The Lab · Prototype
Jobs that didn't exist in 2022
A running list of work that wasn't possible before frontier models, as opposed to old jobs with new tools. Each entry says what the job is and why it couldn't exist earlier. Some have postings today, and some are still hypotheses.
AI does the task
Automation. The worker moves on, up, or out. This is the most studied case.
AI helps with the task
Augmentation. The worker gets more done. Most current research and hype sits here.
AI makes new tasks possible
New work that wasn't affordable or possible before. It's barely measured, and this list sits here.
The list
18 entries · last reviewed July 2026
Expertise Encoder
Sits with a retiring machinist, nurse, or loan officer and turns thirty years of judgment into a corpus a model can actually use.
No system could hold, let alone apply, a career of unstructured judgment.
Evals Designer
Writes the exams for AI systems. Decides what "good" means for software with no spec sheet, then measures it.
Software did what it was told. There was nothing open-ended to examine.
Agent Operations Lead
Runs a fleet of autonomous agents the way a plant manager runs a line: throughput, error budgets, escalation paths.
Software never acted on its own long enough to need an operations discipline.
Context Engineer
Designs what a model sees before it answers: the documents, the ordering, the framing. Real job postings exist today.
Nothing could read a million tokens, so nobody curated a million tokens.
AI Output Auditor
Checks machine output for a living: provenance, calibration, failure modes. Quality control for cognition.
Output volume was human-scale and mostly self-verifying.
Synthetic Data Curator
Builds training data that never happened to teach models things that did. Data design as a full-time craft.
Generating data used to cost more than collecting it.
Counterfactual Analyst
Runs a policy through a thousand simulated alternate versions before a legislature votes once.
Each scenario used to take a team a quarter. Now the marginal one is nearly free.
Institutional Translator
Makes an organization’s unwritten rules legible enough for agents to follow them.
The unwritten rules could not run on anything.
Model Welfare Researcher
Studies whether and when AI systems warrant moral consideration. A real research program at real labs.
Nothing artificial was a serious enough candidate to study.
AI Red Teamer
Attacks models professionally so the failures get found on payroll instead of in production.
Adversarial testing meant networks and locks, not minds.
Agent Economist
Designs market rules for autonomous buyers, sellers, and negotiators. Mechanism design goes operational.
Markets had no non-human participants with budgets and goals.
Curriculum Compiler
Turns a master practitioner’s method into a teachable, testable agent.
Teaching could not be compiled.
Simulation Demographer
Builds synthetic populations so a city can test a policy on everyone before trying it on anyone.
Population-scale simulation was a supercomputer grant, not a Tuesday.
Deep Research Editor
Assigns, supervises, and edits a research staff that never sleeps. The judgment moves up a level, and the reading moves down.
An editor’s staff was limited to people.
Memory Architect
Decides what a long-lived agent keeps, forgets, and consolidates. Institutional memory as a built system.
Software had storage. It did not have memory worth architecting.
Multimodal Field Auditor
Points a model at bridge photos, shop-floor video, or crop imagery and gets structured findings back.
Seeing was not something you could delegate.
Capability Forecaster
Tracks the frontier and tells institutions what becomes possible next quarter, not next decade.
The frontier moved slowly enough that strategy could ignore it.
Preference Elicitor
Turns messy human values into training signal a model can learn from.
There was no system that could take values as an input.
Know a job that belongs here? Nominate it in the comments.