Google AI Research's Agile Deliberation: a diverse team reviews complex image classification data on screens.

Editorial illustration for Google AI Research debuts Agile Deliberation for gray-area image classifiers

Google AI Research debuts Agile Deliberation for...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

November 5, 2026 • 3 min read

Google Research has put out a new open framework called Agile Deliberation, built for a problem that trips up most image classifiers: concepts nobody can quite define. "Healthy food," "inspiring artwork," clickbait thumbnails. These aren't categories with clean edges. They're judgment calls, and judgment calls are exactly what standard labeling tools assume away.

Most classifier-building workflows start from the premise that a user already knows where the boundary sits and just needs help labeling examples on either side. Agile Deliberation starts from the opposite premise. It assumes the boundary itself is the thing in question, and builds a process around finding it rather than applying it.

The framework runs on Gemini 2.5 Pro and Gemini 2.5 Flash through the API, no fine-tuning involved, inside a Jupyter or Colab notebook. Retrieval leans on CLIP embeddings and a FAISS index, starting around 260 MB for the demo setup. Google's team says the approach outpaces zero-shot Gemini 2.5 Flash by 10.5 F1 points on average across test concepts and 18 live sessions, with the gains and tradeoffs varying sharply depending on what's being classified.

Think ‘healthy food’, ‘inspiring artwork’ or clickbait in content moderation. Most tooling assumes users already know the concept boundary. Agile Deliberation assumes they do not, and helps them find it.

Google AI Research Releases Agile Deliberation Framework for Gray-Area Image Classifiers - MarkTechPost

Why this matters

Anyone who has shipped a content moderation system knows the real cost isn't the obvious violations, it's the thousands of borderline cases where reasonable people disagree. Google's approach is notable for what it doesn't do: it skips fine-tuning and instead prompts Gemini 2.5 Pro and Flash to actively surface the disagreements, rather than waiting for them to surface in production and blow up your review queue. Running this out of a Colab notebook also lowers the barrier for teams that want to prototype a policy before committing engineering resources to a full pipeline.

That said, we'd flag the obvious tension here: a prompted generative VLM with uncalibrated confidence scores is being used to find the exact cases where calibration matters most. If Gemini's sense of "borderline" doesn't match your actual edge cases, you could end up with a tidy set of examples that teaches you nothing about your real failure modes. Worth testing against known hard cases before trusting it to define your gray areas for you. We'll be watching whether this gets adopted for policy work beyond Google's own examples.