Introduction

No. Google does not use Customer Data to train or fine-tune any AI/ML models without prior permission or instruction, as outlined in Section 18 “Training Restriction” in the Google Cloud Service Specific Terms. This applies to all managed models on Gemini the platform, including GA and pre-GA models.  

Your data remains completely isolated within your Google Cloud project and is processed under your control, in accordance with the Google Cloud Terms of Service and Cloud Data Processing Addendum (CDPA).

For more information on how prompts are handled, see our overview of How Gemini products in Google Cloud use your data.

No. The training restriction in Section 18 of the Google Cloud Service Specific Terms is active by default for all enterprise Google Cloud and Workspace customers. Accordingly, no manual opt-out is necessary.

Google's foundational AI/ML privacy commitments mandate strict security and customer control over data, extending across all Google Cloud generative AI products. In addition to internal data governance practices, these controls are independently audited and verified under our ISO/IEC 42001 certification, complementing our existing ISO/IEC 27001, 27017, 27018, and SOC 1/2/3 certifications. Customers can review the exact scope of covered services and download current certificates and audit reports through the Compliance Reports Manager.

Customer Data is retained in Gemini the platform for limited periods of time and under certain scenarios and conditions.  

For more information on achieving zero data retention within specific Google Cloud generative AI products, see guidance on actions Customers must take.

We believe in a shared fate with our customers and recognize that responsible AI requires an ecosystem approach. We offer enterprises protection with generative AI copyright indemnification, and we don’t charge egress fees, so that customers are free to pick the most responsible provider without fear of being locked-in to models or cloud providers. We have also developed Responsible AI tooling, enablement and support so customers can tailor their own risk and safety posture for each use case and each deployment.

To provide logical isolation, the Gemini platform leverages Google Cloud’s secure-by-default infrastructure alongside robust Access and Identity Management (IAM) controls. Customers retain full ownership and control over prompts, grounding datasets, fine-tuning data, and generated outputs, ensuring your assets are never used for public model training. Operationalizing this protection is simple, as organizations can configure security controls such as Customer-Managed Encryption Keys (CMEK), VPC Service Controls, Client-Side Encryption (CSE), and Zero Data Retention configurations to block unauthorized access.

Responsible AI is essential for generative AI. Google remains a leader in responsible AI, guided by our AI principles that uphold our commitment. To help organizations build and use AI responsibly, we offer a comprehensive suite of generative AI tools focused on helping them test and understand models, and reduce the risk of negative or harmful outcomes, if any.

Google employs a multi-layered approach to AI governance that combines human expertise, user feedback, and automated systems that help scale our work to manage risk across the model development lifecycle. To learn more about our specific governance frameworks, policies, research-driven methodologies, and our ongoing commitments, please refer to our annually updated Google Responsible AI Progress Report.

The Secure AI Framework (SAIF) is Google's conceptual framework designed to ensure AI systems are secure-by-design throughout their lifecycles. It is structured around six core elements: expanding strong security foundations, extending detection and response, automating defenses, harmonizing platform-level controls, adapting controls via fast feedback loops, and contextualizing surrounding business risks. Customers can implement SAIF on Google Cloud by leveraging secure-by-default platform protections while configuring custom controls across their data, infrastructure, models, and applications. This unified approach, which now extends to managing emerging agentic AI risks, allows organizations to safely deploy and scale trustworthy AI solutions.

Read more to understand how to implement Secure AI Framework (SAIF) controls on Google Cloud.

Building multi-agent architectures requires strict defense in depth, treating every LLM and tool as a potential vector for a range of security threats, including but not limited to prompt injection or data exfiltration. 

We treat AI agents as digital workers that require the same accountability, oversight, and boundaries as any other member of the workforce. We assign every agent a unique machine identity to enable granular permissions, and provide built-in observability and a centralized registry. This layered approach gives leadership total visibility and absolute administrative control over their AI portfolio, ensuring that as they scale, their agents remain predictable, compliant, and always aligned with corporate standards.

To safely transition autonomous agents from isolated pilots to production fleets, we provide built-in sandboxing and unified policy enforcement via Agent Compute Engine and expanded agent governance capabilities. This framework delivers silicon-rooted hardware isolation (via microVMs), central policy controls (via Agent Gateway and Agent Registry), and standardized OpenTelemetry-compliant tracing to keep your agent portfolios entirely secure, predictable, and auditable.

Read more on how Google Deepmind is Securing the Future of AI Agents and Google’s Policy Approach to Agentic AI.

Google is committed to complying with applicable legal obligations across all jurisdictions in which we operate, including the EU AI Act. Google signed the European Union’s General Purpose AI Code of Practice in 2025 and signed the Transparency Code of Practice for AI-generated content (Article 50) in 2026. We actively track regulatory milestones, harmonized European standards, and implementation guidelines as they are phased in. Google has taken appropriate steps to comply with the EU AI Act, such as carrying out risk assessments and mitigations for Gemini 3 and documenting those in its Frontier Safety Framework and Model Report.

Accordingly, while Google Cloud ensures its foundational models and services comply with applicable laws, customers are responsible for conducting their own assessments to ensure their specific deployment and use cases remain compliant.

For an overview of readiness, visit our EU AI Act Compliance Card and EU AI Act Blog.

