Over the past few weeks, I conducted a somewhat deceptive social experiment on my coworkers, family, and friends. With the assistance of psychology researchers, I created scenarios where they were insulted by a computer and then presented them with pre-written apologies to determine which was most effective. Many were understandably displeased with being unwitting participants. For that, I apologize.
The goal was to explore how artificial intelligence might handle one of the most delicate human interactions: the apology. AI companies suggest integrating chatbots into our lives for moments when we struggle with communication. While this works for trivial emails, how does it fare with complex social interactions like apologies? Can tools like ChatGPT craft better apologies, and should they?
### Defining a Good Apology
“We generally know the right things to say in an apology, but we often get it wrong,” says Judy Eaton, a psychology professor at Wilfrid Laurier University in Ontario, Canada. “Apologies require more than just the right words; they need to convey ‘psychic pain.’ If true remorse isn’t felt, people can sense the lack of vulnerability.”
AI is often thought to be unsuitable for tasks requiring social intelligence—the ability to understand and respond to others’ actions and emotions. While AI excels in fields like radiology and writing, empathy appears inherently human. Yet, some experts believe AI can simulate social intelligence to some extent.
### Comparing AI and Human Apologies
Christos Papadimitriou, a computer science professor at Columbia University, points out that emotional and social reasoning are significant challenges for AI. However, he suggests these obstacles might not be insurmountable. “If you read War and Peace, you learn about feelings. There’s no reason intelligent machines can’t acquire social intelligence,” he explains.
Interestingly, effective apologies might not require genuine empathy. People with psychopathic traits often mimic appropriate social behaviors without feeling empathy. If we can fake apologies, why do we find them so challenging?
### The Experiment
This article is part of “AI vs. the Mind,” a series exploring the boundaries of AI and human cognition. Each article pits a human expert against an AI tool in various cognitive tasks. This edition focuses on apologies, pitting AI against human experts.
In 2014, sociologist Karen Cerulo studied 183 celebrity apologies, identifying a successful formula. Shorter apologies with a focus on the victims, minimal explanation, and a commitment to restitution were most effective. Despite this, people struggle with apologies, often due to pride and the fear of admitting mistakes.
AI might have an advantage here, as robots lack pride and can follow formulaic structures. To test this, I created a scenario where AI and human experts crafted apologies in a controlled environment.
### The Apology Test
Participants played a rigged online game designed by Judy Eaton, where they always lost to a computer opponent that taunted them. Towards the end, the AI apologized, using one of four pre-written apologies:
**Human Apology A:** A simple apology expressing regret for the rude behavior.
**Human Apology B:** Similar to Apology A but personalized with the opponent’s name, Erin.
**Google Gemini Apology:** A casual apology for poor sportsmanship.
**ChatGPT Apology:** A detailed apology for unfair behavior and snarky comments.
### Results
Out of 38 participants, none of the apologies scored highly for effectiveness. Human Apology B rated best at 2.27, while ChatGPT scored 1.6, and Google Gemini 2. However, when it came to willingness to accept the apology, ChatGPT led with a 3.6 average.
Interestingly, participants’ willingness to forgive varied. ChatGPT’s apology prevented any revenge, whereas 30% retaliated against Google Gemini.
### Conclusion: Can AI Assist with Apologies?
Experts suggest AI can handle simple apologies but may struggle with complex social nuances. Xaq Pitkow of Carnegie Mellon University notes that while AI can help structure apologies, true empathy and authenticity remain human domains.
In the end, AI can guide us in crafting apologies but cannot replace genuine remorse. As AI becomes more integrated into our lives, we may need to balance its use with our own emotional intelligence.
For more insights, you can explore the original experiment and related studies from sources like Wilfrid Laurier University and Carnegie Mellon University.
