Research Funded by DARPA: AI Detection of Textual Behaviors Violating Social Norms

2023-07-27

The Defense Advanced Research Projects Agency (DARPA) has temporarily shifted its focus from autonomous flying helicopters and human recognition robots to providing resources for a peculiar AI implementation. With funding from DARPA, researchers from Ben-Gurion University in Israel have created an AI tool that can detect violations of social norms. This tool reliably identifies 10 different types of social norm violations, providing psychologists and sociologists with a unique way to study human behavior. According to a paper published in Scientific Reports, researchers Yair Neuman and Yochai Cohen used GPT3, a natural language interface (NLI)-based zero-shot text classifier, and automated rule discovery to create this system. The powerful language prediction model GPT3 allowed the researchers to establish a set of foundational social norm categories for evaluating violations. As the system is trained based on text, Neuman and Cohen utilized NLI-based zero-shot classification to analyze the premises and relationships in each text example. At the same time, automated rule discovery provided the foundation for violation detection, allowing social scientists to interact with the tool and understand the process, rather than relying on opaque processes generated by most models. This tool can identify behaviors that violate 10 types of social norms: ability, politeness, trust, discipline, care, kindness, success, conformity, dignity, and loyalty. When researchers input a textual context, the tool evaluates whether the text describes a violation of social norms. In a text resembling a soap opera plot, which describes a married couple with tendencies to deceive each other, the tool identifies shame. Conversely, in a text describing academic success, the tool recognizes that the "narrator" adheres to social norms (i.e., performs well in school) and is rewarded (pride). "When people feel ashamed, embarrassed, or regretful, it can be assumed that they are aware of violating social norms," wrote Neuman and Cohen. "These consequential social emotions seem to have an evolutionary basis in social devaluation mechanisms and play a clear role in supporting cooperation." The researchers pointed out that the importance of social norms varies among social groups and even individuals, meaning that individuals may avoid meeting the requirements of these norms when they perceive them as not aligning with what they consider "right." Therefore, tools like this may provide a new research avenue for studying human decision-making, values, cooperation, and even altruism (as violating social norms often has consequences, as mentioned above). While this sounds promising, it does not explain why DARPA would be interested in and provide funding for a system that studies social norms and identifies inevitable violations in any social environment. Source: https://www.extremetech.com/defense/darpa-funded-study-uses-ai-to-detect-violations-of-social-norms-in-texts

RECENT AI NEWS

RECENT AI TOOLS