Introduced in the collection of stories I, Robot 1950, the laws of robotics formulated by the writer Isaac Asimov (1920-1992) are a set of guidelines to regulate the behavior of androids and guarantee the safety of human beings in their interactions with intelligent machines. It presides over all the stories contained in the volume, almost always in a future in which synthetic figures would be a common reality. The first and fundamental rule states that “a robot may not harm a human being or, through inaction, allow a human to come to harm,” and thus becomes the cornerstone of a brave new world. In the 2004 film adaptation of the anthology, detective Del Spooner suspects that one of these creations, Sonny, is the possible murderer of a famous scientist.
If we talk about this year 2025, just begun, the rapid evolution of artificial intelligence (AI) raises similar fantasies in many people and it is not strange to hear that “what will be the limit of AI” if it manages to become independent of human designs .
A partnership between researchers from the Stanford University (United States) and Google DeepMind allowed the creation of a model capable of cloning a person’s personality. It was not even necessary to develop a revolutionary technology, never seen before. Through the capabilities of the language model behind ChatGPT OpenAI, the most popular chatbot of the moment, the experts replicated the character, so to speak, of more than a thousand people, based solely on two-hour interviews. It was simple: after asking the interviewees a series of questions about childhood, memory and career, among other topics, the computer predicted the behavior of the people involved in the work.
“The promise of human behavior simulation (general-purpose computational agents that replicate human behavior across domains) could enable broad applications in policymaking and the social sciences,” the scientists say in their paper. paperwhich was published on arXiv, an open access scientific research repository, but not subjected to peer review, that is, by other experts in the area.
And they describe their work and findings: “We present a new agent architecture that simulates the attitudes and behaviors of 1,052 real individuals, applying large language models to qualitative interviews about their lives and then measuring how well these agents replicate the attitudes and behaviors. behaviors of the individuals they represent. Generative agents replicate participants’ responses on the General Social Survey with 85% accuracy as participants replicate their own responses two weeks later, and have comparable performance in predicting personality traits and outcomes across replications. experimental. Our architecture reduces accuracy biases in racial and ideological groups compared to agents given demographic descriptions. “The promise of human behavioral simulation – general-purpose computational agents that replicate human behavior across domains – could enable broad applications in policymaking and the social sciences.”
The slogans
At the head of the scientific team, Joon Sung Park He conducted the survey as part of his PhD in computer science from Stanford. The people recruited received one hundred dollars for participating in the study and represented different ages, genders, racial groups, geographic regions, educational levels and political ideologies. From interviews with them, the team created agent replicas of those individuals.
To check how well the agents imitated their human counterparts, participants took a series of personality tests, social surveys, and logic games, twice each, two weeks apart; The officers then completed the same exercises. The results of these comparisons showed that 85 percent similarity.
In the article, the replicas are called simulation agents, and the goal in creating them is to make it easier for researchers in the social sciences and other fields to conduct studies that would be expensive, impractical, or unethical if done with real human subjects.
These simulation agents are slightly different from those that dominate the work of major AI companies today, which are called “tool agents.” Trained from emails, text messages and other personal files, which help create a kind of feedback repository, these agents are models developed to do specific tasks in place of a person, not to talk to a person. They can, for example, enter data, retrieve information you have stored somewhere or, one day, book trips and schedule appointments.
The new technology, however, goes further. Simulation agents aim to mimic human behaviors and personalities, allowing researchers to study real-world dynamics in controlled environments. The underlying idea is that if it is feasible to create AI models that behave like real people, they can be used to test everything from, for example, how effective interventions on social networks are to deal with misinformation to what behaviors cause traffic jams. By simulating human behaviors, they provide a scalable and ethical alternative to involving real participants in large-scale or sensitive studies.
For now, the tool has only been tested with a limited number of evaluations. Therefore, it is not possible to know if it would be equally efficient when replicating responses to more complex situations, or that went beyond the scope of the questions asked during the interview. In any case, the development makes it possible to identify potential dangers.
The fine print
Today, the most widespread tools already allow the creation of deepfakes, audios and videos that were created digitally, without the use of large resources or the consent of the real people involved, and that make us believe that someone said something they never said. . They are already used to defraud and to exercise massive manipulation for political purposes. So Sung Park’s team’s research involves some risks: Any agent-generating technology raises questions about the ease with which people can create tools to impersonate others online, saying or authorizing things they had no intention of saying. say.
Furthermore, other works, led by Dongwook Yoon, from the University of British Columbia (in Canada), raised possible risks of artificial “subjects,” such as the growing fear of machines and the difficulty of creating bonds, due to the widespread presence of computerized clones. The evaluation methods the team used to test how well the AI agents replicated their corresponding humans were also fairly basic.
These included the General Social Survey, which collects information on demographics, happiness, behaviors and more, and assessments of the Big Five personality traits: openness to experience, conscientiousness, extraversion, agreeableness and neuroticism. These tests are commonly used in social science research, but they are not intended to capture all of the unique details that make a person who they are, with all their uniqueness and complexity. AI agents were also worse at replicating humans in behavioral tests like the “dictator game,” which aims to show how participants view values like justice.
“Those risks can be minimized through rigorous measures, ensuring that individuals are fully aware of how their clones are used,” says Park, perhaps overly optimistic. Digital twins, that’s where some artificial intelligence models are looking that are designed to replicate individual personalities.
Image gallery
