How to Write an AB 2013 'High-Level Summary' Without Giving Away Trade Secrets
California’s AB 2013 requires training data transparency by 2026. Here is how to disclose without losing your IP. 🛡️
The Dilemma: Transparency vs. IP
AB 2013 mandates that developers of generative AI systems provide a "high-level summary" of the datasets used to train their models. For many companies, this data is their secret sauce. How do you comply with the law without handing your competitors a blueprint of your proprietary advantages?
What Counts as a "High-Level Summary"?
The law does not require you to list every single file or URL. Instead, it asks for categories and sources. This distinction is your shield.
Do This: "Training data includes licensed stock photography, public domain literature (Project Gutenberg), and proprietary synthetic datasets focusing on medical imaging."
Don't Do This: "We scraped the following 50,000 URLs..." or "We used the specific dataset located at [link]."
Strategies for Protection
Best Practices
- Aggregate Sources: Group data by type (e.g., "Public Social Media Posts") rather than specific platforms if possible.
- Focus on Provenance: Emphasize that you have the right to use the data (e.g., "Licensed," "Public Domain") rather than the exact content.
- Synthetic Data: Highlight the use of synthetic data. It shows innovation and is naturally opaque to reverse-engineering.
The "Trade Secret" Exception
AB 2013 does have provisions that respect trade secrets, but relying on this exception is risky. It is better to provide a compliant, broad summary than to refuse disclosure and invite an audit. The goal is to be helpful to the consumer and regulator without being helpful to your competitors.
Conclusion
Drafting your AB 2013 summary is an exercise in precision. Be honest, be compliant, but be vague enough to protect your intellectual property.