Abstract
Asking whether we can know whether a given text or output is coming from a machine or a human, which was at the core of the Turing test, is being progressively retired. This opens a new and no less interesting question: what happens when we know? This paper reviews evidence that when humans encounter exactly the same content, presented here as coming from a human and there from an AI, they do not behave similarly. The evidence across different kinds of interaction, advice giving, compliance, reciprocal interaction, and preferential choice, is that humans respect other humans more than they respect machines. This difference is often described as a 'machine penalty' where we distrust machines or lack familiarity with them, and predicted to decrease as AI becomes more prevalent. The paper offers a different interpretation, showing that we treat humans better for good reasons, including utility calculations, and that this human premium is likely to stay.