People resist AI even after they are shown that it works, and they resist it across uses that have little in common — generating art, giving medical advice, grading an essay. Pragmatic worry does not explain that pattern. In this Heterodox Academy virtual colloquium, University of Toronto psychologist Michael Inzlicht argues that much of the opposition to AI is moralized: a conviction that using AI is wrong, full stop, rather than unwise. Most people who oppose AI say their view would not budge even if the benefits grew and the risks shrank.
Inzlicht also shows what the alternative looks like, walking through his own cost-benefit work on AI-generated empathy, where AI responses were rated more compassionate than those written by people — including trained crisis-line responders. His talk closes on an issue that should interest anyone who cares about open inquiry: moralizing the empirical questions around AI ignores evidence, forgoes benefits, and forecloses discussion. His advice is to stay empirical — study AI, debate it, say how bad it is — but keep your views open to revision. A transcript of Inzlich’s talk follows below.
Dylan Selterman: Mickey Inzlicht is a professor in the Department of Psychology at the University of Toronto with a cross-appointment as professor in the Department of Marketing at the Rotman School of Management. He is also a research lead at the Schwartz Reisman Institute for Technology and Society. His research sits at the boundaries of social psychology and cognitive science, exploring the paradoxes of human motivation, particularly why people avoid and find meaning in mental effort and how digital technologies are reshaping behavior and well-being. He has pioneered research showing that empathy is cognitively demanding and often avoided because of its mental costs, challenging the common assumptions about human compassion.
His current work examines how exerting effort paradoxically increases feelings of meaning, how rapid content switching on digital platforms increases boredom, whether artificial intelligence can express empathy more effectively than humans, and the psychological effects of recreational cannabis use. Dr. Inzlicht completed his bachelor’s degree in anatomical sciences at McGill University, his PhD in experimental psychology at Brown University, and his postdoctoral fellowship in applied psychology at New York University. He has published more than 180 peer-reviewed journal articles and book chapters and edited two books with his work cited over 34,000 times.
His research has been featured in major media outlets including the New York Times, the Atlantic, the Guardian, NPR, the Washington Post, BBC News, the Globe and Mail, Time, Forbes, and Science, among many others. His research and teaching have been recognized with the Carol and Ed Diener Mid-Career Award in Social Psychology, the Wegner Theoretical Innovation Prize, the ISCON Best Social Cognition Paper Award, and Professor of the Year. He’s also been recognized as among the top 1% of most cited psychologists in the world for four consecutive years, from 2022 to 2025. He also co-hosts the podcast Two Psychologists Four Beers and writes the Substack newsletter Speak Now, Regret Later.
I’ll also share that I’ve long admired Mickey as an exemplar of a generalist, someone who is really actively curious and engaged with a lot of different topics. And as an aspiring generalist myself, it’s great to have role models like him to look up to. So without further ado, I will hand things off to Mickey. He’s going to give us his presentation, and then we’ll switch gears to live audience Q&A. Please use the hand-raise feature on Zoom to speak up when we get to the Q&A portion. If you must stay off camera and microphone, then you do have a backup option to type questions into the chat. All right, Mickey, take it away.
Michael Inzlicht: All right. Make sure you can hear me, Dylan? Yes? Okay, excellent. I’m going to minimize this so it’s not distracting to me. So first, thank you for the warm introduction, Dylan. Heterodox Academy, I joined it very early on when I just kind of got it started, but I haven’t really done much with my membership, but perhaps that will change. And maybe this is my first kind of foray into doing a bit more with Heterodox. I love the mission, and I’m happy to support it. So today I’m going to talk to you about something, a topic that’s close to my heart, because I feel I’ve been living it the past year or two. And that is the topic of the moralization of artificial intelligence. What I’m talking about today is really, can we still talk clearly?
Can we see clearly, think clearly when we talk about AI? And if you spent any time online, read newspaper headings in the past year or two, well, you know, AI is everywhere, but you also know that people have strong attitudes, strong views about AI, some of which are not positive. So here is one I’ll begin with. This is written by someone named Becca Rothfeld. Becca Rothfeld today is a New Yorker staff writer. Before then, she contributed at the Chronicle of Higher Education and also was at one point was a philosophy PhD student. And she wrote in an essay, a viral essay, I hope you’re ashamed of yourself and your dwindling humanity.
