A barista’s manager yells at him, leaving him flustered and fearing he’ll be fired. A colleague might reassure him by pointing out that the boss was probably just having a bad day. But today, an AI chatbot might suggest an even better choice of words to comfort him.
Since 2016, psychologist Amit Goldenberg, an assistant professor at Harvard Business School, has been studying how cognitive reappraisal—a reframing technique that changes the emotional impact of a negative experience—could help employees overcome difficult challenges in workplace settings. These reappraisals are one of many tools that human therapists use to help clients who are navigating work and social situations. Recently, Goldenberg decided to test whether a large language model could help generate such reappraisals and developed a narrow study to investigate the idea. As it turns out, ChatGPT-4 (the version retired in 2025) is very good at this.
Both the human participants in the study and ChatGPT were prompted with short texts describing negative social situations, such as, “My friend said she’d buy me dinner for my birthday, then forgot. I feel unwanted.” Then they were asked to construct emotionally helpful cognitive reappraisals, such as, “She probably had a busy week. The offer alone shows that she cares about you.”
Each human participant was then asked to judge the effectiveness of the other reappraisals. ChatGPT’s reappraisals were described as more effective, empathic, and novel than all but 20 percent of the reappraisals generated by humans.
Why are AI’s reappraisals so compelling? Subsequent text analysis revealed that AI favors certain words while humans lean toward others. “Dedication,” “love,” “boundaries,” and “opportunity” all showed up frequently in AI’s reappraisals, whereas humans’ top words included “trying,” “new,” “feel,” and “maybe.” When the researchers dug a bit deeper, they noticed a trend: AI’s word choices sound more erudite than the average human’s. For example, the algorithm liked to substitute “maybe” with “perhaps.” Goldenberg’s conjecture is that humans want support from people who will be reliable and judge people who use sophisticated language as more likely to be reliable.
However, in a subsequent experiment, when Goldenberg and his colleagues told participants which reappraisals were generated by AI, ChatGPT’s advantage dropped. The researchers call this the “AI penalty,” and it indicates to Goldenberg that “what people are really after is other humans—human understanding, human validation, human empathy.”
This suggests that the best use of AI’s skill at generating cognitive reappraisals could be as an aid in limited situations: for example, helping volunteers providing peer support services, such as answering grief and substance abuse hotlines. When discussions become too dark, Goldenberg says, peer support volunteers sometimes freeze up and end the conversation because they don’t know what to say. AI, he suggests, could propose supportive language when these volunteers get overwhelmed.
But Goldenberg cautions against the idea that interactions between humans could be replaced, rather than supplemented, by interactions with machines. A 2024 study conducted by researchers at Stanford found that 24 percent of AI users in the United States turn to the technology with their mental health challenges—a startlingly large number, and a cause for concern given several documented instances when AI has encouraged a human interlocutor to commit suicide.
AI’s appeal likely stems from its availability, Goldenberg says: “It’s readily available and cheap and will always want to hear what you have to say.” But the quality of support users receive from AI will always be inherently limited, he adds. Because as flashy as its language may be, there’s no person behind that language. His own study’s results, he says, indicate that what people are truly looking for is “what’s standing behind that response, and whether that person is reliable and will be there for them. AI can’t come and meet you and give you a hug.”
