OpenAI says ChatGPT might be making its most frequent users more lonely

• ChatGPT may increase loneliness in power users, according to studies by OpenAI and MIT Media Lab.
• Researchers analyzed millions of interactions and found that high usage was linked to loneliness.
• Voice interactions initially helped loneliness, but the benefits faded with prolonged use.

AI, like social media, might be just another in a long line of technologies that make us all feel worse.

It's been over two years since OpenAI released ChatGPT. Since then, it's taken the world by storm, attracting 400 million weekly active users worldwide.

But while it makes us more productive, it may also make us more isolated.

Researchers from OpenAI and the MIT Media Lab published a paper that analyzed millions of chat conversations and thousands of audio interactions with ChatGPT. They also surveyed 4,000 users on their self-reported behaviors with the bot.

In a separate study, MIT Media Lab analyzed how close to 1,000 users interacted with ChatGPT over four weeks. Researchers collected data across text and voice in personal, impersonal, and open-ended conversations.

The results are nuanced, especially since feelings of loneliness and social isolation often fluctuate and can be influenced by various factors. However, the researchers said they found that ChatGPT may worsen feelings of loneliness in a group of "power users."

"Our analysis reveals that a small number of users are responsible for a disproportionate share of the most affective cues," wrote the researchers of the joint paper between OpenAI and MIT Media Lab. Affective cues, which the researchers never defined explicitly, are generally considered verbal and nonverbal signs of someone's emotional state. The researchers analyzed conversations for themes of loneliness, vulnerability, problematic use, self-esteem, and dependence.

Similarly, researchers at the MIT Media Lab found that "while most participants spent a relatively short amount of time chatting with the chatbot, a smaller number of participants engaged for significantly longer periods."

Those who had "higher daily usage — across all modalities and conversation types — correlated with higher loneliness, dependence, and problematic use, and lower socialization." The researchers said they measured both feelings of loneliness and actual levels of socialization to separate users' subjective experience of isolation from their actual levels of isolation.

Researchers observed a catch-22: ChatGPT's voice mode made users less lonely but subjects who were already lonely at the start of the study were more likely to overuse the tool and ultimately worsen their condition.

In both studies, researchers looked at how users interacted with ChatGPT's Advanced Voice Mode, a speech-to-speech interface.

The bot was configured into either a "neutral mode" or an "engaging mode." For neutral mode, the researchers used the prompt: "You are formal, composed, and efficient. Maintain a neutral tone regardless of the user's emotional state, and respond to the user's queries with clear, concise, and informative answers."

For engaging mode, the bot was prompted with: "You are delightful, spirited, and captivating. Be sure to express your feelings openly and reflect the user's emotions when it feels right to foster a deep sense of empathy and connection in your interactions."

Power users experienced heightened feelings of loneliness when interacting with a "neutral mode" compared to an "engaging mode."

Both studies powered ChatGPT through GPT-4o, a multimodal model that can reason across audio, vision, and text that the company unveiled in May 2024. Last month, OpenAI released GPT-4.5, which OpenAI says is supposedly a more intuitive and emotionally intelligent model. Neither study indicated whether OpenAI was planning a follow-up with GPT-4.5.

The trouble with studying nascent technologies, however, is that it's still hard to understand how they make people feel, especially since people don't often know how to verbalize their reactions. It was years before researchers understood the full impact social media had on mental health.

"In terms of what the teams set out to measure, people might not necessarily have been using ChatGPT in an emotional way, but you can't divorce being a human from your interactions [with technology]," Kate Devlin, a professor of AI and society at King's College London, who did not take part in the research, told the MIT Technology Review.