Honesty in AI Systems: A Terminological and Conceptual Analysis (BSI Research Note 03)

The word honesty is used in artificial intelligence (AI) research to name at least four different things: an alignment criterion, a set of measurable behaviours, a normative specification written by developers, and a moral virtue projected onto machines. This note delimits the term. It proposes an operational definition: in AI systems, honesty is the property of a system whose assertions correspond to its own internal representations, are expressed with confidence proportionate to the available evidence, and are not intended or structured to create false or unwarranted beliefs in the interlocutor. Each of the three conditions is established in the literature; the contribution claimed here is their joint statement and the explicit tie between non-deception and calibration. The note distinguishes honesty from truthfulness, accuracy, explainability and moral honesty, and separates three failures — lying, deception and indifference to truth. A comparative study of Latin, Greek, Sanskrit and Chinese designations shows that the AI notion, centred on calibration and non-deception, is closer to Aristotle’s alētheutikós and to Analects II.17 than to the Latin word that names it. The note then treats honesty as a candidate target of alignment rather than a component of its definition, and examines its role in agentic delegation. Honesty is not obedience to instructions but one of its limits; in delegation it bears on the relation between what an agent is instructed to do, what it does and what it reports. On this reading honesty is the condition under which alignment can be verified, and it returns the human parties to their own responsibility for the instructions and data they provide.