Four years after ChatGPT arrived, the job titles it was supposed to create mostly have not appeared. Stanford's 2026 AI Index shows the demand went somewhere else: into marketing, engineering, project management and analysis roles that now expect AI fluency as part of the existing job. The numbers underneath are dramatic in places — mentions of AI agent skills in US postings went from 151 in 2024 to 16,541 in 2025 — and near-zero in others. AI ethics and governance, the cluster that was supposed to become a profession, appears in 0.05 percent of US postings, the lowest of the ten AI skill clusters Stanford tracks.
Agent skills are the growth story, and the shape of that growth matters. A jump from 151 to 16,541 in a single year is not a market expanding; it is a market being created from nothing, because almost nobody was writing "AI agent" into a job description in 2024. The most concentrated piece of it is LangGraph, the open framework for deploying agents: postings naming it rose from 194 to 4,294, growth of more than 2,000 percent, and roughly a quarter of all agent-skill mentions now attach to that one tool. What employers mean by the skill is broader than the framework — coordinating teams of agents, getting agents to operate websites and make purchases, designing interfaces and tools that both agents and humans can use — but the hiring signal is already clustering around specific software rather than a general competence.
Prompting followed a different path. The prediction was a profession; the outcome is a line item. US postings requiring the skill rose from 6,152 in 2024 to 22,227 in 2025, a real increase, but read that number in a national labour market and it is small. Companies are not hiring prompt engineers. They are hiring marketers, analysts, project managers and engineers who can write prompts while doing the job they were hired for.
Machine learning remains the most frequently mentioned AI skill in postings, which is the least fashionable fact in the index and probably the most durable one. It demands programming, mathematics and statistics, and it sits underneath everything else on the list. The vocabulary has moved to human-machine interaction and workflow automation, while the algorithms doing the work inside recommender systems, predictive maintenance, fraud detection and risk modelling are the same ones, maintained by the same kind of specialist.
Chatbots show what happens to a skill that wins. Postings requiring ChatGPT skills grew from 5,535 to 14,376 over the past year, but as a share of AI-related postings the vocabulary collapsed — "chatbot" down 50 percent, "conversational AI" down 68 percent. Those two movements are not in conflict. The category dissolved into a product name, and the product name is on its way to becoming what "proficient in Microsoft Word and Excel" was on a résumé twenty years ago: not a qualification, just an absence of a problem.
My read is that the index describes an absorption, not a boom. The new-profession thesis assumed AI would be a department. Instead it became a tool, and tools get added to the job description of whoever already does the work. That is good news for anyone employed and bad news for the retraining industry built on the other assumption, because a skill that lives inside an existing role is much harder to certify, price or teach in six weeks. It also means the growth rates in the agent cluster will not repeat. You can only go from 151 to 16,541 once.
The part nobody is asking about is the 0.05 percent. AI ethics and governance is not a slow-growing cluster; it is a cluster employers are effectively not hiring for, at the exact moment the EU AI Act is turning oversight into a documented legal obligation rather than an opinion. The plausible explanation is that companies do not consider their own AI strategies mature enough to staff governance, and are waiting for an incident to force the question. That is a defensible calculation right up until the first incident, at which point the people who could have built the oversight function are being recruited by everyone at once, under a deadline, at a price nobody budgeted for.