When Anthropic published its first external job listing for a Prompt Engineer at $175,000–$335,000 per year, the technology press reacted with a mixture of disbelief and excitement. The listing asked for no formal computer-science degree. It asked instead for "a talent for explaining complex technical concepts" and "a deep familiarity with how large language models behave." Within weeks, similar postings appeared at Google, Microsoft, and dozens of startups. A job category had materialized almost overnight — one whose entire purpose was to communicate with systems that did not exist four years earlier.
Not every "new" job title is genuinely new. When ATMs spread in the 1970s, banks created "ATM service technician" roles — but these were repackaged versions of existing repair and cash-handling work. The technology gave the job a new name; the underlying cognitive demands were familiar.
A subset of AI-era roles is different in kind. Prompt engineering, AI alignment research, AI red-teaming, and machine learning operations (MLOps) involve reasoning about systems whose outputs are probabilistic, whose failure modes are subtle, and whose capabilities change month to month. The knowledge required did not exist in any usable form before roughly 2017–2020. These are not repackaged jobs. They are structurally new.
LinkedIn's 2023 Jobs on the Rise report listed "AI Specialist" as the fastest-growing U.S. job title, with a 74% year-over-year increase in postings between 2021 and 2023. "Prompt Engineer" went from zero listings in LinkedIn's database in early 2022 to over 2,000 by mid-2023.
Economists who study technological employment distinguish three mechanisms by which new technology creates new work. All three are visible in AI's current expansion.
Complementarity: AI tools amplify existing human roles, creating demand for people who specialize in pairing human judgment with AI output. A radiologist who supervises AI diagnostic scans is doing something qualitatively different from one who reads films alone — and hospitals are now hiring specifically for that hybrid capacity.
Oversight: Any system powerful enough to make consequential decisions requires humans to audit, challenge, and correct it. AI's rise has generated sustained demand for AI safety researchers, model evaluators, and algorithmic auditors — roles dedicated to ensuring AI systems do what they claim to do.
Infrastructure: New technology requires new infrastructure workers. The cloud computing boom created hundreds of thousands of DevOps jobs. AI's compute demands are generating analogous roles in MLOps, GPU cluster management, and AI-specific data engineering.
A persistent misconception is that AI-native jobs require deep software engineering skills. Some do. Many do not. Anthropic's prompt engineering role explicitly said a computer science degree was not required. What it required was calibrated curiosity about model behavior — the ability to notice when outputs are subtly wrong and to hypothesize why.
Similarly, the role of AI Trainer — people who provide feedback on model outputs to improve RLHF (reinforcement learning from human feedback) pipelines — relies primarily on domain expertise and judgment. A lawyer who evaluates whether an AI's legal summaries are accurate is performing AI training work. So is a nurse who flags clinical errors in AI-generated care notes.
The common thread across AI-native jobs is evaluative intelligence: the capacity to judge the quality, safety, and appropriateness of AI outputs in a specific domain. That capacity is distributed across many fields — not concentrated in computer science departments.
Evaluative intelligence — the ability to assess AI outputs against domain-specific standards of quality, accuracy, and appropriateness — is the foundational skill underlying most AI-native roles. It is domain-portable and does not require coding ability.
Use the AI assistant below to explore the boundaries of AI-native job categories. Your goal is to distinguish roles that are structurally new from those that are simply relabeled versions of existing work.
In March 2023, an open letter signed by over 1,000 researchers — including some who had built the very systems under discussion — called for a six-month pause on training AI systems more powerful than GPT-4. The letter cited "profound risks to society and humanity." It did not slow hiring at Anthropic, OpenAI, Google DeepMind, or Meta AI. It accelerated it. Every signatory of that letter implicitly acknowledged that the problem of making AI safe required more expert humans working on it — not fewer AI systems in the world.
AI safety is not a department that exists despite AI development — it exists because of it. The more capable AI systems become, the more consequential their failure modes, and the larger the workforce needed to understand, test, and constrain those failures. This creates a structural employment dynamic unlike most technology fields: the riskier the technology, the more safety jobs it generates.
OpenAI's "Superalignment" team, announced in July 2023, committed to dedicating 20% of the company's compute resources to safety research. Anthropic was founded explicitly as a safety-focused AI company and structured its entire research program around the Constitutional AI framework. DeepMind's Safety team predates GPT-4 by several years. These are not PR exercises — they represent sustained investment in a technical problem that does not yet have a solution.
In 2023, 80 Levels reported that AI safety and alignment researchers at top labs earned between $200,000 and $900,000 in total compensation — among the highest ranges in any technical field globally. Demand exceeded supply by an estimated 10-to-1 ratio at leading organizations.
