Motivation
The social simulator is a technology for studying human behavior and social interactions through digital models. Especially with the advent of Large Language Models (LLMs), it has become possible to approach more complex phenomena at relatively low cost. Here’s a brief summary of why social simulators are useful, what sets them apart from previous methods, and what they can be used for.
To give some conclusions in advance, ultimately, we hope to gain insights through social simulators like:
Generalized understanding of humans
The insights we’ve gained about humans are based on the existing Earth environment, historical developments, and reliance on limited materials.
Discover new behaviors in human societies
For example, introducing concepts like ” human rights” that can’t be directly observed, but have a significant impact on human societies and can successfully guide group behavior. We want to discover norms, cultures, and concepts that can improve human societies.
Better understanding on known social phenomena
There is believed to be an impact from unconscious biases in society, but we want to know how strong this impact is.
What does Social Simulator mean for psychology?
Psychology is a representative discipline for understanding the mechanisms within humans. Born about 150 years ago, it is a field that has produced many insights. However, it faces several structural constraints and difficulties.
Because humans are complex, grasping the full picture of a phenomenon requires a vast amount of data
For example, when trying to explain why some food tastes good, various causes can be identified. These include: “It was strongly flavored (food component),” “I was hungry (physical state),” “I saw a commercial,” “I have memories of eating it with my family (memory),” “It was a situation where it was expected to taste good (social environment),” “The temperature and humidity (external environment),” “Genetic and upbringing environments that prefer strong flavors.”
There could be this many factors just off the top of one’s head, and when these interact with each other, there is a need to investigate every possible combination. When comparing each of these factors, whether they exist or not, whether they are strong or weak, and examining their impact, a vast amount of research is necessary, which can exceed practical costs. Therefore, in practice, we focus on points of interest and control other conditions so they do not influence the results (through control and sophisticated statistical processing).
Controlling conditions and focusing on points for investigation is difficult.
Continuing from the above example, let’s say we want to focus on the impact of hunger on tastiness. In order to do this, we must be able to ignore the impact of other conditions, but this requires a great deal of effort and cost. For example, it is difficult enough to gather people who are eating in exactly the same situation and environment, who have never seen the commercial, and who are eating this food for the first time, with only their levels of hunger differing. Even doing all this, there might still be influences from their upbringing environments and internalized beliefs that we don’t know about.
The process by which phenomena occur is not understood.
We cannot directly see inside humans, so we attempt to grasp it through observable actions or responses to questionnaires, but we cannot know the exact moment someone thinks something tastes good or the timing when they recall a memory with their family.
There is distortion due to humans observing humans.
Both the researcher and the subject are human. The researcher naturally has motives such as wanting to prove something or be recognized by society, and the subject may want to be liked by the researcher or at least not to upset them. When these motives inconveniently combine, there is a risk of creating a flawed understanding of humans.
There are inherently difficult themes to study.
Experiments that place humans in negative environments or situations, such as those that are likely to induce depression or bullying, are not ethically allowed.
Social simulators potentially alleviate some of these constraints and difficulties.
The desired complexity can be designed, and it is possible to assume that influences of experiences and knowledge that would otherwise be noise are not taken into account.
Since conditions can be freely set, it is possible to simulate human behavior in environments and timelines that are difficult to set in reality.
All processes can be logged, making it possible to observe the processes by which thoughts and emotions occur.
Since it does not involve real humans, it is possible to tackle difficult themes and eliminate influences caused by human observation/participation.These eased constraints and difficulties, coupled with computational resources, make it possible to conduct experiments as much as needed, allowing for faster and cheaper acquisition of insights.
How is Social Simulator different from previous simulations?
Simulations have been a valid method for generating insights. Robert Axelrod’s research on prisoner’s dilemma is well-known, and models used to predict COVID-19 infections are still fresh in memory. However, these models deal with humans in a very simplified form. For instance, a typical model for COVID-19 infection prediction considers only three states of human condition: healthy, infected, quarantined (recovered).
However, there are limitations when dealing with more complex social phenomena. For example, when trying to understand human relationships, even interactions with the same person involve numerous factors such as work relationships, friendships, neighborhood relations, and relationships with mutual acquaintances. Just capturing one of these factors does not fully explain human relationships.
