What makes work exposed: routine, well-defined, text- or data-heavy
The clearest pattern across the research so far is that exposure tracks task structure, not job prestige or salary. A task is more exposed when it's well-defined (you could write reasonably clear instructions for it), when it's primarily about processing or producing text, code, or data, and when getting it 80% right on the first try is genuinely useful even if a human still checks the last 20%. That describes a lot of entry-level white-collar work: first-draft writing, routine coding tasks like boilerplate and test-writing, initial customer support triage, basic data cleaning and summarization, and first-pass legal or financial document review.
Notice what's on that list: a lot of it is exactly the work junior employees traditionally cut their teeth on. That's part of why the entry-level effect shows up in the data — the on-ramp tasks that used to be assigned to new hires as a way to build skill and prove themselves are the same tasks AI now does a reasonable first pass on. That doesn't mean entry-level jobs vanish, but it does mean the path into a career built on 'do the routine stuff until you're trusted with more' is genuinely getting disrupted, and that's worth taking seriously rather than waving away.