Google provides comprehensive, mandatory AI literacy training to our full-time employees. This role-tailored curriculum ensures that our product, engineering, and support teams possess a strong understanding of core AI concepts, security baselines, and safety practices.

When customers deploy Google Cloud AI models and systems, they are responsible for ensuring that they use the services in compliance with applicable laws, including the EU AI Act and Article 5 (Prohibited AI Practices). Please also refer to the agreement under which you access Google Cloud Services and in particular the Google Cloud Platform Acceptable Use Policy (AUP) and the Generative AI Prohibited Use Policy (PUP).

To help customers evaluate and understand model capabilities, Google provides comprehensive transparency assets including reports for EU AI Act and CA AB 2013:

• Model Cards: We publish comprehensive, technical Model Cards for our foundation models provided on Gemini platform outlining their capabilities, limitations, safety evaluations, and recommended use cases. For deeper foundational details, we point customers to the comprehensive research papers and technical reports published by Google DeepMind.
• Training Data Summaries: High-level summaries of datasets and sourcing methodologies used to train our general-purpose foundation models (including the Gemini model family) are made available through the Google Transparency Report Center.

We enforce rigorous internal checkpoint governance, tracking foundation model versions from pre-training through fine-tuning and deployment. This model lineage information is detailed in our Model Cards and reports, helping enterprise CISOs and risk teams maintain a clear, auditable inventory of their deployed AI assets.

Google Cloud’s infrastructure, platform controls, and AI services undergo rigorous independent third-party audits to validate security, privacy, and governance. To find out if a product has a compliance certification, search for the product name in the security pages; for example, the following:

• FedRAMP
• ISO/IEC 42001
• ISO/IEC 27001
• ISO/IEC 27017
• ISO/IEC 27018
• ISO/IEC 27701
• SOC 1
• SOC 2
• SOC 3
• PCI DSS
• BSI C5:2020
• HIPAA

We want to make it as easy as possible for you to use your choice of models on our platform, while keeping the boundaries clear:

• Model Choice: Google Cloud has always been defined by choice, enterprise-grade governance, and market-leading price-performance. Key to the foundation of openness and choice is giving customers flexibility to run agentic workloads across their choice of models and infrastructure to achieve the right balance of cost, capability, and performance. Google Cloud offers access to all three leading closed source LLMs on a unified enterprise AI platform.
• Partner & Open Model Documentation: Customers can review the respective third-party model cards, technical documentation, and evaluations directly within Model Garden to conduct their own evaluations.
• Deployment Choices for Claude on Model Garden AI: To provide maximum deployment choice and flexibility, customers have two distinct options for Anthropic’s Claude models to choose from on Google Cloud:
• Google-managed: Delivers latest Claude models with Google Cloud operating as the data processor under your existing customer agreements, providing VPC Service Controls, Customer-Managed Encryption Keys (CMEK), and native integrations with services like BigQuery and Model Armor.
• Anthropic-managed: Delivers day-zero access to new Claude Platform features (such as Claude Managed Agents and code execution). Anthropic operates the stack as the data processor under Anthropic’s commercial terms, while you continue to benefit from Google Cloud billing, IAM authentication, and Cloud Audit Logging.

This question is best answered by your organization’s legal team, given it involves determining the company’s specific roles and related compliance under the Act. When customers deploy Google Cloud AI models and systems, please refer to the agreement under which you access Google Cloud Services and in particular the Google Cloud Platform Acceptable Use Policy (AUP) and the Generative AI Prohibited Use Policy (PUP).

Google applies SynthID watermarking across multiple modalities of Google models and C2PA Content Credentials are cryptographically signed and handled where applicable at the platform level across Google Cloud services. 

For technical details, see the Google DeepMind SynthID Overview, the technical description of the algorithm, and the Google AI Responsible Safeguards Guide.

In 2025, we launched the SynthID Detector, a verification portal, to verify if a piece of content was watermarked with Google SynthID. After testing with journalists and media professionals and collecting their feedback, we’re expanding access to everyone. On October 7, 2026 Google announced the new SynthID Detector tool was available globally in English to check if an image, video or audio file was made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple. The SynthID Detector checks uploaded media for a SynthID watermark. For more information on SynthID detection or to try it out, you can visit synthid.com.

Read more technical details on limitations in tools for detecting LLM-generated Text and how Google is building transparency in the age generative AI.

To support enterprise requirements, Google Cloud offers an AI Content Detection API currently in private preview, available to Google Cloud customers when using our platform. We plan to expand access to the AI Content Detection API over time and are working on support for video, audio and text modalities. To request access to the endpoint, complete the application form.

C2PA is an open standard that provides cryptographically signed Content Credentials added at the platform and system layer across Google Cloud services during output packaging. It records provenance details, signing authority (Google LLC), and modification history for supported media containers.

You can review Google Cloud's active list of models that support Content Credentials to see which media generation pipelines automatically package C2PA metadata out of the box.

To verify the Content Credentials for a media file, go to the Content Credentials website. If the media file was created, edited, and published using only C2PA-aware tools, the Content Credentials are verifiable. Any modification to a C2PA-compliant media file using a non-C2PA tool is considered tampering, resulting in a validation failure, making it easy for customers to flag untrusted or altered assets.