It's the nationaldesignstudio/rampart PII-removal model. Not weird at all to load that browser-side if your goal is to avoid PII hitting your endpoints.

It's weird to be loading 50 MiB. This is a government website. It should be engineered to serve people on extremely old hardware or in extremely remote locations.

The other options are sending PII to the chatbot, making users install a native app or getting browser/OS vendors to ship the classifier. Or create a national Linux distro and give out physical install media to people who can't download it, that would be great but not very realistic.

What is your proposal for a low-tech way to prevent senior citizens from volunteering unnecessary personal information to a chatbot? Because there are a lot of people with living senior parents who would love to be let in on the secret.