And I think it should be legally required for people to say when, if, they’ve used AI in the process of writing something so that I can avoid it at all costs. Now, I picked a not particularly outrageous quote, but this essay was full of outrageous quotes. She called people who use AI human-shaped voids. She said that if someone used AI, you know, if her husband used AI to say something to her, she would immediately divorce him. You know, it was hyperbolic. It was funny. I don’t think it’s meant to be taken tongue in cheek a little bit. But it’s clear she’s got strong moral intuitions about AI and she thinks AI is bad. Here’s another quote from Bluesky, a place with only reasonable discussion.
This is from Tage Rai, former editor at Science and now a professor at the University of California, San Diego. He wrote, I don’t think it’s worthwhile to engage with the possibility of AI being used for good, rhetorically, when the guys with power over the technology and who profit from it are white supremacists, eugenicists, and pedophiles. Now, I don’t know if Tage meant this literally or seriously. I suspect he was also being tongue-in-cheek, but he’s using the words of a QAnon supporter to describe a technology. And that’s odd. At the very least, it’s odd. And just to show you that the speech, like whether it’s Tage’s speech, Becca’s speech, or hyperbole, it has consequences. It doesn’t just stay on the internet.
A number of months ago, on the same weekend, these were headlines from U.S. news sources. So the San Francisco Standard, Sam Altman’s house targeted in second attack, two suspects arrested. That means — Sam Altman, who’s the CEO of OpenAI, the biggest AI company, his house was attacked on the same weekend, not once, but twice, with Molotov cocktails. Second headline from CBS News, Indianapolis city councilor says his home was shot at 13 times, no data centers sign left behind. Right, so a politician, a low level politician mind you, experienced violence in his own home because I guess that city council was considering whether they’d open up a tax-generating data center. What is going on here?
What is going on where a normal technology is causing so much outrage? How, like, this is, again, you don’t have to be concerned for this, although I think the last slide with the threats of violence should make you concerned. But at the very least, it should be, huh, this is weird. And, you know, it’s coming from what I would describe as normal technology. And then some of you maybe in the audience would respond: “Normal technology? AI is an amoral machine wearing the skin of a human. Nothing good could come of it. It is evil and should be destroyed.” And by the way, I’m paraphrasing something that someone wrote to me — I think it was on X. Clearly people have strong views here. So, kind of trying to understand what the hell is going on here.
So before kind of really getting to the heart of the matter about what I think is going on, I want to just show you another way, another way of kind of talking about AI that I believe is more responsible, is utilitarian. It’s a cost-benefit analysis of a normal technology, in this case of AI-generated empathy. And this is actually, like, I’ve done a lot of research on this right now. So I’m just going to give you, it’s a bit of a sidestep, a bit of a tangent, but I just want to kind of demonstrate how it could be done potentially. So I often will start talks on AI-generated empathy with noticing a few facts. First, people are lonely, often friendless. So we hear a lot about a loneliness epidemic.
I think to some extent that’s a bit overplayed, but I definitely agree that there is a, whether you want to call it an epidemic or not, there is a concern of friendlessness. So 8% of Americans, according to Pew, say they don’t have one single friend, not even one person who they could rely on, they could speak with who’s not a family member. This is more true for men than for women. It’s also more true for men over 40 or 50. I’m 54 — I just turned 54. So the percentage for me would be much higher. And that breaks my heart that there are people out there who have nobody, no one to turn to. And who are those people, by the way, who might be especially likely not to have friends? These are the people that society forgets. So we have the elderly.
I consult with a small tech company here in Toronto, where they’re a medical tech company and they deal with truly marginalized people. And their main mission is to help elderly people stay at home. Many of their clients do not see a single person ever other than the food delivery people. Some of these people, you know, the way even people realize they’ve died is because the food delivery person finds them. All right. That’s not uncommon. You’ve got people who live in the country, people who like they’re literally kilometers or for Americans, miles between their next neighbor. And, you know, it’s literally hard to find friends, find people, find community.
Or what about the people who literally are essential workers, who are keeping our cities running, who need to work in the middle of the night? That’s when a lot of maintenance happens. If you are sleeping during the day, where are your friends? Where’s your community? Where’s your place of worship? It’s very hard to connect with people. And one thing I wondered was, could a large language model, could an AI bridge the gap here? Could it help some people out? I don’t view this as a replacement for friends, replacement for therapists, replacement for community. But could it help people in need when they are in need? And again, thinking about those people who might be housebound, who don’t leave the home for disability reasons, for health reasons.
