JPMorgan Chase, Amazon Seek Patents to Make AI Unlearn Bad Habits

AI models tend to pick up bad habits from their environments, and some firms are working on ways to help systems “unlearn” them.

JPMorgan Chase and Amazon both filed patents for “unlearning” in generative models, tech that could help developers avoid having to entirely retrain models as well as strengthen AI guardrails against data security slipups. 

To start, JPMorgan Chase is seeking to patent “machine unlearning in generative models.” This  essentially feeds a model a “forget dataset,” which trains it to treat certain data points as meaningless noise. The model is also fed a “retain dataset,” to help it hold onto relevant knowledge. 

Similarly, Amazon filed an application to patent “unlearning in pre-trained generative machine learning models.” Amazon’s tech helps models unlearn specific concepts or data by feeding them “negative prompts,” or data they should forget, and “positive prompts,” or data they should remember. Instead of retraining the whole model, the system uses these prompts to create an “adapter,” or a filter that adjusts how the model responds. 

Generative models are essentially parrots: If they’re approached with the just the right prompts, it’s possible to get them to reveal the data they were trained on. These techniques are just a few of several that tech firms are eyeing to get AI’s data security problem under control. Google, Microsoft, Intel and others have all sought patents for ways to lock down AI’s training data. 

But with the rate at which enterprises are seeking to adopt AI, especially with the rise of autonomous AI agents, there’s a constant push and pull between innovation and security. To avoid slipups that can cause both financial and reputational damage, one has to come before the other.