A Great Hollowing

It's Research Roundup time again! This week let's explore how changes in our life experiences change our life outcomes, for bad and for good.

Is "Early Career" Gone?

Erik Brynjolfsson and colleagues published one of the first papers showing the LMMs lift productivity of lower-skilled workers while having little impact on higher-skilled, a finding replicated up and down the “skills ladder”. Now Erik et al. are back…but the story has gone from lifting to hollowing out.

You may have seen the news on this study; it’s still making the rounds: since ChatGPT's rise began, employment for “early-career workers (ages 22-25) in the most AI-exposed occupations” has dropped a stunning “13% relative to” their peers.

Experienced workers in those same jobs are fine. It's the entry-level roles—ones you learn by doing, making mistakes, building a foundation, and eventually becoming an experienced worker—that are being automated away.

The standard approach to modeling technologies impact on labor is to model it all as a series of tasks. But if these early findings hold, we're not just automating away tasks; we're automating away the first rung of the career ladder. By choosing automation over augmentation, we risk pulling up the ladder behind us, creating a generation of young talent with no clear path to entry.

While these numbers need some validation and replication, I have some initial questions:

1. What is the long-term impact of this potential hollowing out of the first job? First job placement is a significant predictor of life outcomes, from the obvious—life wages—to the unexpected—divorce and suicide risk.

2. Why automation rather than deprofessionalization? Given the original finding on productivity lift, one might expect an employment shift to lower-skilled workers. Perhaps these most automatable jobs don’t have a deprofessionalized middle-ground to fall back on. Or, perhaps management is simply showing its profound distaste for uncertainty and full automation is just a cleaner management story.

3. The methodology is intriguing but unproven: They “distinguish between automation and augmentation empirically using estimates of the extent to which observed queries to Claude, the LLM, substitute or complement for the tasks in that occupation”. This approach is wonderfully empirical and naturalistic; it’s also open to much interpretation. Are early- vs late-career individuals more likely to submit substitution-type queries (self-automators) for the same challenge? There might be a great deal of heterogeneity at the individual worker level rather than at the occupation level.

For everyone entering this chaotic workforce (and their parents) you must be wondering, “How do I robot-proof myself or my kids?” Stay tuned—Robot-Proof is out next year…if there’s anyone left to read it.

Corn Syrup for the Soul

I recently asked the question, “Is GPT the new GPS?” implying that both heightened dementia risk. (And boy did some people get angry.) But my core point remains: it is how you use technology that matters. Take for example, books.

A new analysis of the American Time Use Survey found that “reading for pleasure” has declined by a staggering “3% per year”. That’s a compounded collapse in an activity so strongly tied to cognitive development and life outcomes.

It is possible that people are shifting to equivalent alternatives—perhaps organizing community action of world hunger—but I don’t see too much of that when I glance around the subway car. It seems like a whole lota TikTok and whatever the current version of Dopamine Game Yaya Fun Time! is called (probably no longer minesweeper or Candy Crush).

Unsurprisingly, the decline is steepest for lower-income and less-educated communities, creating downward spiraling feedback loops that widens the opportunity gap before a career even begins.

A society that stops reading doesn't just lose a pastime; it loses its capacity for deep thought; its simplest lesson on the value of patience; its opportunity to buy my next book. (Who am I kidding—“pleasure” has nothing to do with my books.)

Go Deep-->Go Long

Everyone worries about cognitive decline, and for good reason: copious research shows raw cognitive ability begins declining by age 30. But functional cognition may not have such a clear expiration date.

An analysis of literacy and numeracy skills of thousands of adults over several years found that “skills decline at older ages only for those with below-average skill usage”.

Functional cognition increased in the the 40s for most adults, “before decreasing slightly in literacy and more strongly in numeracy”.* For workers in cognitively demanding jobs who consistently exercised their mental muscles, skills didn't just hold steady—they continued to increase well beyond their 40s. Age wasn't the driver; engagement was.

AI may well be disrupting entry-level work and declining reading habits are almost certainly eroding cognitive foundations, but the experiences that shape our lives aren't just the ones that happen to us; they are the ones we actively choose. Go deep..

“ Curious and worrying sidenote: Women have larger skill losses at older age, particularly in numeracy.”

Vivienne L'Ecuyer Ming