Module 4: Question about the PII redaction example

Course Q&A Agentic AI
, ,
1

Hi everyone,

First of all, thanks for this course. I’ve taken several DeepLearning.AI courses and always find them really practical and clearly explained.

I had a small thought about the example in Module 4, “How to address problems you identify”. The example shows an LLM that doesn’t reliably follow instructions to redact PII, and one suggested fix is switching to a different LLM provider.

I found the example a bit surprising, because it implies sending (other people’s) personal data to an external LLM and relying on the model itself to remove it. Depending on the provider and settings, prompts may be logged or retained, so the PII has already left your control before any redaction happens. In many settings (e.g. under GDPR) that could be a real compliance issue.

I completely understand the example is about error analysis and improving a component, not about privacy best practice. But since so many people learn from these courses and may reuse the patterns, it might be worth adding a short note, for example that PII is usually better removed before it reaches an external model (e.g. with a rule-based tool or a locally hosted model), or that you’d want a provider with appropriate data-retention and processing agreements in place.

Curious whether others had the same reaction, or if I’m missing something about the intended setup!

Thanks again for the great material.

2

Hi Maud,

Good point.

I guess it depends on the trade-off between privacy and the advantages of using a leading model. Sending PII to a leading model does not have to be problematic as long as the information is not sensitive. At the same time, awareness about issues of privacy and security is extremely important.

As you indicate, the example in this course is about error analysis and improvement in the specific context of making agentic AI work. Privacy and security is a general topic to be taken into consideration with every interaction over the internet.

There is a course on quality and safety for LLM applications that addresses some of these issues here: Quality and Safety for LLM Applications - DeepLearning.AI. Another course shows how to use a local model instead of having to send information over the internet: AI Coding Workflows: From Cloud to Local - DeepLearning.AI

Thanks for your post!