Creative Content and AI Training: A Practical Guide

Artificial intelligence is transforming our world at a breakneck pace, touching everything from our daily lives to the cutting edges of science and art. As AI continues to evolve rapidly, it's crucial to consider how we can foster a balanced approach to using creative content in training AI models. This isn't just about legalities; it's about shaping the future of AI innovation and human creativity.

Every new technology that helps create or share knowledge and art—from the printing press to the internet and cable TV—has sparked debates about how to generate and distribute value. When it comes to AI, developers can take several steps to support creative industries and build a thriving AI ecosystem that benefits everyone. So, what strategies should we consider for AI model outputs, training processes, and the new ways AI can create shared value?

Evaluating AI Outputs

Whether you're writing with a pen, a typewriter, or AI, or creating art with a paintbrush, computer graphics, or AI, the key question is whether the new work infringes on the copyright of an existing one. This can be tricky, depending on factors like how similar the new work is to the old one, the nature of both works, and whether the new one competes in the same market as the original. Tools like output filters can help prevent outputs that are too similar, even as the models learn to make more nuanced judgments about these factors.

Provenance information, such as watermarks or metadata, can also help reduce the risk of misleading people about who created a piece of content. For instance, Google has led the way with its SynthID tool and is part of the steering committee for the Coalition for Content Provenance and Authenticity (C2PA). These efforts can help consumers make better-informed decisions about the content they encounter.

Training AI Models Responsibly

Training foundational AI models on content from the open web is considered a transformative fair use under U.S. copyright law, and many other countries have similar text and data mining exceptions that encourage new uses of information. However, adopting good practices can help build acceptance for new AI uses of existing content.

It's important to acquire content responsibly and legally, such as by allowing websites to opt out of having their content used for AI training. Existing industry standards for web crawling are a key way to do this. These standards are straightforward and scalable, building on well-established machine-readable robot.txt protocols that are widely used across the web to control how content is accessed by web crawlers. Nowadays, thousands of web publishers are also using the Google-Extended protocol and similar AI-specific protocols offered by other companies. AI developers should be open to evolving these standards as the ecosystem grows and should take reasonable steps to avoid improperly training general-purpose AI models in ways that bypass these standards or similar technical measures like paywalls.

When it comes to avoiding the use of individuals' voices and likenesses, legislative frameworks can build on existing "notice-and-removal" systems for copyright, including safeguards to prevent abuse. New tools can also help creators tap into AI's creative potential while maintaining control over their voice and likeness.

Sharing Value, Expanding Opportunities

AI has the potential to benefit everyone, and collaboration between AI developers and content publishers can expand the market and generate new income for creative industries.

AI developers are looking to share the value of outputs by driving related traffic to content providers. The ecosystem is also working together to find new ways to create value from emerging AI applications. For example, there may be opportunities for commercial partnerships when AI services "ground" responses on facts from websites.

AI developers and content publishers are also collaborating on new content agreements for using specialized or non-public data for training purposes. AI developers are increasingly learning how to assess the usefulness of individual content for different AI applications. For our part, Google has already entered into agreements with several publishers for broad data rights and continues to explore new opportunities.

AI developers are actively working with media and creative industries to design new generative AI tools that add value to these industries. For example, Pinpoint, an AI tool for journalists, helps reporters search through text, audio, image, and video files to spot patterns in data, identify new angles, or find a quote in a video or audio file.

AI is a shared opportunity with the potential to expand the realms of science, commerce, and creativity. We're committed to working with all stakeholders in the ecosystem to create a shared framework where both creators' rights and innovation can thrive.