benjamin verschuere @benjamin_versch

Thanks for sharing our work! Here is the dynamic in heatmap form: within AI-adopting industries, AI hits lower-skilled jobs at every age, and the young even in skilled jobs. The shielded are experienced workers in higher-skilled, higher-paid roles, where AI complements rather than substitutes. Interactive version: liminal-capital.com/apps/ai-amdahl…

Super interesting! "AI and (Not Only) the Young: Amdahl at Work" by Benjamin Verschuere and Angus Cameron. "The diffusion of generative AI across industries raises the question of who is affected by the resulting labour-market adjustment. Using 12.3 million person-months of US Current Population Survey data from 2015 to 2026 and the task-based codifiable substitution decomposition of Verschuere and Cameron (2026), we identify two patterns at the subset of US industries that have adopted generative AI (both of which also remain detectable on the full 23-industry panel). First, AI substitutes lower-skilled tasks for every worker regardless of age: at clerical occupations in AI-adopting industries, substitution loadings are −0.37 for young workers and −0.30 for prime workers (log-point employment growth, both p < 0.01). Second, at higher-skilled work, age and educational credential shield prime workers (35–49) from substitution exposure while young workers (22–25) at the same positions are not shielded: the young–prime differential in substitution loading at non-clerical occupations in AI-adopting industries is +0.50 log points (p = 0.016), against a null +0.09 (p = 0.60) in low-AI industries. The age-shielding gradient appears across task position, educational credential, and occupation wage, and, within task tiers, concentrates at occupations with high judgment content. We show the organizing mechanism is a mediation of Amdahl’s Law into labor economics within the job. The complementation share binds as the substitutable share is accelerated." Working paper: papers.ssrn.com/sol3/papers.cf… Interactive app: liminal-capital.com/apps/ai-amdahl…

Agreed, it seems there was a change in productivity trend before but post ChatGPT there has also been a widening of the gap between AI exposed sector (high C*) and non AI-exposed ones (low C*). We ran some DiD on that, full paper here (where also also look at 18 other economies): papers.ssrn.com/sol3/papers.cf…

Like every other econ reporter, I've spent much of the past couple years trying to understand what A.I. will mean for the economy and for workers. But it's been frustratingly hard to reach any firm conclusions, because so much is uncertain -- how the technology will evolve, how quickly it will be adopted, how policymakers will respond.

Great historical perspective, and it seems history might be rhyming... The teamster episode is strikingly similar to what our research currently finds with white-collar jobs. 1) The AI job impact is happening via lower hiring: entry-level job starts (22–25) in the most AI-substitutable occupations are down about a third, ~227,000 per year. 2) The young are not shielded even in skilled jobs, while experienced workers in judgment-heavy roles are retained and expanded. Your young men who stopped entering teamster work while older drivers finished their careers at rising wages, in 2026 form (heatmap below) Two papers for your reference: – AI is adding and taking jobs, so the aggregate effect looks quiet: papers.ssrn.com/sol3/papers.cf… – Who is being impacted by AI (today’s equivalent of the teamster example): papers.ssrn.com/sol3/papers.cf…

@soumitrashukla9 @RonnieChatterji Thanks! Any feedback welcomed.

@int_mon_econ Regulatory arbitrage...

And in a follow up paper drilling deeper into the dichotomy we find where AI is adopted, employment is declining for the young and also generally any low skilled automable jobs. While shielded worker have so far been older employees with college degrees where AI complement them (chart below). Paper: papers.ssrn.com/sol3/papers.cf…

@RonnieChatterji this echos some of our empirical findings. AI can act as a substitute for some task and a complement or other (while being inert for physical jobs) Partitioning jobs the same way we find that, in the US since ChatGPT, employment in AI-substitutable work has been growing ~3.3pp/yr slower than untouched work and AI-complemented work is growing ~3.3pp/yr faster. Essentially the complemented jobs are the bottelneck after AI-substitute jobs are being accelrated. Paper here: papers.ssrn.com/sol3/papers.cf… The mapping of 798 occupation below: liminal-capital.com/apps/ai-job-ma…

Strong convergence with what we find in CPS data and looking fwd to read that paper. We partition every O*NET task into what AI substitutes vs complements . In employment data the bottleneck already binds: where AI is adopted, the workers holding the judgment tasks AI cannot do are retained and expanded, while young and lower-skilled workers in substitutable work are not (chart below).Amdahl’s Law, applied within the job. Link to paper: papers.ssrn.com/sol3/papers.cf…

Completely agree with the premise: the debate has been too focused on the jobs AI takes and not on what it augments. And actually there are already some sign in the data using US CPS data shows that both margins are visible: sure employment in AI-substitutable work is growing ~3.3pp/yr slower than untouched work, but AI-complemented work is growing ~3.3pp/yr *faster*. The AI-augmentation side is real and already measurable... Here is how the mapping looks like: liminal-capital.com/apps/ai-job-ma…

@econcallum Fwiw data from OpenRouter points toward Chinese models generating around 25% of revenue on their platform... Data here: liminal-capital.com/apps/ai-tokeno…

@AngelAuYeung @AngelAuYeung And there is also some revenue grab from Chinese models... Full tracker here: liminal-capital.com/apps/ai-tokeno…

@socialcapital @socialcapital worth noticing the fact Chinese models are now also grabbing some revenue share away from US model Full tracker here: liminal-capital.com/apps/ai-tokeno…

@ZeffMax @AngelAuYeung @ZeffMax worth noting that Chinese models have not only the upper hand in terms of token usage but now they are also grabbing revenue share.... Full tracker here: liminal-capital.com/apps/ai-tokeno…

@albrgr @albrgr in case you are interested we built a tracker and are seeing Chinese models taking more revenue share (not just tokens) like Ramp data. Full tracker here: liminal-capital.com/apps/ai-tokeno…

@Noahpinion Fwiw we are seeing the Chinese models grabbing more share of US revenue models (not just token) Full tracker (with model capabilities) here: liminal-capital.com/apps/ai-tokeno…

@StealthQE4 Fwiw we are seeing the Chinese models grabbing more share of US revenue models (not just token) Full tracker here: liminal-capital.com/apps/ai-tokeno…

@EpochAIResearch @cherylwoooo @ansonwhho Fwiw we are seeing the Chinese models grabbing more share of US revenue models (not just token) Full dashboard here: liminal-capital.com/apps/ai-tokeno…

Hiding in the mean: AI is adding jobs and removing them at the same time Benjamin Verschuere & Angus Cameron (Liminal Capital) ow.ly/aBQF50ZUm8B