Google Analyzes 2.1 Million Chats for 5 Uses of AI Image

Researchers at Google have categorized 2.1 million exchanges with Gemini Nano Banana into five uses without inspecting the images. The result: the presence of people dominates, memes remain marginal, and marketing proves to be very prescriptive regarding text and layout.

Trends Contrary to Common Beliefs, Memes in Retreat

Memes and humorous images rank among the least observed formats in the social and viral sphere. In contrast, the inclusion of people is the most widespread characteristic, appearing in nearly 80% of cases. Requests concerning the structure of the image or artistic styles also appear to be common, while soliciting a new generation from the model occurs approximately twice as frequently as making small modifications. The identified use cases encompass commercial domains, educational illustration, technical instruction writing, and the creation of visuals for products. According to the researchers, generative image AI promotes a plurality of creative and professional activities, while enabling a broader audience to accomplish tasks that were previously the domain of specialists. They anticipate other forms of use as experimentation progresses, while emphasizing uncertainty about the proportion of genuinely new activities compared to existing practices that are simply less costly.

Marketing: Visual and Typographic Requirements at the Forefront

In the Brand and Marketing category, a large portion of conversations revolves around original, fictional, or unrecognized brands, which may correspond to identity creations for new structures or speculative projects. Marketing exchanges are characterized by detailed visual requirements: control of composition, placement of elements and people, management of lighting, and exclusion lists. Requests here display the highest prevalence of precise text and typographic instructions among the five categories. Small businesses, individuals, and side activities are also more represented than large brands. Use cases go beyond creation ex nihilo to include logo refinement, text and color changes, production of promotional materials, and defining coherent visual identities, with variations intended for online engagement as well as traditional media.

Observation Without Images: An Analysis That Preserves Privacy

The researchers explain that they identified uses based on the text of requests and exchanges, without ever consulting the uploaded or generated images. This approach aims to address privacy concerns, especially since sensitive content regarding users and their relatives may be shared within the tool. The Urania framework, already deployed at Google for analyzing interactions with language models, enabled a descriptive analysis while preserving privacy and revealing usage patterns.

A Massive Corpus and Five Families of Uses

The analysis is based on a corpus of 2.1 million exchanges with Gemini Nano Banana, divided into five groups: Brand and Marketing, Product Visualization and E-commerce, Architectural and Spatial Design, and Educational and Technical Information. The authors explain this collective choice by the need to identify burgeoning uses that are still poorly described outside of self-reports, in order to catalog applications and guide potential tool developments. The analysis specifically aims to contribute to commerce, education, social connections, and scientific endeavors.

Mobilized Functions: Editing, Style, Composition, and Iteration

The researchers compared the categories according to several capabilities of the tool. In terms of editing, requests cover the removal of objects or people, background changes, the addition of elements, resizing, and restoration. Stylistic orientations include explicit mediums, such as watercolor, pencil sketch, photographic style, or oil painting, as well as atmospheres like minimalist, film noir, or vintage, sometimes with conceptual blends. Composition encompasses scene arrangement, subject placement, perspective, lighting, negative constraints, and indications on the number of people. Interaction dynamics observe how users iterate and modify outputs over the course of a session. Finally, multiple selection features allow for annotating additional dimensions, such as the presence of text and typography, type of transformation, work context, and creation stage, with subclasses ranging from complete specification to content only, generic style only, or absence of any textual mention.