We Designed a Game to Test AI's Apology Skills
I need to apologize. Over the past few weeks, I conducted a social experiment without the knowledge of my co-workers, family, and friends. Collaborating with psychology researchers, I created scenarios where participants would be insulted by a computer and then exposed them to various pre-written apologies to determine which was most effective. Understandably, many of my subjects were not pleased about being used as guinea pigs. It was deceptive, and for that, I sincerely apologize.
The purpose of this project was to explore how artificial intelligence might influence one of the most sensitive human interactions: the act of apologizing. Companies that develop AI chatbots suggest using them when we are unsure of what to say. This approach might work for trivial emails, but what about more nuanced human interactions? Can tools like ChatGPT craft better apologies? Should they even be used for such purposes?
What Constitutes a Good Apology?
Judy Eaton, a psychology professor at Wilfrid Laurier University in Waterloo, Ontario, Canada, explains that effective apologies aren’t just about the right words—they must convey genuine remorse. “Apologies should reflect ‘psychic pain’,” says Eaton. If the apology lacks true vulnerability, it comes off as insincere.
A prevalent belief about AI is that it struggles with tasks requiring social intelligence—understanding and responding to human emotions. While AI can assist in various fields, including healthcare and creative writing, empathy-related tasks feel inherently human. Nevertheless, some experts believe AI could develop the capability to mimic social intelligence to a degree.
Christos Papadimitriou, a computer science professor at Columbia University, acknowledges that AI currently lacks the complexity needed for personal interactions. However, he argues that AI might eventually acquire social intelligence. Even now, it can mimic human behaviors, much like individuals with psychopathic traits who simulate empathy without genuinely feeling it.
The Challenge of Apologizing
If the elements of a good apology are well-known, why do people struggle with it? Karen Cerulo, a sociologist at Rutgers University, studied 183 celebrity apologies and identified a formula for effectiveness. Key components include brevity, prioritizing the victim’s feelings, avoiding excessive justification, and ending with restitution. Despite its simplicity, many fail to execute it correctly, often due to pride or fear of consequences.
AI, devoid of pride, might excel in this area. Apologies, being somewhat formulaic, are well within AI’s capabilities. To explore this, we devised a game where participants were repeatedly insulted and had money stolen by an AI opponent. The AI would then apologize using one of four scripted apologies.
The Apology Game
In collaboration with Judy Eaton, we adapted an online game designed to study apologies. Although the sample size was small and non-random, the experiment provided interesting insights. Participants competed against an AI in a rigged game where the AI consistently won and chose to steal money while sending taunting messages. Eventually, the AI issued an apology.
We used four different apologies:
- Human Apology A: A generic apology expressing regret and a desire to act more friendly.
- Human Apology B: A personalized apology from “Erin,” expressing regret and acknowledging the hurt caused.
- Google Gemini Apology: A casual apology for poor sportsmanship.
- ChatGPT Apology: A comprehensive apology expressing regret for unfair behavior and hoping to move past it.
Results and Reactions
Despite the setup, many participants suspected the game was rigged. However, they still reacted emotionally to losing. The effectiveness of the apologies varied, with Human Apology B being rated highest overall, while ChatGPT’s apology led to the least desire for revenge.
Implications for AI in Apologizing
The experiment showed that AI can craft passable apologies in controlled scenarios. However, more complex situations might challenge AI’s capabilities. Xaq Pitkow from Carnegie Mellon University suggests that while AI may struggle with intricate emotional contexts, it can still assist in formulating apologies.
Using AI to draft apologies isn’t inherently bad. It can help people express regret more effectively. As AI becomes more integrated into our lives, the ethical implications of using it for sensitive interactions will need careful consideration. Ultimately, effort and authenticity remain crucial in making a meaningful apology.
