Google releases AI framework to help people define subjective image categories

Google Research introduced Agile Deliberation on November 4, 2026, a framework that helps people clarify subjective image categories before asking AI to apply them. Rather than requiring a finished definition upfront, it presents borderline images and updates classifier instructions from human feedback. Google also released the underlying code and algorithms.

The starting problem is that a label such as “healthy food” does not carry one fixed meaning. Google’s researchers describe boundaries that change with dietary restrictions, cultural context and personal beliefs. People may begin with an intuitive definition, then discover exceptions only when an actual image forces a decision.

That creates a mismatch with workflows that treat human labels as settled instructions. Agile Deliberation instead makes defining the category part of the work. The human can revise the rules as ambiguous examples appear, while the system handles the corresponding changes to the classifier’s prompt.

From a broad label to a difficult image

The first stage, concept scoping, breaks a broad category into smaller groups and explicit exclusions. For healthy food, a user might include salads, grain bowls and lean proteins, while excluding processed food or raw, unprepared ingredients. Those choices establish an initial definition, not an assumption that everyone agrees about nutrition.

The second stage tests those boundaries. Vision-language models—AI systems that work with images and text—find images that challenge the current wording. Google illustrates this with a fruit smoothie: should it count as healthy because it contains fruit, or fall outside the category because of processing or high sugar content?

The user’s answer feeds an automatic prompt-refinement loop. This differs from conventional active learning, which selects examples to reduce a model’s uncertainty. Here, the examples are also meant to expose uncertainty in the person’s definition. Users give feedback on a gallery of images rather than finding edge cases and rewriting prompts themselves.

18Live user sessions

Google evaluated the prototype in 18 live user sessions.

1.5 hoursLength of each session

Participants used the framework and alternatives during sessions lasting 1.5 hours each.

Testing definitions as they change

Google used live sessions because a fixed, pre-labeled dataset cannot capture how a person’s definition changes during deliberation. Participants tried the approaches in randomized order to reduce learning bias. The tasks covered healthy food and “Pay-to-Play”: identifying entertainment or gaming clickbait images that promise misleading financial or digital rewards.

The comparisons included a one-shot classifier using an initial prompt, an automated collaborator that decomposed concepts without user correction, and manual deliberation. In the manual approach, users inspected spreadsheets, searched for anomalies and edited prompts by hand. The comparison therefore tested different ways of defining rules, not just different image models.

Google reports that Agile Deliberation produced more consistent human labeling and classifiers that outperformed the baselines, especially on difficult subjective images. Expert participants also said the borderline examples exposed blind spots in their original definitions. These are results from the small live study, not evidence of performance across every subjective classification task.

Participants reported statistically significant reductions in effort and stress compared with manual prompt engineering, according to Google. Manual editing caused anxiety that fixing one error would break something already resolved. Every participant preferred Agile Deliberation to the manual methods.

A prototype others can adapt

The research was conducted in partnership with Google Ads. Alongside the code on GitHub, Google provides an interactive Colab/Jupyter demo built around the healthy-food example. It suggests adapting the workflow to domains where users have subject expertise, such as misleading clickbait, or personal preferences such as inspiring artwork. Those are suggested applications; the reported evaluation covered only the two study domains.