Is Your AI 'Substantially Modified'? The AB 2013 Trigger Points
Updating your model in 2025? You might have just triggered a mandatory public disclosure. ⚙️
The Trigger
AB 2013 requires transparency for any Generative AI system released on or after January 1, 2026. But it also applies to existing systems if they are "substantially modified" after that date. This creates a "compliance trap" for companies that think their legacy models are safe.
What Counts as Substantial?
The law defines a substantial modification as a new version, release, or update that materially changes the functionality or performance of the system.
- Retraining: If you retrain your model on a new dataset (e.g., adding 2025 medical records), this is almost certainly a substantial modification.
- Fine-Tuning: Fine-tuning a base model (like Llama 3) with RLHF (Reinforcement Learning from Human Feedback) to change its behavior or safety guardrails is likely substantial.
- Architecture Changes: Changing the model architecture (e.g., moving from GPT-3.5 to GPT-4 level complexity) is substantial.
What Does Not Count?
Minor bug fixes, security patches, or UI updates that do not affect the underlying AI model's logic or training data generally do not trigger the requirement.
The Documentation Burden
Every time you trigger this clause, you need to update your transparency documentation. This means your compliance team needs to be in sync with your engineering team's release cycle. "Continuous deployment" now means "continuous compliance."
Conclusion
Track your model updates carefully. Maintain a version history that maps specific model versions to specific training datasets. If you can't prove which data trained which version, you are in trouble.
Frequently Asked Questions (FAQ)
Does adding RAG (Retrieval Augmented Generation) count?
This is a gray area. RAG changes the output by providing new context, but it doesn't necessarily change the model weights. However, if the RAG system fundamentally changes the tool's capability (e.g., allowing it to diagnose a new condition), it might be seen as a substantial modification of the "system."
Do I need to keep old disclosures?
Yes. You should maintain an archive of past disclosures so that users (and regulators) can see what data was used for previous versions of your tool.
Who decides if it's "substantial"?
Ultimately, a court or the Attorney General. But you should document your own internal rationale for why an update was or was not substantial. This "good faith" effort can be a defense.