Google Calls for AI Training Exceptions and Creator Opt-Outs at Singapore IP Forum

Google is pressing policymakers to allow AI models to train on publicly accessible content under clear text-and-data-mining exceptions, paired with machine-readable ways for creators and publishers to opt out. The proposal, set out by Global Affairs President Kent Walker at Singapore IP Week, puts a legal default in favor of training while making refusal a technical choice for rights holders.

Walker’s keynote at the Global Forum on Intellectual Property argued that existing intellectual-property frameworks can adapt to AI rather than be replaced. For copyright, Google’s position is that the law should focus on infringing outputs and on how a tool is used, while recognizing AI’s transformative character.

A default permission, then a refusal signal

The mechanism Google favors starts with a broad permission to train on material that is publicly accessible. Walker compared model training to learning patterns from prior works, and said requiring a commercial license for every item of public training data would halt AI innovation. Google would pair that default with an opt-out that software can read automatically.

The controls Google points to

• Google-Extended lets rights holders choose not to have their content used for training or to ground a model’s output.
• Robots.txt is a longstanding web protocol that Google says creators can use to decide whether content may be used for training.
• Updated Search Console protocols let website owners manage how links and content appear in generative AI Search features.

Licensing remains part of the picture

The proposed exception is not a rejection of paid content deals. Walker said clearer rules could facilitate commercial negotiations between AI developers and rights holders, and that Google is exploring new partnership and value-exchange models. That places voluntary licensing alongside, rather than in place of, the company’s preferred legal rule for public web content.

A different legal lane for deepfakes

Walker drew a sharper boundary around deceptive digital replicas. Google says copyright protects original creative works rather than a person’s identity, making it the wrong tool for unauthorized AI-generated imitations. It supports U.S. measures including the NO FAKES Act of 2025 and the TAKE IT DOWN Act, while calling for comparable protections elsewhere.

Google also presented product safeguards as part of that response. SynthID embeds imperceptible watermarks in AI-generated images, audio, text, and video, while YouTube’s likeness-detection tools scan for videos that may contain creators’ faces. Those systems are meant to identify or limit misuse, but Walker’s policy argument still seeks tailored laws to address the underlying identity harms.

The policy choice inside Google’s proposal

Google’s case rests on a particular division of responsibility: copyright rules should let models learn from public material unless a rights holder signals otherwise; output infringement should remain actionable; and impersonation should be handled through identity-focused law. Whether governments adopt that division—and what counts as a workable opt-out—will determine how far the company’s middle-ground model travels beyond its own services.