Single-user clusters use a different security mode which is the reason for this difference.

On single-user/assigned clusters, you'll need the Fine Grained Access Control service (which is a Serverless service) - that is the solution to this problem (also allows you to read tables with RLS/CLM, Dynamic Views and DLT's Streaming Tables and materialized Views - all stuff that is problematic to read on single-user(ie assigned) clusters today)

So, assuming you need Shared ML you need:

1. Service Principal Cluster Preview

2. Fine Grained Access Control Preview.

You can talk to your Databricks Account team to enable these.