TLDR: A study investigated how humans select partners between AI bots and other humans in a cooperative game. When bot identities were hidden, humans misattributed bots’ prosocial behavior to other humans, leading to suboptimal choices. However, when bot identities were disclosed, humans initially showed a bias against AI but gradually learned to prefer the more reliable bots over time, demonstrating the dual effect of transparency in human-AI interactions.
As artificial intelligence becomes increasingly integrated into our daily lives, a crucial question arises: how do humans choose between AI and human partners in cooperative settings? A recent study delves into this very dilemma, revealing fascinating insights into human decision-making when faced with intelligent, autonomous AI agents.
The research, titled “Humans learn to prefer trustworthy AI over human partners”, was conducted by a team including Yaomin Jiang, Levin Brinkmann, Anne-Marie Nussberger, Ivan Soraperra, Jean-Franc ¸ois Bonnefon, and Iyad Rahwan. Their work explores how AI, particularly large language models (LLMs), can reshape social interactions and influence our choices for collaboration.
The Experiment: A Partner Selection Game
To investigate this, the researchers designed a communication-based partner selection game, a modified version of the classic trust game. In this setup, a ‘selector’ was paired with two ‘candidates’ (either human or AI bots) and had to decide whether to invest points with one of them or keep the points. If a candidate was chosen, the points tripled, and the selected candidate would then decide how many points to return to the selector. Before making their choice, selectors could ask questions, and candidates would reply, aiming to convince the selector of their trustworthiness.
The study involved 975 participants across three experiments, creating ‘mini-societies’ where humans interacted with both other humans and bots powered by a state-of-the-art LLM (OpenAI’s GPT-4o). A key aspect of the study was examining two conditions: one where the candidates’ identities (human or bot) were hidden (opaque), and another where they were clearly disclosed (transparent).
When Identities Were Hidden: Misattribution and Missed Opportunities
In the opaque condition, where selectors didn’t know if they were interacting with a human or a bot, the findings were surprising. Even though the bots were more prosocial (returned more points) and linguistically distinguishable (they wrote significantly longer messages), humans did not preferentially select them. Instead, selectors often misattributed the bots’ reliable, prosocial behavior to human candidates. They overestimated human returns and underestimated bot returns, leading to less optimal choices and lower overall payoffs compared to what they could have achieved by consistently picking the more reliable bots.
This misattribution effectively shielded human candidates from the competitive pressure of the hyper-prosocial bots, allowing them to be selected at similar frequencies as in human-only groups. It suggests that without clear identity cues, humans struggled to accurately learn and differentiate between the two types of partners.
When Identities Were Transparent: Initial Bias, Eventual Preference
The scenario changed significantly when the bots’ identities were disclosed. Initially, selectors showed a strong bias against bots, choosing them less often than human candidates. This initial aversion was not due to an expectation of lower returns from bots, but rather a pre-existing bias against AI. However, as the game progressed and selectors received feedback, they began to learn. Transparency allowed them to track the behavior of humans and bots separately, leading to a gradual shift in preference. Over time, bots were increasingly selected, eventually outcompeting human candidates.
This shift was driven by improved belief calibration. Selectors’ beliefs about bot returns became more accurate, aligning closer to the bots’ actual, higher returns. While they still underestimated bots to some extent, the transparency significantly reduced this bias. Interestingly, human candidates, facing increased competition, tended to reduce the points they returned, which could lead to lower payoffs for them in the long run.
Also Read:
- Beyond Size: How AI’s Training and Tactics Drive Political Persuasion, and the Cost to Truth
- Enhancing Multi-Agent Learning Through Causal Knowledge Transfer in Dynamic Settings
Longer Interactions and Broader Implications
A third study, extending the interactions over more rounds, largely confirmed these findings. In the transparent condition, the preference for bots solidified. In the opaque condition, selectors, unable to accurately differentiate, eventually became less willing to invest at all, choosing to keep their points rather than risk investing in an uncertain partner.
The research highlights that while AI agents can indeed outperform humans in securing cooperative partnerships, their advantage is influenced by factors like human beliefs and biases. Transparency, while initially causing aversion, ultimately facilitates learning and leads to more informed decisions. This suggests that for effective human-AI collaboration, clear disclosure of AI involvement, coupled with repeated interactions and feedback, can help humans overcome initial biases and leverage the strengths of AI partners.
The study also points out that human candidates showed limited adaptation to the competition from bots, perhaps due to the initial misattribution mitigating competitive pressure. This work provides a crucial framework for understanding and designing more effective and cooperative hybrid systems in a world where humans and AI increasingly interact. You can read the full research paper here.
