When prompts learned to move: how AI video generation rewrote the grammar of human-AI co-creation

When generative artificial intelligence moved from producing still images to producing moving images, the way people wrote prompts changed with it. A comparative observational study by Yigang Liu and Jiayang Ma of Nanjing Forestry University, published in Frontiers in Human Dynamics, examines how prompt expression shifted as text-to-image creation gave way to text-to-video generation between 2022 and 2025.

The researchers compared two large-scale public prompt corpora, one drawn from image generation and one from video generation, under a single neural topic-modeling protocol. The protocol combined lexical analysis of style, quality, technical, motion, camera, and format vocabulary with manual family coding and validity checks, and added an exploratory user-focused topic structure.

The central finding is that video prompts weaken the style, quality, and technical decoration characteristic of image prompts. Instead, meaning in video prompts organizes itself around action, scene, and screen format. The terms that organized image prompts did not carry over intact to video prompts.

Artist-style imitation, which had been prominent in image prompting, loses its earlier prominence in video prompting. According to the study, this decline survives the coding reliability checks. Meanwhile, everyday action and scene description forms the largest family in the adjudicated solution, although the authors note that this position is sensitive to those checks.

The authors are careful about causation. Because the two corpora differ not only in modality but also in platform ecology, model generation, period, and corpus construction, the observed differences are described rather than causally attributed to the shift from image to video. The screen-format result is identified as the most interface-sensitive of the findings.

Overall, the study concludes that prompting practice moved from selecting aesthetic styles toward specifying how a scene moves and how it is framed. The prompting grammar observed here appears bound to the modality and the period in which it was collected.

These patterns suggest that the symbolic interface of human-AI co-creation is not a fixed grammar but an adaptive arrangement that responds to the salient capabilities of a generative system. As models gain new capabilities, the vocabulary and structure users employ to direct them shift accordingly.

The study contributes a comparative account of how the division of creative labor is reconfigured across media modalities, with implications for co-creation theory, generative interface design, and media literacy education. As a comparative observational study based on two public corpora, it documents an association between modality and prompting practice rather than a causal effect, and its conclusions are bounded by the differences between the corpora beyond modality alone.

Subject of Research: How prompting practice changes when generative AI shifts from text-to-image to text-to-video creation

Article Title: From text-to-image to text-to-video: the reconfiguration of prompting practice in human-AI co-creation in 2022–2025

Article References: Liu, Y., & Ma, J. (2026). From text-to-image to text-to-video: the reconfiguration of prompting practice in human-AI co-creation in 2022–2025. Frontiers in Human Dynamics, 8, Article 1990681. https://doi.org/10.3389/fhumd.2026.1990681

Image Credits: AI Generated

DOI: 10.3389/fhumd.2026.1990681

Keywords: generative AI, text-to-video, text-to-image, prompting, topic modeling, human-AI co-creation, prompt literacy, BERTopic, semiotic system, media modality, reconfiguration, practice