So we wondered. And with my, another graduate student I worked with, my lab manager, Daria Ovsyannikova, we ran a series of studies examining the quality of AI’s empathic utterances. So what did we do? We first created what we call empathy scenarios. By that I mean, we described situations where empathy might be evoked in an interlocutor. So you might hear a story of, you know, here’s an example of a story. I’m, you know, I’m worrying about my mother. She’s in her eighties, she lives alone. And last week I discovered that she got lost in the neighborhood she’s lived in for the past 30 years. And I don’t live in the same city as her. I’m just worried. That could be one empathy scenario.
Another one could be joy because you also feel empathy towards joy. I’m so proud of my son. He got a 95 on his physics exam. And I think he’s well on his way to get accepted in a great university. So we had these scenarios, five positive, five negative. And then we had, I think, about 10 to 20 humans respond to these scenarios. All we told the people were, read the scenario and respond with compassion using between 50 and 75 words. And we also told people, don’t give advice, just respond compassionately. And then we prompted, at this point, this is ChatGPT, I think it was 4o. And we prompted it iteratively over and over again. And then we generated five responses.
So I should also say that of the 10 or 20 humans we used, we picked the five best responses. And then a second group of participants now examined the responses produced by the humans, by the AI, and we asked them, which one is more compassionate? Which one do you think would make people feel loved and cared for, et cetera? And what we found was astonishing, at least at the time. So, what you see here are ratings of empathy as a function of empathy source, ratings of empathy is on the y-axis. You’ve got ratings of AI’s empathy and human empathy. And what you see is very clearly that the AI is producing words that people are saying are more empathic, more compassionate, leads them to feel something more than humans.
Now, that was true when people didn’t know who was producing the statements. So what happens when you make the generator of the statements transparent? You label it as AI, you label it as human, what happens? What we see are two effects. First, non-transparent, that’s when they don’t know, it’s not labeled. We see the same effect as above. AI is rated as being more empathic than humans. But it’s also true when the identity of the agent is known. AI is still beating humans, although, and very consistently, you see an AI penalty. As soon as people find out it’s an AI producing those statements, people rate it as being worse. And then finally, we wondered, well, these are just random humans, selected, mind you, for producing good statements.
What if we got expert empathizers? So we decided to get, we recruited crisis line responders who are trained to dole out compassion to people on telephones, also to make sure they’re safe, people aren’t engaging in self-harm. And we had them do the exact same thing, produce these responses. And what we found was identical to the above. The AI beat our expert humans, both when the identities were known and when they were not known. So this is an example of, you know, a study that we ran. And despite the positivity here, we can run a cost benefit analysis of AI empathy.
So pros could be, you can help people feel heard, you know, and AI is always available, even in the middle of the night, even for those shift workers, even when, you know, you’re at your home alone, it’s less biased, we’ve got lots of research now suggesting that it won’t respond with the same kinds of integral biases that humans do. And also, it doesn’t get tired. Humans, your human friends will eventually be tired of hearing you complain about the same thing over and over again. The LLM will not feel tired. Now, what are some of the cons? The first one is one I really worried about. If we, because AI is so good at empathy, perhaps we’ll start preferring the AI, the LLMs.
And we won’t, you know, go out in the world and actually meet real live flesh people. And I think that would be a net loss. Sycophancy, a major problem that computer scientists are focused on, which is a GRE word or an SAT word. It means being excessively flattering and agreeable. And LLMs are found to be that. And that could be problematic, especially in a therapeutic context. AIs could manipulate you. If they’re your friends, and we know that these AIs are software from a company and those companies might have political agendas. They could then suddenly manipulate their AI, and so manipulate you. Not great. And then finally, this is something that Paul Bloom, a friend and collaborator, has raised: is loneliness functional?
Perhaps if we dull the pain of loneliness, we are preventing people from taking action to remediate the loneliness, which is going out into the world, right? So I view this as a kind of a, I hope, an even-handed listing of pros and cons. Now, someone might say, and in fact, someone did say this to me, lots of words just to say, you’re a human hater, right? So an analysis like this, a pro-con analysis like this, could be responded to this way, and in fact, has been responded to this way. So what is going on here? What is going on with AI attitudes where a simple cost-benefit analysis could generate so much heat, so much hate, so much vitriol. And we believe, and we have data to examine this question, is that AI has become moralized.
So what do I mean by morals, moralized? What do I mean by something, you know, an attitude becoming moralized? Well, first thing to note is something that’s moralized is unlike other attitudes. It’s not a mere preference. So I have a preference. Preference, I like pineapple on my pizza and you might not like pineapple on my pizza, but you would just agree, okay, you know, Mickey likes something that’s gross to me. But you don’t think I’m wrong. At least you shouldn’t, because it’s just a matter of preference. Whereas moralized attitudes, there are rights and wrongs. You know, if I was to say that I think animal torture is fine, you would say I’m wrong because I’m doing something unethical, right? Moralized attitudes are prescriptive.
Everyone should abide by this rule. Everyone regardless of station should not be torturing animals, for example. It’s consequence insensitive. Regardless of how much good can come of this thing or how much bad, how little bad would come of it, you know, a moralized attitude is insensitive to that evidence. So again, no matter how much you tell me about some of the benefits of animal torture, and I can’t think of any right now, that will not change your mind for something that’s been, you know, moralized. It’s inherently bad. And then finally, violators of moral attitudes elicit intense emotions in others, disgust, outrage, and those emotions can lead to actual behaviors, including violence. And that’s what we think is going on here.
So we tried to examine this empirically. And when I say we, I really mean my student, Victoria Oldemburgo de Mello, who wrote this brilliant paper in my mind, which is currently under review at Nature Human Behaviour. And she asked a few questions. First, does the media discuss AI with a moral lens? Is the media using moral language to talk about the pros and cons of large language models? Is opposition to AI moral? There are certain questions we could ask people to determine this, and I’ll show you that in a minute. Is opposition to AI domain general? Meaning that once you’re opposed to AI writ large, each individual application you might be opposed to as well. And finally, are there costs to moralization, personal costs?
So to examine the first question about the way the media talks about AI. We examined headlines from 2018 to 2024 and note that AI, ChatGPT, I guess it was 3.5, came out in 2022. So right in the middle of, we’re right towards the end of this period. And we used like an online dictionary to determine how, and we had like nearly 70,000 headlines. And we examined the extent to which moral language is used in various headlines. So here’s an example of a moral headline. AI systems are making decisions that destroy people’s lives. That headline could be phrased in a different way. Could it displace jobs, for example, would be a non-moralized way of describing the same thing.
And what we found, as we looked at the timeline, is that moralization surged after DALL-E and ChatGPT were launched. No surprise. But when we compare the way news headlines, and this is in English language, by the way, when we compare how news headlines in the English language talk about AI, we see that they’re using moral language to do so. So our control would be interior design, where there’s very little moralization of interior design. And what you see is that it’s quite low. And you see that there’s something very high. We know that abortion is moralized. And indeed, it’s reflected in the way headlines speak of it. And what we see is that AI is just below climate change as a moralized topic.
It’s ahead of genetically modified organisms or COVID-19 or vaccines. So the media is discussing AI and large language models using moralized language. Here’s another example. The great AI art heist. Right? Copyright law is pretty complicated. And AI, especially large language models, are generative. And it’s not clear to what extent, you know, if a model is trained on a corpus that was purchased by the company, which is not always the case, by the way, but let’s assume it was, in what way is that a copyright infringement? It’s not clear to me. So to say art heists is already moralizing what I might argue is potentially a neutral thing.
Second, the second thing we did in our second series of studies is we got a representative sample of Americans, nearly a thousand. We actually got two samples of nearly a thousand samples each. I’m just going to show you one study. And what we did was we gave people examples of applications, various AI applications. So a chatbot you might see in a store, a website, AI used for creative art, AI in legal decision making, and AI companions, romantic AI. And what you see is that only a minority of Americans oppose AI. Despite the vitriol you see online, despite those sensational headlines, most Americans are okay, at least with individual applications. Only 30% oppose, although you see variability.
