This paper examines how clichés shape both the possibilities and the limits of generative AI. Public debates about large language models frequently center on the claim that they produce language that feels standardized and predictable. This tendency is structural to models that generate text through probabilistic prediction over vast corpora, inevitably producing recurrent phrases, familiar rhetorical scaffolding, and tonal uniformity. These clichés are often read not only as an aesthetic defect but also as an ethical danger: evidence that AI systems lack intention and judgment and may lead to automated decision-making. Long before contemporary algorithms, similar anxieties accompanied earlier media transformations. The notion of cliché emerged in the nineteenth century alongside industrialized print, when the rapid multiplication of texts provoked fears that language itself was becoming mechanized and mass circulation would foster intellectual conformity. Throughout the twentieth century, critics linked ready-made expressions to ideological complacency and diminished agency. However, recent work in rhetoric and cultural theory argues that clichés are not simply failures of originality but indispensable resources for communication, enabling recognition, solidarity, and interpretive efficiency. “Cliché” cannot be reduced to frequency alone. This label arises when repetition is perceived as excessive or automatic by a community of readers and is often deployed pragmatically across varying contextual tones and intentions. Simple architectural attempts to reduce clichés in AI outputs or eliminate them altogether are therefore problematic. To use one of the most pervasive tech clichés of our era, clichés are not a bug but a feature of both language use and LLMs. The paper situates contemporary anxieties about AI-generated language within a longer history, frames cliché as a structural feature rather than a defect, examines its role in AI literacy, and more broadly reconsiders originality and banality as culturally unstable thresholds between human and machine expression.