Work

From Honesty to Observable Duties: How EU, Chinese and US Law Translate an Epistemic Requirement for AI

Year: 2026
Type: article
Abstract: BSI Research Note 03 defined honesty in artificial intelligence (AI) systems as an epistemic property: a system’s 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. Law can effectively impose and enforce only obligations that can be evidenced through observable conduct, outputs, records, procedures or attributable decisions. Because the correspondence between a system’s assertions and its internal representations cannot be directly observed with sufficient reliability, current law translates honesty into external and legally assessable proxies. This note examines how three legal orders, treated as ideal types, perform that translation: the European Union through transparency and, for high-risk systems, accuracy; China through requirements concerning the accuracy and authenticity of generated content and training data, together with labelling; and the United States through ex post enforcement against deception. None of the provisions examined uses the word honesty to qualify an AI system. They regulate observable proxies — disclosure about the interlocutor, accuracy of content, labelling of outputs and prohibition of deception — and bind legally responsible actors, including providers, deployers, operators and other natural or legal persons, not AI systems as legal subjects. The note then identifies what current law does not directly reach. Within the system, law governs inputs and outputs and may indirectly shape internal processes through documentation, testing, risk management and oversight, but it does not directly regulate or reliably render observable the correspondence between internal representations and assertions. Honesty is also exposed to operational pressures of cost and speed, which a duty of proportionate diligence could address if its required measures and the authorities responsible for defining them are specified. The note also shows that a poorly drafted accuracy obligation could reward the appearance of accuracy at the expense of calibration. Between legal orders, no binding or commonly accepted multilateral framework currently harmonises their regulatory logics. The note proposes three directions: a requirement to declare and validate the reference bases on which deployed systems rely, thereby giving truthfulness, legal alignment and auditability an identifiable reference; a duty of proportionate diligence, with the required measures and the competent authorities specified; and, at international level, technical interoperability and mutual readability of labelling and marking formats, followed where appropriate by mutual recognition of compliant formats. It also outlines, without developing it, one further line: legal alignment at each step of agentic delegation. Throughout, it distinguishes three terminological registers: the conceptual definitions adopted by the BSI Lexicon, technical AI usage and legal usage.
Language: English
Subfield: Law
Open Access status: green
OpenAlex ID: W7226027448
FWCI:
Cites:
Cited by:
Related to: 10