In an international collaboration, scientists of the University of Tübingen have uncovered how people use social information to make decisions, even when the goals or preferences of others differ from their own. This research, now published in the renowned scientific journal Proceedings of the National Academy of Sciences (PNAS), was led by University of Tübingen PhD candidate Alexandra Witt and her advisor Dr. Charley Wu, along with colleagues from the University of Konstanz, RIKEN (Japan), and the University of St Andrews (United Kingdom). The results not only allow the researchers to better understand a crucial aspect of human learning, but also open up new paths to incorporate similar principles into artificial intelligence (AI).
Imagine you are visiting a new city for the first time, and dinner time rolls around. How do you choose a restaurant? You could just check reviews and go to the highest-rated option. But how can you be sure that the reviewers share your food preferences, your spice tolerance, or your budget? And how do humans in general manage to learn from others when preferences can vary quite substantially between individuals?
Until now, much research on how people learn from other people has focused on settings where everyone has identical goals and preferences. But in the real world, that's rarely the case. Whether we’re choosing restaurants, planning holidays, or making career moves, the individuals we rely on for advice often have their own, unique tastes and circumstances. This new study closes this gap by investigating how humans use social information to make decisions when the preferences of those around them aren’t a perfect match.
Socially correlated choice task
To study this phenomenon, the researchers created a new socially correlated choice task where goals are similar, but not identical, across participants. In an online experiment that resembled a computer game, groups of four participants were asked to collect as many samples of different salts from alien planets as possible, while they could see the progress of everyone else on their research team. Participants were told that the process generating these salts was similar, so that regions with high salt concentrations tended to overlap, which meant that incorporating information about the other salts could help them find spots with high concentrations of their own salt. This mimics how preferences in real life can be similar across people (e.g., have a delicious dinner), but with certain individual differences (e.g., how spicy the dinner should be).
Results and Conclusion
The results show that humans use social information to guide their decisions, but “with a grain of salt” — in the experiment, they treated social information as noisier, and thus less reliable, than information they collected themselves, following the predictions of the Social Generalization (SG) model. This novel model introduced in the paper outperforms a number of other models from previous theories in predicting behaviour. “Unlike models from the previous literature, our SG model assumes that social information should be integrated similarly to individual information, rather than blindly copied” explains lead author and PhD student Alexandra Witt.
Consider how you yourself might use online reviews — while it can be helpful to know that others liked a product, you cannot be sure that they have the same standards as you. It is necessary to take your own preferences into consideration. However, having the product ratings is still preferable to having to try out every product by yourself. This is precisely what the SG model captures—humans treat social information as a guideline, not a rulebook.
Additionally, the researchers found that humans used social information as an exploration tool. Individual exploration can be costly, both cognitively and in terms of risk. When social information was available, participants relied on it to guide their choices, saving themselves the costly individual exploration process.
So why does this matter? “Although recent advances have demonstrated the power of Artificial Intelligence, it still struggles to learn socially in a similar capacity as humans” says senior author Charley Wu, who leads the Human and Machine Cognition Lab and is a Core Faculty member of the Tübingen AI Center at the University of Tübingen. “Indeed, it is our capacity for social and cultural learning that have truly made humans stand out from other animals. A better understanding of this ability could let us incorporate similar principles into AI, such as in virtual assistants or recommendations algorithms”. Ultimately, social learning is one of humanity’s most powerful tools, and this research brings us closer to understanding this impressive ability.
Publication:
Witt, A., Toyokawa, W., Lala, K. N., Gaissmaier, W., & Wu, C. M. (2024). Humans flexibly integrate social information despite interindividual differences in reward. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2404928121
Open access preprint: https://osf.io/preprints/psyarxiv/e4g3q
Contact:
Dr. Charley M. Wu
Independent Group Leader of the “Human and Machine Cognition Lab”
University of Tübingen
charley.wu@uni-tuebingen․de