RESPONSIBILITY & SAFETY
For more than a decade, AI has played a critical role in how we deliver safe and responsible experiences across all of our products. As AI continues to advance, it introduces exciting new capabilities, but also risks – making our responsibility to keep you safe online more important than ever. We’re committed to addressing these risks so that we can maximize AI’s benefits for people and society.
We rigorously test our models and infrastructure at every layer of the stack, combining the best of AI with our world class teams of safety experts. This end-to-end approach enables advanced AI experiences that put safety first.
Core to our security strategy is a technique called automated red teaming (ART), where our internal Gemini team constantly attacks Gemini in realistic ways to uncover potential security weaknesses in the model. This technique has helped significantly increase Gemini’s protection rate against indirect prompt injection attacks during tool-use, making Gemini 2.5 our most secure model family to date.
Our safety and security teams provide a follow-the-sun model for 24/7/365 monitoring of all Google products, services, and infrastructure.
Through our Bug Bounty program, we collaborate with and incentivize the security research community across 68 countries to identify and address vulnerabilities in our generative AI products. In 2023 alone we awarded $10 million to more than 600 researchers who contributed to the safety and security of our products.
While AI is an important tool in identifying potentially violative content at scale, human review plays a critical role in the process. We fight abuse and continually improve our models with input from independent experts as well as over 25,000 human reviewers from a range of disciplines. Together, they augment automatic safety systems to address more nuanced cases.
Being able to identify when something is AI-generated is a critical component of trusting content and information. We proactively watermark synthetic media generated by Google’s AI products and provide built-in tools so you can easily evaluate the accuracy of information.
SynthID uses a variety of deep learning models and algorithms to embed imperceptible watermarks directly into any image, audio, text, and video generated with Google’s AI tools. This helps systems identify if content, or part of it, is AI-generated by Google’s AI tools. The SynthID Detector enables quick and efficient identification of AI-generated content made with Google AI. The portal provides detection capabilities across different modalities in one place and provides essential transparency in the rapidly evolving landscape of generative media. It can also highlight which parts of the content are more likely to have been watermarked with SynthID.
Through our SynthID toolkit and features like YouTube labels, we help people and organizations responsibly create and identify AI-generated content. As an active member of the Coalition for Content Provenance and Authenticity (C2PA), we collaborate with industry partners to build and implement a standard that improves the transparency of digital media.
Built-in tools across Gemini, such as Double-Check response, and About this image in Search, make it easy to double check information and access helpful context, including original content sources. For example, About this image can indicate whether an image was generated using Google's AI tools when you come across it in Search or Chrome.
To set youth up for success in an AI-first future, we’re building AI experiences tailored to the unique needs of younger users. For example, we implemented more stringent content policies that work to prevent age-inappropriate responses and partnered with experts to create AI literacy education to help younger users access AI responsibly.
As we advance the future of generative AI, we leverage the same industry-leading security infrastructure that protects billions of users across all of our products. We strictly uphold responsible data practices, put you in control of your information, and are actively implementing privacy safeguards tailored to the unique needs of our AI products.
Today, and for decades, the main challenge in cybersecurity has been that attackers need just one successful, novel threat to break through the best defenses. Defenders, meanwhile, need to deploy the best defenses at all times. With AI, we are reversing this dynamic by enabling security professionals and defenders to scale their work.
We've pioneered state-of-the-art techniques and solutions that protect your information. For instance, Private Compute Core processes on-device data privately and we leverage Federated Learning as part of a suite of steps we take to protect privacy when we train our AI models. And because we rely only on our own AI models and data servers, your data is always within Google's secure data center architecture, unless you request otherwise. We’re committed to providing users with transparency through our Gemini Apps Privacy Hub, and independent third-party certifications regularly assess Google's security and privacy practices.
As we push the limits to deliver even more helpful AI products, we work to ensure that privacy safeguards are baked in from the start. For example, when developing Circle to Search, we used our privacy principles as the north star for decision-making and worked with internal and external privacy experts to design a new way to search on Android that puts trust and safety at the core of the experience.
Building a safe AI future is a collective effort. We partner with NGOs, industry partners, academics, experts, ethicists, and more at every stage of the product development phase. And to drive a safer AI ecosystem, we openly share our expertise and tools with partners, organizations, and competitors around the world to ensure that everyone benefits from the advances we're making in responsible AI.
This industry-leading set of protocols developed by Google DeepMind proactively identifies future AI capabilities that could cause severe harm and puts in place mechanisms to detect and mitigate them. The Framework will evolve as we deepen our understanding of AI risks and collaborate with industry, academia, and government.
Our industry-leading SAIF Framework provides security practitioners with guidance for integrating security measures into ML-powered applications for AI/ML model risk management. We also helped launch the Coalition for Secure AI with other industry partners to establish best practices, standards, and open-source tools in the industry, and are planning to donate our SAIF risk taxonomy and assessment to further democratize the way Google keeps its own AI systems safe and secure.
Our Responsible Generative AI Toolkit provides guidance and essential tools for creating safer AI applications with Gemma. For example, the Toolkit includes a section on model Safeguards, offering developers a series of safety classifiers to filter the input and outputs of their applications and protect users from undesirable outcomes.
Sharing information about responsible AI practices benefits the field. We first published research on model cards, which are like food nutrition labels that describe essential facts on AI models, in 2018. Since 2019, we’ve published an annual report on our progress implementing our AI Principles. The most recent paper details our lifecycle approach to building safe and responsible AI.