Kush R. Varshney

We propose an overarching framework for agentic artificial intelligence (AI) governance by drawing technical parallels from Advaita Vedanta philosophy. Our framework maps key concepts — atman, prakrti, dharma, jiva, and karma — to modern AI components. We use it to suggest solutions for the identity problem and the control problem of agentic AI. We approach the identity problem through the concept of citta, an immutable record of conduct that serves as a unique identifier for an agent. To address the control problem, we introduce an intrinsic governance model based on an asymmetric actor–critic architecture, where an internal reward model ensures alignment without the need for high-cost external guardrails. Moreover, we suggest technical specifications including Merkle-based structures for secure identity and simulated student phases for robust adversarial training.

AI researchers and ethicists have long worried about the threat that automation poses to human dignity, autonomy, and to the sense of personal value that is tied to work. Typically, proposed solutions to this problem focus on ways in which we can reduce the number of job losses which result from automation, ways to retrain those that lose their jobs, or ways to mitigate the social consequences of those job losses. However, even in cases where workers keep their jobs, their agency within those roles might be severely downgraded. For instance, human employees might work alongside artificial intelligence (AI) but not be allowed to make decisions or not be allowed to make decisions without consulting with or coming to agreement with the AI. Here, we argue that this is a kind of epistemic harm (which could be an injustice if it is distributed on the basis of identity prejudice). It diminishes human agency (in constraining people’s ability to act independently), and it fails to recognize the workers’ epistemic agency as qualified experts. Workers, in this case, aren’t given the trust they are entitled to. This means that issues of human dignity remain even in cases where everyone keeps their job. Further, job retention focused solutions, such as designing an algorithm to work alongside the human employee, may only enable these harms. In response to this, we propose an alternative design solution, adversarial collaboration, which addresses the traditional retention problem of automation, but also addresses the larger underlying problem of epistemic harms and the distribution of trust between AI and humans in the workplace.