How we safeguard creator rights while delivering on the promise of AI
At that same time, and at this conference, we have to ask “Can we safeguard creator rights while still delivering on the promise of AI?”
I think we can.
Let’s look at patents and copyrights, both highly relevant to AI.
First, patents. The patent system today is working through important issues.
Google has been working in AI for many years, and we have the largest AI patent portfolio, with some of the most foundational patents on AI technologies.
Of course, whenever there’s a new technology, people rush to try to patent old ways of working in this new setting. Just as we saw with first computers, and then again with the internet, we’ve seen a dramatic surge in global AI patent applications.
And this time, Generative AI tools have compounded the challenge, by making it easier than ever before to write an application.
As a result, according to the World Intellectual Property Organization, more new patents were published in 2024 and 2025 than in the prior ten years combined.
We’re seeing claims across all aspects of AI technology, including model architectures themselves as well as novel applications of those models in different fields.
Of course not all of those claims may be valid, and we need to ensure we have tools in place to evaluate claim quality, including using AI to analyze prior art and deconstruct claims generated by AI.
If LLMs are turning three bullet points into patent applications, patent offices may need LLMs to turn patent applications back into three bullet points.
But at a high level, the patent system is supporting the development and deployment of AI tools.
There’s no need to tear down an architecture that is working, but there is a need to work together to adapt to a moment of rapid change.
When we turn to copyright, the conversation becomes more complex. One key insight is that traditional copyright has always focused on outputs, not inputs.
For our part, we are prioritizing safeguards for AI outputs as we aim to protect intellectual property while preserving freedom for creative expression.
Our approach includes a range of efforts, from deploying advanced, modality-specific filters that help to prevent AI models from exactly replicating content that might be in the training data, to actively removing infringing material via notice-and-removal systems.
And we believe we don’t have to reinvent the wheel: Existing copyright principles are robust.
AI is a tool that assists creation. If a user creates an infringing output, it’s infringing regardless of the technology used to create it.
It doesn’t matter whether a work is created with a pencil or a typewriter, a personal computer or an AI tool.
The legal standard remains.
Just as it does with traditional creative tools, the law should focus on how an AI tool is used, while recognizing the transformative nature of the technology itself.
When it comes to liability, the law has always sought to draw a clear line between the tool and how someone chooses to use it.
AI training is learning to recognize patterns
In evaluating the training necessary to create AI models, we can likewise draw on our laws governing the creation of words and images inspired by prior works.
Human creativity has always drawn on what came before.
If students go to a public library, read the books on the shelves, learn how to create plot twists, and then go home to write their own original novels, they have not infringed copyright. They have used those works to learn the craft, the art of how words and passages typically relate to one another. Generative AI training works in an analytically identical way, recognizing patterns in what’s come before.
A legal regime that required developers to get a commercial license for every piece of publicly available data used to train a model would end AI innovation.
The better approach — similar to those pioneered by Singapore, Japan, and the EU — is to have clear rules around AI training by having clear text and data mining exceptions for training on publicly accessible content. And the courts in India have just followed this approach as well.
The importance of reasonable opt-outs
Balanced regimes also recognize the unique scale and nature of content on the internet, and give publishers and creators the ability to opt out from having their content used to train or ground a model’s output.
A balanced copyright framework, with clear text and data mining exceptions, does not preclude commercial negotiations between AI developers and rights holders for access to content — in fact by establishing clear rules of the road, it facilitates those negotiations.
While Google believes that both training a model and grounding it to improve accuracy are transformative uses, we are also engaging with the ecosystem to explore new types of partnership and value-exchange models.
And we have implemented various rights to opt out, including through controls like Google-Extended, which gives rights holders the ability to say, “I choose not to participate in this ecosystem;” Long-established international protocols like robots.txt, which allow creators to decide whether they want their content to be used for training; and our Updated Search Console protocols, which let website owners manage how their links and content appear in generative AI Search features.
Having a broad right to train and ground, coupled with a machine-readable right to opt-out, offers the reasonable middle way, allowing the benefits of cutting-edge AI while protecting the rights of copyright holders.
The role the private and public sectors can play in responding to deceptive digital replicas
Before I close, I want to also touch briefly on an issue that is often raised in conversations around intellectual property and AI: deceptive digital replicas — the unauthorized generation of misleading AI deepfakes of an individual’s voice, face, or likeness.
Using an AI tool to create these replicas could harm the reputation of the person or deceive the audience.
We believe that both industry and regulators have a role to play in safeguarding against deceptive digital replicas.
Industry has a responsibility to build technical guardrails that prevent misuse of our tools and foster trust.
For example Google pioneered the industry-leading SynthID tool, which embeds imperceptible watermarks directly into AI-generated images, audio, text, or video, reducing the risk of deception about who created a particular material. We also developed advanced Likeness detection tools on YouTube, scanning our system to identify videos that potentially contain the face of creators.