Moreover, to increase complexity, a vast amount of input is needed. Even if we talk about work relationships, it is difficult to define rules for the expected behaviors in relationships with supervisors, subordinates, colleagues, clients, and customers, and the differences in each industry.
This is where LLMs (Large Language Models) provide some relief. Because LLMs have already undergone extensive training, we only need to describe the parts we want to emphasize or control. The need to articulate and describe everything that is considered common sense is eliminated, opening the path for complex simulations.
Limitations and Anticipated Criticisms of LLM-Based Social Simulators
Of course, we cannot free-handedly enjoy the benefits mentioned above. Through future research, we must find ways to overcome these limitations and respond to criticisms.
It is difficult to erase or limit knowledge from an LLM.
This is a downside of having already undergone extensive training; it is difficult to simulate a state of not knowing what has already been learned. This leads to the unnaturalness of a human simulation that can mobilize extensive knowledge. For example, it might be difficult to accurately simulate a young child.
The training data of LLMs is likely biased.
The data is primarily from the internet, and thus it naturally reflects the bias of that data. It is expected that the data is overwhelmingly modern and in English. Additionally, the process of tuning is influenced by the companies or organizations that conduct the tuning. In practice, the speech and behavior of agents controlled by GPT often give an “American-like” impression. Because it is difficult to erase or limit prior knowledge, this problem becomes more significant.
LLMs currently only know what humans have articulated.
Humans have bodies that are influenced by the environment, which naturally affects thoughts and emotions, and indirectly influences actions without the need for words. However, LLMs do not have bodies and cannot currently represent physicality in simulations. They live entirely in a world of words and are limited to what humans have articulated thus far.
It is merely a simulation and does not fully represent humans.
This simulator could offer various benefits to psychology, but it faces challenges that psychology has not previously encountered. In psychology, the subject of research is humans themselves, and it is taken for granted that experimental results apply to humans. However, in simulations, the extent to which the simulation resembles humans is a new question that affects the generation of insights.
How far should humans be represented?
Among the limitations and anticipated criticisms of social simulators, the point that “it does not fully represent humans” is particularly fundamental, and we need to organize our thoughts early on.
After all, we do not fully understand real humans.
As mentioned, psychology has faced constraints in its research methods, and we have not been able to construct theories or models that fully explain human behavior in all situations. Therefore, whether it is a simulation or real humans, what we can study is limited to certain aspects of humans in certain situations.
If the trends of the group are similar, it’s good enough.
Up to this point, the emphasis has not been on simulating individual humans but on social simulators. Therefore, even if each individual does not resemble an actual human, as long as the overall behavior is similar, it serves the purpose.
In summary, social simulation is useful to the extent that it represents one aspect of humans in a limited situation as a trend of the group, and this is considered to be compatible with the limitations of our understanding of humans and the probabilistic behavior of LLMs.
Conversely, vaguely conducting simulations carries a high risk of not gaining insights, and the importance of well-designed experiments based on clear hypotheses is high (as is generally the case with research).
Patterns of utilization
Considering the possibilities and limitations mentioned above, we have typified the applications and listed corresponding applications.
Generating hypotheses in complex phenomena.
Taking advantage of the ability to conduct various parallel experiments with different situation settings
Applications
- The impact of office environments on employee communication.
- The influence of tenant placement and configuration on the shopping experience.
- The impact of team composition (combination and number of members) on labor productivity.
Computational science・information engineering backup for phenomena where effects could not be obtained with real humans
Recognizing the existence of certain psychological effects but updating knowledge about how strong they are.
Applications
- Research on factors causing mental health changes and conformist behavior.
- Studies on the impact of mood and emotions on perception, evaluation, and decision-making.
- The effect of unconscious biases on interpersonal cognition.
Experiments in extreme environments or large time scales.
Taking advantage of ethical and physical constraints being weak.
Applications
- The impact of long-term stays in space environments on group human relations.
- Observation of how cultures and norms change over long time scales.
In addition to the above, we are also challenging applications to automatic research simulations and entertainment such as narrative generation.
Want to know more?
To make our intentions clear, we have written in plain language, but this lacks accuracy, precision, and consideration for previous research. We are preparing more detailed content, research plans and experimental results articles worked out with experts, to address such concerns. Stay tuned!