Shadow AI thrives under punitive social evaluation

Generative Artificial Intelligence (GenAI) tools, such as ChatGPT, offer significant performance benefits across professional tasks. Yet, their adoption in work-related contexts is complicated by social disapproval and penalties, especially under conditions of mandated transparency. In three studies (one pre-registered; n = 1,678 applicants and n = 477 evaluators), we investigate how people navigate this augmentation-approval tradeoff in an incentivized mini-job application scenario. We find that mandatory disclosure substantially reduces visible AI adoption, but prompts a covert behavioral strategy we term shadow adoption, that is, using AI in ways that avoid detection and disclosure. Strikingly, these shadow AI users produce the highest-quality applications, as rated by HR professionals who are unaware that the outputs were AI-assisted. As the knowledge about the tradeoff spreads, shadow adoption becomes more prevalent, with nearly twice as many people choosing to use shadow AI. These results reveal a misalignment between well-intended transparency rules and user incentives in work-related contexts. Policies and technologies designed to enforce ethical AI use may inadvertently encourage covert behavior, rewarding concealment over compliance.