<p>This paper investigates how people perceive and evaluate authorship in AI-generated texts, situating these practices within broader questions of AI literacy. Using data from a two-year experimental study (<i>N</i> = 88) on blind assessments of paired human and AI-authored texts across three genres, we examine the interpretive strategies respondents mobilize. Theoretically, we draw on Gregory Bateson’s metacommunication and Erving Goffman’s frame theory to analyze how communicative frames become unstable when the nature of the author is uncertain, and what interpretive labor is required to navigate that instability. Our results show that overall attribution accuracy seldom exceeds chance, and that reliance on surface-level heuristics, such as grammar checks and formatting cues, frequently backfires. In contrast, domain expertise, sender familiarity, and contextual knowledge are stronger predictors of correct attribution. Participants achieve their highest detection rates where multi-patterned heuristic evaluation is possible, as in genre-rich, culturally familiar contexts, and their lowest where formal institutional conventions constrain interpretation to surface features alone. Paradoxically, incorrect respondents are more likely to report high certainty than correct ones, suggesting that reliance on isolated cues produces false confidence. These findings recast authorship attribution as a cultural practice rather than technical competency. Situated and tacit responses, including whether a text feels captivating, prove more predictive than formal linguistic analysis. We conclude by arguing for a reconceptualization of AI literacy centered on what we term metacommunicative agility: the capacity to shift fluidly between interpretive frames, navigating meaning across unstable communicative registers rather than seeking definitive detection markers.</p>