But they will write more efficiently and less thoughtfully. So if we're thinking about how the outcome of our work changes when we have AI assistants, in this case, the outcomes were less verbose but also less colorful, a little bit more stereotypical. So one thing that I found in a follow-up study was that it doesn't have to be this way. And so part of being in my field, human interaction, we can critique the systems and how they're currently working. But we're also engineers, so we can build better systems. We don't have to just point out the problems. We can try to solve them. So here, we had a study where we had people do creative writing to write a story. And we had AI-generated text that might be relevant to the evolution of their story, AI-generated images, AI-generated sounds. It was kind of a multimedia experience inspired by whatever they had been writing. The suggestions-- because this was two years ago-- were not as good as they are today. But the point is not how good they were. It's what people did with them. And because they were part of the environment and not directly-- additional gray text in the text buffer, people-- one person was writing about a situation. They weren't sure how to solve the plot point. And they saw that the system had generated the image of an elephant, which felt pretty random. But it was their human creativity, where they made it what we called an "integrative leap," where they said, ah, I will solve the problem that I've created in this plot. I realize that one of the characters has stolen an elephant, and this is why the detective has come to their door. So this was not limiting. This was helping them get past stuck points, but it was really them in the driver's seat, and it was not impacting the predictability of what they were writing. So just some other places where you may not think of being AI assisted in your decision making, but just so you call it out-- we receive driving recommendations-- even with metadata-- about the impact on the environment or our pocketbook when we're driving. The media that we consume may be recommended to us by algorithms that are surfacing some content they think would be more appealing to us than others. They may be wrong. Even when you come to college and perhaps engage in some of our data match and other algorithms for matching people-- we should continue to study the effects of these systems and how they affect us. Sometimes it's more easy to recognize the system not quite serving our needs correctly. Does anyone notice what happened here when I was trying to google for-- not use AI-generated stuff anymore, just see if I could find some clip art. SPEAKER: Humans and it's not humans-- SPEAKER: It's not humans. There-- SPEAKER: The first one's human. SPEAKER: There are two images on here that have humans interacting with humans working together, which was my query. The rest are humans interacting with robots. So anyway, I hope now that I'm pointing it out, you will continue to notice these things. You have to keep your brain on when you're interacting with these systems. Another really critical situation-- how many people have asked ChatGPT to summarize a document for you? Yeah, OK. Be careful. [LAUGHTER] SPEAKER: And then read the document and scrutinize-- SPEAKER: Exactly, you have to read the entire document. And that's tedious-- you should come to my class, OK? I'm serious. SPEAKER: I probably will. SPEAKER: So AI-generated summaries can omit something that it decides is not important enough but actually would fundamentally change how you interpret the sentence. They may hallucinate things that are plausible but not found in the original document. And they may even just subtly misrepresent something that's critical for you to understand correctly. And they're just onerous and memory intensive to assess. Many people don't read the entire original document. So when you're-- again, Google gives me so much material. I was googling-- a lot of my friends are also in their late 30s. Many of them are starting families. And I remembered sometime hearing that twins were more likely as you got older. So I googled twin probability with maternal age. And it said, "According to BabyCenter, the chances of having fraternal twins are 6.9% for women, 35 to 37." And I was like, wow, that's quite high. I should be meeting more twins. But I accepted this. And it wasn't until much later that I actually went and looked at the source. It turns out these are numbers-- left off for some reason women under 35-- these are numbers for people who are receiving assisted reproductive technology interventions where they intentionally put multiple embryos in to maximize the chances of a live birth. So this completely changes the interpretation. These are intentional as opposed to what I was asking about, which was some natural feature over reproductive age. So there was no information sent to signal to me the missing context that completely changed the meaning of the numbers that it gave me. And I know it says generative AI is experimental, but we have to be really, really careful. How do we design interfaces that help people have the information sent to-- or give them the context to notice when the AI is making a choice that doesn't make sense for their query or for their situation? So who's driving? Well, if you're not cognitively engaged or supported in noticing AI choices or given enough context to judge the AI choice well, the AI might be driving more than you think. But again, we're engineers. I'll tell you one solution we've come up recently called grammar-preserving text saliency modulation, which is a totally different approach. What we do is we recursively cut the least semantics-modifying words from a sentence. And then we reify that in the saliency of the sentence. So we initially cut just-- it said, "The world is at present accumulating carbon dioxide in the atmosphere from well-known two sources, the combustion of fossil fuels and deforestation." You can see how the later a word was cut while still preserving the grammaticality and the overall meaning of the sentence, the darker it remains in the original sentence. If you do this for an entire paragraph from the GRE, it looks like this. What's beautiful about this is that there are extractive summaries and multiple levels of granularity all within the same thing. You can change the level of detail as you read. And if it mistakenly makes less salient a piece of text that you realize for your context is exactly what you need to read you, you see it. It's still there. It's still legible. It's not hidden. So we call this an AI-resilient alternative to AI-generated summaries. We found that people were able to read faster and answer questions more accurately when reading this rather than normal text, as well as one other control condition which modulated text salience a different way. So, in summary, I think we should be inviting, facilitating, or even forcing-- even though it's unpleasant sometimes for users-- cognitive engagement. We need to design interfaces that help people better notice AI choices that are made on their behalf and, at the same time as it makes those choices visible, provides the context necessary for you to decide whether that choice is right for you. So, on that note, I will pass it on to our final speaker, who I'm very excited about. [APPLAUSE]