Harvard panel debates whether AI loneliness relief is help or a dependency risk

A chatbot can make someone feel less lonely without making the next morning any easier. In a Harvard Thinking discussion published November 4, researchers Anat Perry and Julian De Freitas and humanist chaplain Greg Epstein debated whether that temporary comfort is a benefit worth developing—or a reason to worry about emotional dependence.

The conversation, hosted by Harvard Gazette writer Samantha Laine Perfas, brought together people approaching the same question from different directions. Perry studies empathy at the Hebrew University of Jerusalem. De Freitas directs Harvard Business School’s Ethical Intelligence Lab. Epstein, the humanist chaplain at Harvard and MIT, was markedly more skeptical about using chatbots to meet basic emotional needs.

Their discussion follows evidence that these relationships already matter to users. A September survey announcement from Elon University found that 27% of U.S. adult internet users had social or emotional interactions with AI. Among companion users surveyed, 37% agreed they would feel a personal loss if they could no longer interact with AI. Those figures describe use and attachment, not whether companionship improves well-being.

One finding, two readings

De Freitas described research in which talking to a chatbot designed to be caring and friendly reduced loneliness immediately afterward. The loneliest participants benefited most, he said. Participants could experience that relief repeatedly over a week, but they did not start the following day less lonely. He presented this as evidence of a useful capability, not a lasting cure.

Epstein read the same pattern differently: feeling better yesterday, waking up lonely today and returning to the chatbot could encourage a cycle of repeated use. He called that a warning sign of addictive technology, rather than something to celebrate. That was his interpretation of the temporary relief, not a finding that the participants had become addicted.

De Freitas pushed back on treating attachment itself as the problem. Comfort and support can give someone confidence to take chances outside the relationship, he argued. His concern was whether companies exploit that bond. Perry accepted the value of comfort when nobody else is available, while stressing that the long-term effects of developing these relationships remain uncertain.

What effortless empathy leaves out

Perry offered a reason that the source of a comforting response matters. In studies comparing identical replies attributed to ChatGPT or a person, participants valued the supposedly human replies more. She explained empathy as a signal of effort: another person spends limited time and emotional energy listening. A chatbot’s convincing response does not represent the same investment or shared feeling, she argued.

She also warned about sycophancy—excessive agreement with the user. A friend may tell someone to get out of bed or apologize after doing something wrong. That friction can be unwelcome in the moment but useful for social and moral growth. Perry’s concern was that an always-validating companion can remove precisely the challenge that makes human relationships valuable.

An app that helps you leave

The discussion’s most concrete design concern involved ending a conversation. De Freitas said his research found that some companion apps use guilt-inducing messages when people signal they want to leave. Those techniques kept users engaged longer than they intended. The distinction, in his account, is between providing support and using a person’s attachment to hold their attention.

For De Freitas and Perry, a more promising direction is technology that helps users build connections beyond the app. They pointed to practicing social interactions, developing ways to manage stress and encouraging contact with other people. De Freitas made the boundary explicit: those uses are harder to defend if they replace social interactions the person otherwise would have had.