I update my app's privacy policy on the fly using AI based on what its code does

"Continuous integration" for GDPR

I'm building a customer-facing web app for a European audience and GDPR requires you to be transparent about all the ways your personally identifying data is processed.

Passing any personally identifying information on even to GDPR-compliant third parties - like a newsletter provider, an analytics provider, or even a backup datacenter - needs a mention in the privacy policy.

Passing PII to non-compliant third parties outside the EU needs authorization from the user (which is why it's best to avoid this if possible).

That becomes insanely complicated really quickly if you take it seriously.

For this project, I set out to do exactly that - and to get AI to help with the process. I built an experimental AI-driven development workflow that contains a new step:

• Planning the feature
• Generating and checking the new code
• Security-testing the new code
• Building tests for the new code

... and now also updating the privacy policy, if the change calls for it, automatically, against a set of well-drafted guidelines.

The goal: when a feature goes live, the updated privacy policy ships with it, rather than becoming a to-do for later... because we all know what happens to those.

Claude's frontier models Opus and Fable are fantastic at this and have helped me catch subtle omissions in new and existing code that might have easily slipped through human hands. Not unlike a security review - but for data protection.

For example, when using an external error monitoring service like Sentry, then an error report, unknown to me, by default transmits the entire request body of the page the error occurred on... including user inputs... which of course can be Personally Identifying Information! You can turn this behaviour off - or you need to document it properly.

As always, the AI output obviously requires close oversight. Given the steep potential fines imposed by data protection bodies, letting nonsense leak into the privacy policy could be more dangerous than most bugs!

And you'll need to constantly adjust the guardrails and rules the AI operates on - it no human sense of what is too much detail and needs to be told all the time to pipe down and instead of describing several dozen processes in epic detail focus on what *really* matters in a privacy context to a reasonable user. 🙄

But overall, the results are... pretty nice.

And, I would argue, of far better quality than a solopreneur or small team typically can come up with in this field.

You'll be able to see it in action in the web app soon, watch this space!