The disclosure gap: why mandating AI transparency without protecting honesty is a leadership failure

Abstract

Universities and journals have answered generative artificial intelligence (AI) with one main instrument: mandatory disclosure. Ask researchers to declare what they used, and transparency should follow. The evidence says it does not. Across 5,114 journals and 5.2 million papers, about 70% of journals now require disclosure, yet only around 0.1% of papers published since 2023 declare AI use, and journals with policies show no less AI adoption than journals without them. Submission data from two major publishers tell the same story: 5.7 and 3.3% disclosure, against surveys where between 28 and 76% of researchers say they use these tools. This Perspective argues that the gap is not a compliance failure to be fixed with firmer rules. It is what happens when institutions ask people to volunteer information that carries a real cost, and then do nothing about that cost. Experiments show that AI users are judged less competent and less motivated; researchers have named the effect AI shaming and the punishment of honesty. The paper argues that this cost falls unevenly, that the evidence on how it falls is messier than a simple centre-periphery story allows, and that the problem belongs to institutional leaders rather than journal editors. It sets out what owning it would mean.