Most opposition for romantic AI, the least opposition for simple chatbots. What about moralization? Well, one way we can examine moralization is we can ask people a series of questions about their attitudes. So here’s one question we ask people. AI should be prohibited no matter how great the benefits and minor the risks from allowing it. This is something we’d call consequence insensitivity. But we ask people a series of questions. Here’s just one example. And what we find is that the majority of opposers moralize. So only 21% of the overall population from these two samples generally moralize. They’re a majority of the opposers.
What this means is if you sampled a random American, there’s a good chance they’ll be fine with AI, more or less, if you give them a specific application. But if you find someone who’s against it, there’s a good chance that they’re opposing it on moral grounds. They wouldn’t be willing to change their mind regardless of the benefits and risks of each of these applications. And then finally, when we examine structural equation modeling of the various forms, we look at the structure of AI opposition, there could be different kinds of opposition for each of the applications, or it could be that one factor explains them all.
And what we find is the latter, which suggests that, in fact, we might think we’ve got nuanced attitudes about this application versus that application, and you might think that way, but in the end, the person who moralizes one application moralizes all of them. And they come up with reasons after the fact for their moralization, but it doesn’t seem like people’s reasons are firmly held. The attitude comes first, it seems like, just like Jonathan Haidt describes in his infamous paper or famous paper. And then finally, we asked, is moralization costly? And by that, we mean personally costly. Would people forego benefits to the extent that they moralize?
So what we did is we recontacted those people from that study I just described three to 18 months later. And now we know the people who moralize AI versus the people who don’t. We then put them in a scenario where they read and grade an essay that we’ve composed. And we tell them that we had a group of expert graders grade these essays. And if they get within five percentage points of these expert graders, we will give them an incentive, a bonus pay. So they’re incentivized to be accurate. And then just as they’re about to go off, we tell them, by the way, if you’d like, you can grade this yourself, or you could use Grade AI, which is an LLM we created that is superior in our testing to human graders on average. So what would you like to do?
And what we find is that people will use Grade AI. They’ll use the LLM. But if people moralize it, way fewer of them will actually use it. It’s quite a big difference in terms of the use of this technology, dependent on whether you moralized three to 18 months earlier. So in other words, people forwent bonus, a cash bonus. And these are Prolific workers who are very incentivized by cash payment. They forwent the possibility of an extra reward because they didn’t want to use the AI. So it does have some personal costs. All right, so almost to the end here. And let’s go back to my favorite person, Becca Rothfeld.
She wrote that viral essay, but just a few days ago, I think it was just last week, she had a back and forth with the philosopher, Daniel Greco, who’s a professor of philosophy at Yale University, who wrote an article saying that AI should be taught how to do philosophy. And Becca Rothfeld argued philosophy is inherently a human activity. Anyway, she then writes this, whether you think it is permissible to train AIs will hinge on your ethical views. Do you think we have some deontological obligation to refrain from certain sorts of complicity in evil? And look, she’s loading that question. She’s already assuming that AI is evil. But she raises the word deontology, which she’s correct.
A deontological moral basis is simply based on certain things being inherently wrong. But proper deontology still requires some form of reflective scrutiny. You can’t just have moral intuitions that just hang there without being scrutinized and tested against other moral intuitions. And if you were to do that, you might come to different conclusions. So my question to Becca Rothfeld is, where’s the reflective scrutiny? Moral intuitions could be a starting point. I’m not against moral intuitions here, but they need to be held lightly and perhaps even open to revision. So what are the challenges AI poses to civic life? I don’t want to be here as an AI booster.
I think there are lots of opportunities, also lots of challenges, but I think there are perils to deontology. So challenges of AI: safety, dignity, meaning, employment, connection. These are real. Let’s study these. Let’s understand these. But there are benefits, too. Like, it can make our world more safe, like with autonomous cars, for example. It can improve important life-saving decision-making. It can improve productivity. It might even make our lives more comfortable. What I’m really against is not, I’m not for or against AI, but I am against, I believe there’s great perils in foreclosing on a cost-benefit analysis by prematurely moralizing. I think that’s what’s happened now.
It ignores evidence, it forgoes benefits, it politicizes decisions, and forecloses discussion, which is the biggest problem of all in my mind. So my advice and my kind of take-home here is stay empirical. Study it, debate it, raise objections, say how bad it is, but keep an open mind. Your views should not be immune to revision. With that, I thank you for your attention — and sorry for going a little bit over time.
