Google’s November 4, 2026 explanation of Agile Deliberation tackles a problem for editors, curators and teams selecting images: an AI filter may struggle because its human owner has not fully defined what belongs. The research explores how borderline examples can help people clarify those rules before asking a model to apply them.
Here, an image filter means a system that selects or categorizes pictures, not a photo effect. This is also not a newly launched Google service. The paper first appeared in December 2025 and was revised in April 2026; November’s blog explains that earlier work.
Make the boundary visible
Agile Deliberation works in two stages. First, the person develops a structured definition, including what should count and what should be excluded. Then the system brings forward borderline images for review and uses the person’s feedback to revise the classification instructions. Google describes this as a way to develop the definition itself, rather than assume the first instruction already captures it.
Consider an illustrative example, not a study finding: an editor wants pictures of outdoor dining. Does a street scene qualify when a café table appears only in the background? One editor might accept it; another might require dining to be the main subject. A rule that simply says “outdoor dining” leaves that choice unstated. Recording the distinction gives both the reviewers and the model a clearer standard.
The practical implication is to separate two questions. If reviewers disagree about whether an image belongs, the policy may need clarification. If they agree on a written rule but the model fails to follow it, that points to an execution problem. A model comparison alone cannot settle the first disagreement.
More precise selections can mean more missed matches
The study involved nine internally recruited participants without professional moderation experience, completing 18 sessions across two concepts. Sessions lasted about 90 minutes, and evaluation used 200 held-out images per concept. These were 18 sessions, not 18 independent participants.
One task concerned images promising unrealistic rewards for online entertainment. The researchers’ reported mean precision—the share of flagged images matching a participant’s definition—rose from 35% to 57%. But recall, the share of intended matches caught, fell from 80% to 63%. Both comparisons use the same group’s initial-prompt baseline.
That tradeoff matters to an editor: a cleaner selection can still leave out more wanted images. Which mistake is more costly depends on the collection’s purpose; a better combined score cannot make that choice.
The authors acknowledge the small study’s limits. They evaluated the combined workflow, not each component separately, and did not test the full pipeline across other models. The findings do not establish universal accuracy or objectively correct judgments about subjective categories.
Clarify the policy before adopting the software
A useful first step is to write down what belongs, what does not and why borderline images remain unresolved. For a team, that could be a shared rule with examples and the reasoning behind each decision. If a pilot follows, keep separate examples for checking it, and record unwanted inclusions and missed matches separately.
Adopting the code is a separate decision. The available implementation is a Jupyter/Colab research prototype, explicitly not an officially supported Google product. It needs an image-retrieval index and model calls; its documentation warns about variable waiting times and rate limits. The documentation gives no fixed service price. Teams considering a pilot would need to establish their own setup, model-usage and human-review costs.
Editors should also keep selection separate from provenance: whether a picture fits a collection is a different question from how it was made. RohitAI’s guide to checking media with SynthID explains what watermark checks can and cannot establish.
The broader implication is about responsibility. A classifier can become more faithful to one person’s preferences without making those preferences appropriate for everyone affected. Choosing whose judgment governs, and who reviews unresolved cases, remains a human decision.