Grounded AI Safety is a principle-driven approach adopted by the Indian Society of Artificial Intelligence and Law for The Bharat Pacific Stack, to managing risks in AI systems, rooted in the fundamental understanding that current AI, such as large language models, functions as statistical pattern-matchers without true comprehension or reasoning ability. This approach:

• Anchors in Observable Limitations: Risk mitigation begins with empirical evidence of AI’s inherent constraints, such as struggles with tasks requiring conceptual understanding—like misinterpreting time differences across regions or failing to follow rules in strategic games—focusing on these measurable shortcomings rather than assumed capabilities.

• Centers on Human-Driven Risks: The primary dangers arise from human over-reliance on or misuse of these limited systems, such as deploying them in critical areas like scheduling or decision-making where their errors could lead to significant consequences, rather than from AI autonomously causing catastrophic outcomes.

• Rejects Speculative Existential Narratives: AI safety must exclude unproven predictions of AI-driven doomsday scenarios that lack evidence and inflate AI’s potential, as these narratives misguide priorities and empower those who might exploit fear for profit, influence, or excessive control.

• Prioritises Evidence-Based Safeguards: Solutions involve systematic testing to identify and address specific failure modes—like errors in visual representations or logical reasoning—paired with transparent improvements, ensuring AI systems are used responsibly within their known boundaries.

This definition is inspired by a post by Dr Gary Marcus, on X.