At a glance
- What it is: The AI Jobs Transition Framework: Mapping AI’s Near-Term Impact on Jobs, a report by OpenAI’s economic research team.
- Who: Alex Martin Richmond, the report’s author, with an opening note by Ronnie Chatterji, the company’s chief economist.
- Where: published by OpenAI Economic Research in April 2026. It is not a peer-reviewed paper and has no DOI: it is a report by the company itself. openai.com/index/modeling-ai-jobs-transition
- Type: a theoretical framework built on 921 occupations (99.7% of US employment) and checked against real ChatGPT usage data.
First reading: what it does and what it finds
It helps to start with the problem the report says it sets out to solve, because it organizes everything else. Most analyses of AI’s impact on employment start from a single question: how exposed is a job to AI, that is, how many of its tasks could a model do? The report argues that this measure, exposure, is too crude. Knowing that AI can do many of a job’s tasks does not tell you whether that job will be automated, redesigned or even grow.
The proposal is to add two more layers to exposure. The first is human necessity: even if AI can do the cognitive work, is a person still needed for physical, interpersonal or accountability reasons? The report distinguishes three types. Regulatory or accountability necessity, when someone with a license has to sign off or answer for the work, as with judges and lawyers. Relational necessity, when the value depends on human contact, care or teaching, as with teachers and nurses. And physical necessity, when someone has to act in the real world, as with plumbers and physical therapists. The second layer is demand elasticity: if AI makes a service cheaper, do people buy so much more of it that the same staff, or even more, is needed? It is the old Jevons paradox: when something gets cheaper, it is sometimes consumed so much more that employment in the sector rises instead of falling.
With those three layers, plus a fourth question about whether AI is already being used today for those tasks, the report sorts occupations into four groups. The headline result: 18% of jobs face a high risk of near-term automation, 24% will be reorganized (workers are still needed, but there could be fewer of them), 12% could grow thanks to AI and 46% will see few immediate changes. The underlying reading is that many highly exposed jobs are not headed for automation, but for redesign or expansion.
The data they use to validate the framework comes from their own house: ChatGPT usage. They cross-reference work conversations from the consumer version, aggregated and anonymized and taken from the second half of 2025, with the tasks of each occupation, without observing who each user is. A clear finding emerges there: in the jobs the framework flags as highest risk of automation, ChatGPT is used about three times more than in those under the least pressure. And everywhere, real use lags far behind what AI could technically do. They call that gap a capability overhang: available technical capacity that is not yet being used. In the highest-risk jobs, theoretical exposure reaches 90% while realized exposure is only 24%, a gap of 66 percentage points. That this is precisely where it is used the most, and at the same time where the most room remains, is the heart of the argument: capability, on its own, does not tell you where the labor market is changing.
There is one finding that the report itself treats with caution and that is worth not skipping. When they look at actual unemployment since early 2024, the group that rose the most is not the highest-risk one, but the “few immediate changes” group, at +0.6 percentage points, versus +0.3 in the high-risk and reorganizing groups. The authors put it this way: it is still hard to clearly link AI and the aggregate labor market, that could change quickly, and so these occupations warrant close monitoring.
One technical point that matters: two of the three layers are not measured but estimated with a model from OpenAI itself. Human necessity is classified by GPT-5.4; demand elasticity is calculated by GPT-5.4-mini from a hypothetical price-drop scenario. The authors are explicit that elasticity is the least observed part of the framework and that these are approximations, not causal evidence.
Second reading: from Latin America
The report is about the United States and does not single out Latin America, so almost everything that follows is my reading, not the document’s. With that caveat up front, there is something that translates well to the region.
The most useful idea is not the 18%, it is the method. Separating exposure from human necessity and from elasticity helps avoid confusing “AI can do this” with “this job is going to disappear.” In much of the region, where public employment, care work and person-to-person services are huge, that distinction matters: a teacher or a nurse may see AI draft materials or summaries for them without the person in front of the class or at the bedside becoming any less necessary. The report would say those jobs are reorganized, not automated.
But the same framework carries a warning that weighs differently in the region. A job can keep needing people and still employ fewer of them: if each worker produces much more and demand does not grow enough to absorb it, there are hands to spare. In economies with more informality and less of a cushion for retraining, that adjustment can hurt more, and faster, than in the United States. This is my own hypothesis: the report did not study these countries, and I leave it as such.
And there is a deeper reason to read it with caution from here, which is the part worth not leaving out. The fine print below details whose report this is and what data it is built on; what I add from the region is what to do about that. The company has an interest in the story coming out as a manageable transition rather than a wave of layoffs, and that choice of framing is not neutral. Before a ministry in the region cites it to justify a policy, it is worth asking for what this report lacks: local measurements and evidence of real-world outcomes.
The question it leaves works for any country thinking about its employment policy. If exposure is not enough to know where it will hurt, the task is not to make lists of doomed jobs, but to build the capacity to measure, in time, where pressure is showing up. The report itself calls for this in the United States: better occupational measurement. In Latin America, where that measurement tends to be weaker, the advice becomes even more urgent.
The fine print
- It is an OpenAI report that measures AI’s impact with ChatGPT usage data and with estimates made by its own models, GPT-5.4 and GPT-5.4-mini. It is best read as internal evidence and not as an independent audit: the company has an interest in the topic and chooses what to measure and how to tell it.
- The four categories are not forecasts of job loss. The authors insist they are a map of where pressure could appear first, not a prediction of how many jobs will be lost.
- Demand elasticity is, in the authors’ words, the least observed part of the framework: approximations generated by a model, not causal measures. The report also does not capture general equilibrium effects, wages or linkages between occupations.
Paper keywords: AI exposure, human necessity, demand elasticity, labor market transition, occupational automation
Automated reading. This text was generated by Claude, an Anthropic model, from the original source, without line-by-line human review. It may contain errors or debatable interpretations; to check any point, see the original source.