Loading page…
Google’s moderation family includes separate text and image models. Pick the modality before comparing capabilities.
ShieldGemma is useful to investigate when an application needs policy-based content checks. The family name covers different inputs, so a directory entry that simply says ‘multimodal safety’ would conceal an important deployment choice.
Google documents ShieldGemma 1 on Gemma 2 in 2B, 9B and 27B sizes, and ShieldGemma 2 as a 4B image-safety model. The documentation links separate model cards for their policies and evaluations.
Source: Google · ShieldGemma family documentationThe family evaluates content against defined safety policies and provides open weights for adaptation. Inspect the selected model’s policy and input format before assuming it matches your product rules.
Source: Google · ShieldGemma family documentationOur recommendation: if a product accepts an image and caption together, test risks arising from that combination. A successful image check and a successful text check do not by themselves show that the combined meaning was evaluated.
A community platform allows illustrated educational posts. Evaluate image moderation on diagrams as well as photographs, and review the accompanying explanation before treating a flag as the final publishing decision.
This family overview does not claim the text model accepts images, the image model verifies factual claims, or any one size is the best choice for a workload.
How to read an AI safety evaluationThe sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
Separates ShieldGemma 1 text moderation from ShieldGemma 2 image classification, with links to their model cards.
Pairs safe prompts with unsafe contrasts to investigate unnecessary refusals. Historical model results are not current rankings.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence