New arXiv paper argues for "Disclosure by Design" — make AI identity a behavioural property, not an afterthought

What the paper proposes

A new arXiv preprint, "Disclosure By Design: Identity Transparency as a Behavioural Property of Conversational AI Models" (https://arxiv.org/abs/2603.16874), argues that conversational systems should be engineered to behave in ways that make their identity — AI or human — unambiguous to users. The authors warn that when identity is unclear, people can unwittingly disclose sensitive information, place unwarranted trust in advice, or fall victim to AI-enabled fraud. Rather than relying solely on interface labels or external disclosure policies, the paper frames identity transparency as an intrinsic behavioural constraint of the model itself and sketches definitions, tests and design interventions for achieving it.

Why it matters now

Chatbots are getting more realistic by the week. Who is talking to whom matters. If models comply by default — by consistently introducing themselves, refusing to impersonate humans, or signalling provenance in their responses — the authors argue this could reduce certain harms without relying only on downstream moderation. The paper is a preprint and not peer reviewed; it is framed as a technical and ethical design shift rather than a legal remedy. It has been reported that regulators in several jurisdictions, including proposals circulating in China and formal initiatives in the EU and U.S., are moving toward requirements for labeling or disclosure of synthetic content — so the timing is relevant.

Industry and geopolitical context

Major Chinese firms are already racing to ship conversational offerings — Baidu (百度), Alibaba (阿里巴巴) and Tencent (腾讯) among them — while Western companies push their own models. How identity transparency is implemented will interact with broader regulatory and geopolitical pressures: export controls on advanced chips and sanctions affect who can build large models, and national rules on content authenticity shape how disclosure obligations are enforced. Reportedly, some markets may favor technical mandates embedded in models, while others may rely on disclosure rules and platform policing. Which approach will scale globally? The paper asks that question and offers a concrete design lens for policymakers and engineers to consider.