What New Jobs Could Emerge in the Age of AI?

In late 2021, an outsourcing firm called Sama began sending workers in Nairobi short passages of text to read and label, nine hours a shift, 150 to 250 passages a day. Some of the text described child abuse, bestiality, and self-harm in graphic detail — the material a major AI lab needed labeled so its chatbot would learn not to produce it. The client paid Sama roughly $12.50 an hour for the work. The workers who actually did it took home between $1.32 and $2. Around three dozen people held that job before the contract ended early, eight months ahead of schedule [1].

That is a new job the Age of AI created. It is real, it exists today, and it complicates every tidy story — optimistic or alarmist — about what "new work" in this technology age actually looks like. It's also not the whole story: the same technological shift that produced that job has also measurably grown employment in far better-paid, better-known occupations. History says both of those things can be true simultaneously, and it says so with a specific, checkable number. Roughly six in ten jobs Americans held in 2018 did not exist in 1940 — not "changed," did not exist — according to a peer-reviewed reconstruction of eight decades of census records tracing exactly when every new job title first appeared in American work [2]. Most of what people are paid to do today is work their great-grandparents could not have applied for. The question this investigation asks is where, specifically, that pattern is showing up right now — and who it's showing up for.

Economists describe automation's effect on work as a race between two forces happening at once. It displaces labor from tasks a machine has learned to do, and it reinstates labor by creating new tasks where humans still hold the advantage — tasks that frequently didn't exist until the technology itself created the need for them [3]. Telephone operators vanished; network engineers appeared. Typing pools dissolved; IT departments materialized in their place. The census reconstruction above quantifies how that race has actually gone: reinstatement has won, decade after decade, which is the entire reason six in ten of today's jobs are new since 1940 [2].

But the same study contains a detail that optimists tend to skip past. The kind of new work changed sharply around 1980. From 1940 to 1980, new occupations mostly appeared in the middle of the pay scale — production and clerical work that paid ordinary people well. Since 1980, new work has split toward the two ends: high-paid professional roles at the top, low-paid service roles at the bottom, and comparatively little in between [2]. The Sama contract and the software engineer down the street who now uses an AI copilot every day are both, in this framework, "new work." History's comfort was never that new work would appear. It was silent on whether it would appear somewhere you'd want it.

Set aside, for a moment, what employers say they expect to happen, and look instead at occupations that already exist at meaningful scale. The federal government's own wage survey counted about 245,900 data scientists employed in the United States as of May 2024, and about 182,800 information security analysts — both categories that barely existed as distinct occupations a decade earlier, both now large enough to be tracked individually in the same survey that counts electricians and nurses [4][5]. Those are measured, current headcounts, not projections of what might exist by 2030.

Better evidence still comes from an actual experiment rather than a headcount. Researchers ran a hiring simulation with 1,725 real recruiters across the UK, US, and Germany, asking each to evaluate synthetic candidate profiles that varied only in whether they listed AI skills. Listing them raised the odds of an interview invitation by roughly 8 to 15 percentage points across three very different occupations — graphic designer, office assistant, and software engineer — and partly or fully offset the usual penalties recruiters apply for a candidate's older age or lower formal education [6]. That's not a postings-price correlation or a survey of intentions. It's real recruiters making real evaluative decisions, and AI fluency moving the outcome.

There's supporting texture from the demand side, too, though it deserves its caveat: job postings asking for AI skills advertise salaries meaningfully above comparable postings without them, and a majority of that demand now sits outside traditional technology roles entirely — human resources, finance, marketing, operations [8]. That's advertised, not paid, compensation, and it's a correlation in postings rather than a causal test. But it points the same direction as the hiring experiment above: whatever is happening is not confined to software engineering. And the field-experiment evidence already established elsewhere in this Convergence — that AI assistance amplifies less-experienced workers most, functioning as collaboration rather than replacement — fits the same pattern [7]: the emerging roles look more like "existing job plus AI fluency" than "brand-new job title."

None of that erases the counter-signal. Employment for the youngest workers in the most AI-exposed occupations is falling in measured data even as these other numbers rise [9]. The creation evidence above is real, current, and measured. So is the displacement at the entry level. Both are true, in different places, at the same time.

Two questions sit on top of these facts, and credible observers land in different places on both. Competing hypothesis Can reinstatement keep pace this time? One view: AI is a general-purpose technology adopted faster than any predecessor, so the new-task engine should also run faster, and the AI-fluency premium already spreading through non-technology occupations is early proof of exactly that [3][8]. The skeptical view: what actually distinguishes AI is that it targets cognitive tasks — the historical refuge into which displaced workers have always retrained — so this time displacement and reinstatement are competing for the same ground, and the post-1980 polarization pattern deepens rather than heals [2].

Competing hypothesis And there's a second, less comfortable disagreement this investigation can't avoid, given where it opened. Is the Sama-style annotation job a real answer to "what new jobs does AI create," or is it the exception that proves the pattern above — new work concentrated at the very bottom of a global pay scale, doing the least visible, least desirable part of building a system whose benefits land somewhere else entirely? Reasonable people land in different places, and the honest answer is that the same reporting used here has documented both real income for workers who had few other options, and conditions serious enough that several outsourcing contracts for this kind of work have ended early.

The new-work question doesn't resolve inside a single labor market — it sits at the intersection of at least three systems already documented above: domestic occupational classification (the BLS categories that took years to even start counting data scientists and security analysts separately) [4][5], corporate hiring and screening (where the recruiter experiment shows AI fluency already changing real decisions) [6], and a global outsourcing market that routes the least desirable, least visible tasks — like the Sama annotation contract — to the lowest-cost labor available anywhere on earth [1]. Those three systems don't run on the same clock: official job categories update on a years-long lag, hiring decisions shift in real time, and outsourcing contracts can start or end within months.

This investigation's companion pieces in the Convergence feed directly into this one: the entry- level contraction documented here [9] is the same contraction the layoffs investigation traces from the employer side, and the hiring-funnel changes explored elsewhere in this Convergence are the mechanism by which "new work" either does or doesn't reach the people who most need it.

History says new work follows a general-purpose technology, but it never says only one kind. Nothing here rules out the others — the honest reading of the evidence is that these can all keep growing at once, at very different rates, in very different places.

Strip away the expectations and the press releases: AI-adjacent occupations already employ hundreds of thousands of people in the United States, measured, not projected [4][5], and the same technology has already created work as unglamorous and poorly compensated as anything documented in a Kenyan call center. History says new work reliably follows a technology like this one [2][3] — it has never promised the new work would be evenly distributed, well paid, or even visible to the people who benefit most from it existing. Public, trackable indicators worth watching over the coming months and years:

• Hiring & recruiting evidence — Whether the AI-skills hiring advantage found in the recruiter experiment replicates in other occupations and countries [6]. Whether entry-level hiring in AI-exposed occupations stabilizes or keeps contracting [9].
• Official occupational data — Whether the next round of BLS occupational data shows measured (not projected) growth in AI-adjacent job categories [4][5]. Whether AI-era titles begin appearing in official occupational taxonomies. The share of total employment in occupations that did not exist in 2022 — the only measure that has ever actually answered this investigation's question [2].
• Labor conditions in the annotation economy — Whether annotation and content-moderation contracts for major AI labs shift toward better-documented labor standards, or keep churning through short-lived arrangements.
• Wage and pay-scale trends — Wage trajectories in AI-adjacent occupations relative to the broader economy — the polarization question [2].

If your own job title didn't exist five years ago — what is it, what do you actually do all day, and did AI create the need for it or just change how you do something older? Tell us. We read every message.