Teaching a model to spot a stop sign is easy because everyone agrees what a stop sign is. Teaching it to spot healthy food, inspiring art or misleading advertising is harder, because the person giving the instructions often does not know where their own line falls until they see a difficult example. Google Research's Agile Deliberation, presented at CVPR 2026 and now released as open-source code, is built around that problem, and it matters to anyone writing content rules or curating images at scale.
The method has two stages. First the user breaks a broad idea into what it includes and what it excludes; for healthy food, the worked example lists salads, grain bowls and lean proteins as in, and processed foods and raw, unprepared ingredients as out. Then the system goes looking for images that sit on the boundary of that written definition, such as a fruit smoothie that is both raw fruit and high in sugar. The user rules on each case, and an automatic loop rewrites the classifier's prompt to match, so the definition and the classifier change together.
The study covered 18 live sessions of 90 minutes each, using two concepts: healthy food and a moderation task for clickbait gaming images that promise misleading financial or digital rewards. Agile Deliberation was compared with a single written prompt, an earlier Google system that breaks down the concept without asking the user, and a manual process of spreadsheets, guessed searches and hand-edited prompts. The blog reports that its classifiers beat all three, with the largest gains on the hardest images, and that manual editing often failed to improve accuracy or made it worse. Effort and stress fell significantly, and every participant preferred it to the manual methods.
The finding about manual editing is the practical one. Many teams now write long prompts for vision-language models to sort images, then patch them one failure at a time, worrying that each fix breaks something else. Agile Deliberation formalises what good moderators already do, testing a policy against awkward cases, and lets the tool hunt for those cases instead of the person. The code, with a demo notebook for the healthy food example, sits in Google's google-research repository.
The evidence is thin in places. The blog gives no accuracy figures, only charts and descriptions, and it does not say how many distinct people took part or who they were beyond some experts. Two concepts is a narrow test of a method meant for any subjective category. The underlying models are not named, so it is unclear how much the results depend on Google's own systems. The paper has the detail, but until others try it on their own policies, it is a promising workflow rather than an established one.