Building Trustworthy AI by Addressing its 16+2 Desiderata with Goal-Directed Commonsense Reasoning

LLMs have a fundamental transparency problem: The paper identifies that large language models like ChatGPT can "hallucinate" (generate false information) and their decisions "cannot be explained or audited," which poses serious risks for trustworthy AI in legal, ethical, and commercial applications.

A hybrid approach may offer better solutions than pure AI methods: Rather than choosing between black-box machine learning or overly complex rule-based systems, the researchers propose s(CASP), a middle-ground technology that combines explainability with practical usability—allowing AI systems to show their reasoning steps while remaining manageable to implement.

Explainability and auditability are non-negotiable for trustworthy AI: The paper emphasizes that trustworthy AI systems must be able to explain their reasoning in a way humans can follow and verify, addressing the 16 desiderata (essential requirements) for trustworthy AI established by leading AI researchers Doug Lenat and Gary Marcus.

Commonsense reasoning can be reliably programmed: The proposed s(CASP) system demonstrates that human-style commonsense reasoning—the ability to make logical inferences and detect contradictions—can be implemented in AI systems in a way that is both reliable and transparent, with applications ranging from chatbots to embodied AI agents.