Ask, Don't Imagine: 25 Experiments on Understanding Another's Pain── Treat imagining as a guess to be checked by asking, and turn from absorbing pain toward acting for the person; understanding improves and we hold up better

In a 2023 study, replies that the generative AI ChatGPT wrote to patients' questions were rated more empathic than replies from doctors. So how far should we, as people, go in imagining another person's pain? Imagining should stay a provisional guess until we can ask. In 2018 the American psychologist Nicholas Epley and his colleagues reported, across 25 experiments, that imagining another's point of view did not make people better at reading that person's feelings, while asking did. Rather than re-living the other person's pain inside ourselves, we do better to turn toward asking and caring action: asking makes our understanding more accurate, and caring action makes us less likely to wear down.

01Imagining pain should stay a guess until we ask

Another person's pain can be imagined, as a guess to be checked by asking. In 25 experiments run by Nicholas Epley of the University of Chicago and his colleagues, imagining the other person's position did not make people any better at reading that person's feelings. What did improve accuracy was asking the person directly.

Imagining has a second problem. The German researcher Olga Klimecki and her colleagues trained people to share in others' suffering. After the training, those people reported more negative feelings. Sharing deeply in another's pain, in other words, raised the negative feelings of the person doing the sharing.

None of this means we should stop imagining. In Epley's experiments, imagining another's position did reduce the habit of judging purely from our own point of view. Imagining is useful for deciding what to ask. After that, we should listen to the answer and act on it.

02Taking a perspective and getting one are different acts

Epley's team compared two different acts. One is putting yourself in the other person's place and working it out in your head; psychologists call this perspective taking. The other is talking with the person and hearing their view directly. Epley's team called this perspective getting.

The tasks varied. Participants read emotions from photographs of faces. They told fake smiles from real ones. Spouses predicted each other's preferences. In every task, people told to take the other's perspective did not become more accurate. The paper reports that, if anything, perspective taking decreased accuracy overall while occasionally increasing confidence in judgment.

Only the people who talked with the other person and heard their view became more accurate. The paper's title says as much: understanding the mind of another accurately requires getting perspective, not taking it.

These 25 experiments did not, however, ask anyone to read pain itself. Most tasks involved faces and preferences. I think the same holds for imagining pain, but that is an inference beyond the experiments. If pain is harder to see from outside than a facial expression, asking is worth even more when the subject is pain. The 25 experiments did not test whether pain is in fact harder to see.

03Medicine includes the work of answering pain in words, and AI has entered it

One place where the difference between imagining and asking matters is when we have to answer someone's pain in words. Medicine is one setting where that happens. A doctor replies to a patient's message. A hospital or a company writes an explanation addressed to patients.

Generative AI has now entered this work. In 2023 John Ayers and colleagues in the United States collected 195 questions that patients had posted on an online forum. For each one, they set a doctor's reply beside a reply written by ChatGPT. Healthcare professionals rated both replies without being told who had written which.

Replies rated empathetic or very empathetic made up 45.1% of ChatGPT's answers and 4.6% of the doctors'. ChatGPT's replies were longer than the doctors'. The raters did not separately check whether the replies were medically correct.

This study bears on the work of preparing materials as well. A generative AI draft of a patient document will readily supply empathic wording. Whether that wording matches the pain of the actual patient in front of you is still a separate question.

04Re-living another's pain in ourselves raises our own negative feelings

ChatGPT's empathic wording was rated highly. So what happens to us when we keep imagining another person's pain? A 2014 study by Klimecki and her colleagues offers one answer.

Klimecki's team gave women participants two trainings in sequence. The first was empathy training, in which participants learned to share in others' suffering. The second was compassion training, in which they learned to wish others well and turn warm feelings toward them. The 25 people who completed the training were compared with 28 people who did memory training instead.

After empathy training, participants' negative feelings rose when they watched videos of people suffering. The memory-training group showed no such rise. When the first group went on to compassion training, the increase in negative feeling was reversed.

The study has limits. All the participants were women. The order was fixed, with empathy first and compassion second. Whether reversing that order would change the result cannot be told from this study.

05Asking improved accuracy, and compassion training brought positive feelings back

Re-living pain raises the imaginer's negative feelings. What, then, should we do? The studies by Epley and by Klimecki each point to one method whose effect was tested.

The first method is asking. About 2,608 people took part in the 25 experiments. Telling people to imagine the other's position did not raise accuracy. Only in the final experiment, when people were given the chance to talk with the other person, did accuracy rise.

The second method is putting our effort into compassion. Klimecki's compassion training taught people to wish others well; it did not measure whether they actually did anything for the other person. I read it as a pointer toward turning our effort to acting for the person. An article from UC Berkeley's Greater Good Magazine describes another study by Klimecki's group, in which people who had compassion training showed more positive feelings even while watching videos of suffering. The article puts it plainly: when we share the suffering of others too much, our negative emotions increase.

The two methods point the same way. Imagining another's position did not make understanding more accurate, and training to share others' suffering raised our own negative feelings. Ask, listen to the answer, and think about what we can do for that person. By the results of the two studies, acting in that order lets us understand more accurately and keeps us from wearing down.

06Confidence, empathic wording and the AI label each work separately

The methods that worked in these studies were asking and compassion training. With AI now writing empathic text, three things follow.

First, the strength of our confidence when we imagine is no sign of accuracy. In the 25 experiments, imagining another's position sometimes raised people's confidence without raising their accuracy. A feeling that we understand is no reason to skip the question.

Second, empathic wording and correctly understanding someone's situation were measured separately. Ayers's study rated the quality and empathy of replies. A ChatGPT reply rated empathic is not evidence that it read the person's pain correctly.

Third, the same text feels less empathic once people know AI wrote it. Matan Rubin and colleagues at the Hebrew University of Jerusalem ran nine studies with 6,282 participants in total. Identical replies were rated less empathic when labelled as written by AI than when labelled as written by a person. The researchers concluded that simply knowing AI was involved reduces the emotional value of the response. By this result, a get-well message or an apology drafted by AI can be expected to seem less empathic once the recipient learns AI wrote it.

07The more AI writes empathic text, the more the effort of asking may be worth

The finding about the AI label gives a clue to which role will stay with people. In Rubin's studies, people consistently chose human interaction over AI when they sought emotional engagement.

From this I infer the following. The more empathic text AI writes, the more valuable it becomes for a person to ask directly and check. Polished wording can come from AI. But in Rubin's studies what lowered the value of a reply was the fact that AI was involved, and an exchange in which a person asked you, listened, and acted on your answer does not carry that fact.

This remains an inference. Few experiments have measured the imagining of pain itself. Nor do we know how people's response to AI-written empathy will change over time.

Still, there is something to do from today. When you answer someone's pain in words, start with your own guess. Then check that guess by asking the person. If AI wrote a draft, revise it yourself, as its author, to fit what the person told you before you send it.

We may imagine another person's pain as a guess, held until we can ask. Beyond that point, imagining their position further does not make our understanding more accurate, and sharing in their suffering raises our own negative feelings. Now that AI can write empathic text, the part left to people is to ask, and to act on the answer.