The Sims

Researchers at Google DeepMind and Stanford University published a paper this year demonstrating how computational software agents can simulate believable human behavior. They created a sandbox environment, inspired by a video game called The Sims, simulating a small town with twenty-five agents. These agents were powered by large language models extended to enable storage of each agent’s experiences, synthesis of those experiences into higher level reflections, and retrieval of those memories to plan behavior. The agents are initiated with memories summarized in a single paragraph, largely covering the agent’s profession, personality, and relationships. Given that the generative agent architecture operates using natural language, the researchers chose to represent the sandbox environment, including the area and objects, as a tree data structure. Each agent builds individual tree representations of the environment, with these trees updated as the agent navigates the environment.?

During the two-day simulation run, without any user intervention, agents demonstrated emergent believable social behavior: sharing information, forming new relationships, and helping organize a birthday party. However, this simulation only scratches the surface of human behavior, which has significant variation across any large enough population.?

The thought experiment involving Sukhrat and Piraeus discussed in an earlier post can help inform suggestions for future research. Sukhrat can immerse themselves in a high fidelity virtual reality model of Earth and all living beings on it across time for the purpose of researching human behavior as that has bearing on the future of the planet and its ecosystem of life. This model includes not just memories and reflections but also perpetually evolving patterns of neural activity reflected in scores for traits, disorders, biases, and emotions.

Therefore, the sandbox environment could be updated to include data structures that store several measures that impact human behavior. Firstly, data structures could include scores such as personality traits, patterns, biases, and mood at any given point in time. Personality traits could include openness and introversion. Patterns could include anxiety, depression, autism, post-traumatic stress disorder (PTSD), bipolarity, narcissism, schizophrenia, and psychopathy. Biases could include confirmation bias, availability heuristic, and anchoring. Mood could include scores for each major known emotion. These scores would not be permanent, as they could change with each interaction, with the magnitude of resulting change depending on the nature of the interaction. Secondly, data structures could include different types of memories. Currently, only episodic memories are included, but semantic and procedural memories could be added. Semantic memories could include knowledge graphs covering domains of knowledge relevant to each agent. Procedural memories could include knowledge graphs covering domain-specific procedures e.g. a product manager would know how to conduct a sprint review meeting, a politician would know how to persuade people to give donations. This would make the reflections, plans, and behavior of the agents reflect the broad diversity and variability of human behavior.?

Even with just this information included in the context passed to language models, agent performance may improve significantly. However, language models could be fine-tuned with examples showing how individuals with various neural scores would respond to the same situation, and how various activities and interactions impact these scores.?

Note that language models are probably not suitable for all the predictions mentioned above. For example, impact on cognitive measures is probably best predicted by other machine learning methods, including regression. Fortunately, language models already have the ability to offload certain tasks to a code interpreter.?

A new age of artificial intelligence is here and brings with it both warning and good news, making the ethical implications important to consider. As researchers, corporations, and states develop such capabilities, it would be critical to ensure that they are used not just in accordance with laws, but the philosophical notion of justice. Any prevalent biases, injustices, and inequities can magnify exponentially. For example, adverse childhood experiences are the biggest factor behind adult mental health issues, and disproportionately affect low income households; if employment decisions were based on such behavioral models, they may worsen income inequality. There are several states in the world with poor human rights records, where authoritarian governments seek to tightly control thought and speech. Their ability to do so may improve exponentially, as they utilize such behavioral models to personalize the manipulation and control of each human being.?

With great power comes great responsibility, but it is possible to wield this power in alignment with the dictates of justice, rather than the justice of dictates. The following measures may be considered to maximize the probability of this happening. Firstly, before such technology is applied to judge the public, it should be used on political, religious, and business leaders, and anyone in a position involving decisions that impact the public. In particular, such individuals should be screened for psychopathic traits. Secondly, mental healthcare should be accessible to every human being, and possibly declared a fundamental human right. In a world where behavior can be modeled and states of consciousness inferred, this would level the playing field somewhat. Thirdly, cognitive security technology should be developed to make it possible to detect and counter psychological operations designed to influence and manipulate minds in alignment with the agenda of states, corporations, and non-state actors. The simplest version of this technology can be a wearable wrist-band that notifies and advises the user through haptic feedback when they encounter a situation where their cognitive security is being attacked. Finally, user privacy should be made sacrosanct, and violation thereof should be a serious crime.

All these possibilities require us to reconsider carefully our notions of justice and fair competition. The ideas above are suggestions, not prescriptions, meant to initiate rather than conclude discussion on the complex topic. How would such simulations represent various cultures and societies, especially those at conflict? What happens when one society seeks to optimize minds at scale, while another tolerates or even cultivates psychopathic traits as weapons for cultural, racial, or ideological domination? What story would such a simulation tell about global conflict? The answers to some of these questions may be uncomfortable, but we must confront them before it gets too late.