API
< Back to all clusters
[TECHNOLOGY] · Japan · 4 sources

started · updated

Generative AI faces risks from unconscious bias and prompt injection

Generative AI technologies are presenting new challenges regarding social bias and cybersecurity vulnerabilities. Research indicates that AI models often reflect unconscious biases, such as gender stereotypes in professional roles and traditional color associations for children's products. Experts warn that these biases could influence the career choices and perceptions of young users who rely on AI for advice.

Simultaneously, security risks are escalating through prompt injection attacks. These occur when malicious instructions are embedded within documents, such as PDFs, to manipulate an AI into ignoring original commands or leaking sensitive data. As AI evolves from a simple response tool into an autonomous agent capable of executing tasks, the potential for unauthorized data access and operational errors increases, necessitating new defensive strategies and literacy training.

Entities

Hiroyuki Tanaka · Waseda University · generative AI