30 Days Gen AI Risk Trial -Start Now
Skip to main content
Chatbots·Free; Google AI Pro $19.99/mo; AI Ultra $249.99/mo; Gemini for Workspace bundled in Business/Enterprise SKUs·gemini.google.com

Google Gemini

Google's multimodal AI assistant across gemini.google.com, the Gemini app, Workspace, and Android, spanning text, image, audio, video, code, and deep research.

Risk Score
Medium
5/10

Independent assessment across data handling, compliance, security and transparency.

Overview

Google Gemini is the consumer- and enterprise-facing interface to Google's Gemini model family. It runs on the web, iOS, Android, inside Workspace apps (Docs, Gmail, Sheets, Meet), and through Google AI Pro and AI Ultra subscriptions, with Veo for video and Imagen for images. Consumer Gemini activity may be reviewed by humans and used to improve Google's public models, while Gemini in Workspace Business/Enterprise and Vertex AI are contractually excluded from training and carry SOC 1/2/3, ISO 27001/27017/27018/27701/42001, HIPAA, and FedRAMP High coverage depending on edition.

Risk by subscription tier

2

The same vendor often carries very different risk depending on the plan. Free tiers typically allow training on prompts, paid tiers usually do not.

PlanRisk
{"PRO"7/10·High
{"FREE"8/10·Critical

Risk factors

2
  • Consumer data may be used for model training
  • Privacy concerns with human review of data

Recommendations

2
  • Evaluate enterprise usage terms
  • Implement data access controls

Data handling

Storage
Hosted in Google Cloud globally; Workspace and Vertex AI editions support regional data residency (US, EU, and others) and CMEK.
Retention
Consumer Gemini Apps Activity defaults to 18 months; disabling it still keeps conversations up to 72 hours. Workspace and Vertex AI retention is governed by the customer's admin settings.
Training on inputs
Consumer Gemini conversations may be used to train Google's public models and reviewed by humans when Activity is on. Gemini for Workspace (Business/Enterprise) and Vertex AI prompts/outputs are not used to train foundation models.