ONE HUNDRED days into the war in Gaza, I was finding it increasingly difficult to read the news. My husband told me it might be time to talk to a therapist. Instead, on a cold winter morning, after having fought back tears reading yet another story of human tragedy, I turned to artificial intelligence.
“I’m feeling pretty bummed out about the state of the world,” I typed into ChatGPT. “It’s completely understandable to feel overwhelmed,” it responded, before offering a list of pragmatic advice: limit media exposure, focus on the positive and practise self-care.
Advertisement
I closed the chat. While I was sure I could benefit from doing all of these things, at that moment, I didn’t feel much better.
It might seem strange that AI can even attempt to offer this kind of assistance. But millions of people are already turning to ChatGPT and specialist therapy chatbots, which offer convenient and inexpensive mental health support. Even doctors are purportedly using AI to help craft more empathetic notes to patients.
Some experts say this is a boon. After all, AI, unhindered by embarrassment and burnout, might be able to express empathy more openly and tirelessly than humans. “We praise empathetic AI,” one group of psychology researchers .
But others aren’t so sure. Many question the idea that an AI could ever be capable of empathy, and worry about the consequences of people seeking emotional support from machines that can only pretend to care. Some even wonder if the rise of so-called empathetic AI might change the way we conceive of empathy and interact with one another.
It is fair to say that empathy is one of our species’ defining traits, evolving as it did in lockstep with social interaction. And yet while we all know instinctively what empathy is, it is actually quite a slippery concept.
One , which looked at 52 studies published between 1980 and 2019, found that researchers frequently declared the concept had no standard definition. Even so, almost all frameworks of empathy, whether they came from philosophy, neuroscience, psychology or medicine, shared certain dimensions, says psychologist at Stockholm University in Sweden, a co-author of the paper.
The researchers found that the empathiser must first be able to discern how the other person is feeling. They must also be affected by those emotions, feel them to some degree themselves, and differentiate between themselves and the other person, grasping that the other person’s feelings aren’t their own while still being able to imagine their experience.
Large language models and empathy
On the first point, in recent years, AI-powered chatbots have made strides in their ability to read human emotions. Most chatbots are powered by large language models (LLMs) that work by predicting which words are most likely to appear together based on training data. In this way, LLMs like ChatGPT can seemingly identify our feelings and . “The idea that [LLMs] could learn to express those words is not at all surprising to me,” says bioethicist at the University of California, Berkeley. “There’s no magic in it.”
But when it comes to the other criteria, AI still misses the mark in many ways. Empathy is , with continued cues and feedback helping to hone the empathiser’s response. It also requires some degree of intuitive awareness – no matter how fleeting – of an individual and their situation.
Consider someone who cries while telling a doctor she is pregnant. If we know her history of trying for years to conceive, we can imagine that her tears mean something different than, say, if she didn’t want to have kids.
Current AIs are incapable of that kind of nuanced understanding. More to the point, says Halpern, is that they are incapable of feeling for the person they are interacting with. “We have no reason to think AI can experience empathy,” she says. “It can produce a product – the language that mimics the actual empathy humans have. But it does not have empathy.”
All of which helps to explain conversations like the ones I had with ChatGPT and, later, with , one of several commercial chatbots that now offer mental health support – and one that, according to its creators, has been used by more than 100,000 people. When I told it how I was feeling about current events, it asked: “How does thinking about these situations affect your daily life and emotions?” “I’m finding it hard to cope,” I responded. “I’m sorry to hear that,” it said. “Are there any specific behaviours or thoughts that you’ve noticed?”
Every time I answered, Elomia immediately generated another question. The prompts gave me an opportunity to reflect. But what I really wanted was for someone to say, “Ugh, I know exactly how you feel,” share their own struggles or give me a hug – not pepper me with questions. In other words, I was looking for a response that was genuinely empathetic.
The big question, then, is whether that is even possible. Some believe that AIs may, in fact, become capable of experiencing and sharing human emotions. One approach is to continue scaling up LLMs with ever vaster and more diverse datasets. Virtual agents build on the language-processing abilities of chatbots by also “reading” our facial expressions and vocal tone. Most complex of all are robots that analyse these features, plus body language and hand gestures. Rudimentary versions of these emotion-detecting bots already exist, but more input can also create confusion in them. “The more modalities we add to the system, the more and more difficult it becomes,” says at , a German start-up studying human-AI interaction.
Making machines that are more adept at reading emotions is still unlikely to create genuine empathy, though, says psychologist at the University of Toronto in Canada, one of the researchers who wrote the paper praising “empathetic AI”. “You need to have emotions to experience empathy,” he says.
Emotions and algorithms
Some computer scientists argue that certain types of algorithm already experience primitive emotions – although it is a controversial position. These algorithms, called reinforcement learning agents, are rewarded for taking certain actions over others. The expectation of receiving a reward drives simple human survival instincts, such as seeking out food and water, too. We also expect and receive social rewards from each other that correspond to an array of human emotions, says Yoshua Bengio at the University of Montreal in Canada. The richness of our emotional palette is a result of the richness of our social interactions, he says. So the idea is that if you can capture similar social complexity in AI, then actual emotions – and perhaps empathy – could emerge.
Another approach hinges on mirror neurons: the brain cells thought to fire when we see someone performing an action or expressing an emotion, activating the same parts of our brain that would light up if we were to do or feel the same thing ourselves. Computer scientists have already built , arguing that these are the key to AI empathy.
But over the past decade, some of the research on mirror neurons has been discredited and, what’s more, we have learned that the brain areas where the cells reside .
Ultimately, you can’t really know what sadness is unless you have felt sad. The evolution of empathy is intertwined with social interactions and a recognition that other minds exist, too. Empathy emerges as a holistic whole, says Inzlicht. “You probably need consciousness to experience empathy.” That is far from a given for AI, now or in the future.
But what if AI doesn’t need genuine empathy to be useful? One online experiment from 2021 found that talking to a chatbot perceived as being caring in people – albeit not as much as talking to their human partner. And in January, another chatbot, designed by Google to conduct medical interviews, was by 20 people who were trained to impersonate patients – all while delivering more accurate diagnoses. Unlike humans, “an AI doesn’t get stressed out or burned out. It can work 24 hours without getting tired,” says Håkansson. However, the participants in the Google study didn’t know whether they were speaking to a chatbot or a doctor. Empathy is a two-way process, so its effectiveness also depends on how we perceive and emotionally connect to the other person – or machine. Research shows, for example, that when people know they are interacting with an AI, they .
“This means AI has two bad alternatives,” says Håkansson. Either the person doesn’t know the AI is an AI and thinks it is human, which might be effective but unethical, or they know it is an AI and this makes it ineffective.
Then again, Inzlicht has just wrapped up as-yet-unpublished research that indicates a preference for AI empathy – even when users know it is AI. In his study, human participants and ChatGPT were given descriptions of different scenarios and asked to write short, compassionate answers. When other participants rated the various responses, they scored the AI responses as highest for empathy.
In the next part of the study, Inzlicht and his team labelled which responses were from AI and which were from humans. “What we found was that people still preferred AI,” he says. Surprisingly, the labelled AI responses were even favoured over a subset of responses by Crisis Line Workers, who are expert deliverers of empathy. (Inzlicht’s work by John W. Ayers at the University of California, San Diego, and his collaborators.)
AI empathy
This may be explained by the human tendency to project feelings onto objects – whether that is children playing with Tamagotchis or people conversing with chatbots like Replika that let you create an “AI friend”. In conversational AI, research has found this by priming people to think that the AI really cares about them. “We begin to feel that it has feelings for us beyond what it is expressing or what it’s capable of,” says , director of the Massachusetts Institute of Technology’s Initiative on Technology and Self. “This isn’t a quality of the machines. This is a vulnerability of people.”
And it is a weakness that many worry could be exploited – particularly given the question of what it means for an AI to be able to discern someone’s emotions, but not to actually resonate with them or even care about them in a genuine way.
“Say an AI detects that you’re heartbroken because your boyfriend has just left you,” says at the Carnegie Council for Ethics in International Affairs in New York. “It could seemingly commiserate with you and then slowly turn the conversation to you buying some product. That’s just a very mild example. Even though the machine cannot feel, you are, in effect, being manipulated through your emotions.” That, say Wallach and other researchers, wouldn’t make AI empathetic. It .
Concerns over the danger of machines that can “read” us but don’t care about us are more than theoretical. In March 2023, a Belgian man after six weeks of discussions with an AI chatbot. Media outlets reported that he had been sharing his fears about the climate crisis. The chatbot seemed to feed his worries and to express its own emotions – including encouraging him to kill himself so that they would “live together in paradise”. Pretending at empathy to too great a degree without the common-sense guard rails that a human is likely to offer can, it appears, be lethal.
The possibility that we will increasingly turn to, and perhaps even come to favour, AI empathy over the human variety also concerns Turkle for a profound reason: it could change the nature of empathy itself.
As an analogy, she uses the Turing test, which was designed as a purely behavioural assessment for a computer’s “intelligence”. But its definition of intelligence – whether the computer could converse like a human – became a stand-in for our understanding of intelligence in general, even though human intelligence is much broader. Turkle suggests that we could start redefining empathy in a similar way – expecting humans to express empathy that is tireless, for example, but also less genuine or connected.
However, Inzlicht counters that we shouldn’t be swayed by the critiques so much that we don’t reap the potential benefits of empathetic-seeming AIs working alongside humans – whether as a quick assist for overstretched health practitioners or to relieve loneliness among older people, as the voice-operated AI “care companion” ElliQ is designed to do. Other researchers, including Turkle, argue that even this hybrid human-AI approach goes too far, as it will reduce key opportunities to practise and improve actual human empathy. We may end up turning to machines, rather than looking for ways to foster genuine connection, she says.
“It’s a sad moment when we are actually thinking that being lonely can be fixed by a back-and-forth interaction with something that has no idea about what it is to be born, about what it is to die,” says Turkle.
At the end of the day, despite talking to multiple chatbots online and getting more advice and reflective prompts than I knew what to do with, I found I still felt sad about the state of the world. I did what I knew I had to do all along: I picked up my phone and called a friend.
Amanda Ruggeri is an award-winning journalist and editor based in Switzerland
91av audio
You can now listen to many articles – look for the headphones icon in our app
