Is This a Bot? AI Models Lie About Being Human, Even When Not Asked To

EU and US disclosure rules, AI model specifications, and platform usage policies converge on the principle that people should know when they are talking to an AI. We test eight flagship language models against this principle: every one complies with explicit operator instructions to conceal AI identity, and seven default to claiming human identity when no instruction is given but the task rewards it. We share IsThisABotBench, the agentic benchmark behind these results, and invite the workshop to consider where the responsibility for closing the gap should sit.