Beyond viral images: What 2.1 million Gemini chats reveal about AI image use

A request for a whimsical portrait and one for a technical diagram may use the same image-generation tool, but they place different demands on it. Understanding those differences is difficult: public examples offer a selective view, while private conversations can contain sensitive information about people, families and businesses.

A team led by Will Tucker-Ray, Marika Swanberg and Alexander Knop examines that problem in an arXiv preprint posted on November 5, 2026. All authors list affiliations with Google teams. Their study describes how people requested images from Gemini’s Nano Banana models, rather than testing whether the resulting images were accurate or useful.

Studying requests without inspecting images

The researchers started with a worldwide random sample of Gemini conversations from 90 days in the first quarter of 2026 that used Nano Banana. They selected 2.1 million conversations, containing more than 5.2 million prompts, within five areas: social and viral sharing, branding and marketing, educational and technical information, product visualization and e-commerce, and architectural and spatial design.

This was a deliberately filtered sample. Conversations with unclear classifications were removed. For the social category, researchers reviewed privacy-protected cluster summaries and retained clusters matching trends identified through media coverage and public posts. Clusters with fewer than 50 conversations were excluded. The final analysis covered roughly 800 clusters.

The team used Urania, a Google-developed framework that groups conversations and produces summaries under differential privacy. This approach mathematically limits how much any individual conversation can influence released results. The authors report a combined privacy budget of epsilon no greater than 0.04, with delta equal to zero.

The privacy unit was a conversation, not a person’s entire history. Researchers did not inspect users’ reference images, generated images or individual prompts. Instead, automated analysis rated conversations for features such as explicit lighting instructions, precise positioning, artistic style and follow-up requests. Results were aggregated into privacy-protected statistics.

People, placement and different purposes

In social and viral clusters, requests involving people were the most prevalent feature, appearing at nearly 80%, according to the paper. Instructions about mood, artistic medium, lighting and the arrangement of multiple subjects were also common.

Explicit meme formats and requests for absurdity or humor were among the least prevalent features in that category. Social image use, in this selected sample, was not synonymous with joke-making. Cluster summaries included birthday cards, sports jerseys, cartoon transformations and age-progression images.

Starting over with the same or a nearly identical prompt was about twice as prevalent as asking for a simple adjustment in social conversations. That describes an interaction pattern; it does not establish why people restarted or how satisfied they were.

Architectural requests emphasized precise placement, lighting and scenes containing multiple objects or subjects. Examples included floor plans, furniture arrangements and exterior modifications. Requests involving people or a particular mood or style were comparatively uncommon within that category.

People were also the most prevalent feature in educational and product-visualization conversations. The authors suggest that anatomy illustrations, hands-on repair instructions and products displayed on models could help explain this. These are possible interpretations, not observations verified against the images.

Text requirements provided another distinction. Educational and technical conversations requested text or typography around a third of the time. Branding and marketing had the strongest emphasis on exact wording and on specifying both wording and visual styling. Most architectural, product-visualization and social clusters showed no text or typography requests.

What the analysis cannot establish

The central limitation is that the study measures expressed intentions. Someone uploading two images and writing “make this look like this” may convey information that text-only analysis cannot recover. Conversely, a detailed request says nothing about whether the model successfully followed it.

The researchers also could not validate their classifications by having people inspect the underlying private conversations. Automated labels can therefore be wrong, while privacy protections introduce statistical noise. The authors describe their findings as directional rather than necessarily precise.

Selection matters too. Filtering for clearly identifiable purposes means these results should not be treated as a census of all Nano Banana use, or as evidence of the overall market share of social versus professional applications. The study captures one service during one quarter, with potential seasonal influences.

Finally, it cannot follow people across conversations or determine whether images were subsequently shared, used in lessons or deployed commercially. It does not establish productivity gains, learning benefits or displacement of existing software. Its contribution is narrower: a privacy-protected description of diverse requests and prompting patterns that future research can investigate for actual benefits and harms.