The field is not monolithic. It encompasses several distinct research tracks, each generating its own hiring pipeline.
In November 2023, the UK government launched the world's first dedicated national AI Safety Institute. Its founding brief included conducting evaluations of frontier AI models for dangerous capabilities — the same work that AI labs do internally, but performed by an independent public body.
The Institute began hiring immediately for roles including Senior AI Safety Researcher, Evaluations Engineer, Policy Lead, and Model Evaluation Scientist. Salaries were published under the UK Senior Civil Service pay framework — significantly below private sector rates, yet the roles attracted hundreds of applicants each. The Institute represented something new: government as an employer of AI safety expertise, not merely a regulator of AI companies.
The U.S. NIST AI Safety Institute, established by executive order in late 2023, followed a similar hiring pattern. Within six months both institutes had published model evaluation frameworks that referenced the need for thousands more trained evaluators globally.
AI safety has created demand for a genuinely interdisciplinary workforce — one that requires combinations of machine learning research, philosophy of mind, cognitive science, formal logic, and domain expertise in law, medicine, or finance. No single degree path currently produces this worker. Most practitioners assemble the skillset through independent research, fellowships, and on-the-job training.
AI safety needs workers from many disciplines — not just ML researchers. Use this lab to explore how someone with your background or a background you're curious about might enter the field. The assistant will help you map realistic pathways.
Scale AI, founded in 2016 to provide labeled training data for AI systems, was valued at $7.3 billion in its 2021 funding round. By 2023 it employed tens of thousands of contractors globally — annotators, evaluators, and domain experts — who reviewed AI outputs, labeled images, transcribed audio, and assessed the quality of model responses. This workforce did not shrink as AI improved. It grew. Better AI required better training data, which required more human judgment at higher quality thresholds. The company's growth illustrated a counterintuitive principle: AI's appetite for human feedback scales with AI capability.
Every major AI model deployed since 2017 has relied on large quantities of human-labeled data. Image recognition models require humans to annotate millions of photographs. Language models require human feedback to distinguish helpful from harmful outputs. Autonomous vehicles require humans to label every pedestrian, lane marking, and traffic sign in millions of video frames.
The global data annotation market was valued at approximately $2.04 billion in 2022 and is projected to reach $17.1 billion by 2030 (Grand View Research, 2023). This growth is not driven by legacy AI systems — it is driven by the continuous improvement of frontier models that require progressively more sophisticated human judgment, not just volume.
Sama, Appen, and Scale AI represent the large end of this industry. But an enormous ecosystem of smaller regional providers has emerged in Kenya, the Philippines, India, and Venezuela — countries where English literacy and low wage costs make annotation work economically viable for both workers and clients.
A January 2023 TIME investigation documented that Kenyan workers hired through Sama were paid approximately $1.32–$2.00 per hour to review and label graphic content — including descriptions of violence, abuse, and self-harm — to train OpenAI's content moderation filters. The work created real jobs in Nairobi but raised serious questions about labor conditions in the AI supply chain. OpenAI subsequently ended its contract with Sama for that work.
Data annotation is the most visible layer of the human-in-the-loop economy, but it is not the only one. Several other categories of human oversight work have grown significantly with AI deployment.
The human-in-the-loop workforce has several characteristics that distinguish it from both traditional tech employment and traditional gig work. It is highly distributed geographically, often contracted rather than employed directly, and compensated at rates that vary enormously — from $1–3/hour for basic annotation to $100+/hour for expert domain review.
Critically, this workforce's existence is not a transitional phase before full automation. The more sophisticated AI systems become, the harder it is to evaluate their outputs without human judgment. A model that writes plausible but subtly incorrect legal briefs requires a lawyer to evaluate it — not another AI. The demand for high-quality human judgment in the AI pipeline is likely to increase, not decrease, as models improve.
The human-in-the-loop economy is not a stopgap measure awaiting full automation. It is a structural feature of AI development. Models that are powerful enough to be deployed in consequential domains are, by definition, too consequential to operate without human oversight — creating persistent demand for evaluative human labor at every capability level.
Data annotation and RLHF work depends on the ability to evaluate AI outputs against clear quality criteria. In this lab, the assistant will show you pairs of AI-generated responses and ask you to evaluate them — just as professional annotators do. You'll explain your reasoning and develop a framework for quality assessment.
In March 2023, GPT-4 scored in the 90th percentile on the Uniform Bar Examination — the standardized test required to practice law in most U.S. states. In the same month it passed the USMLE Step 3 medical licensing exam with a passing score. These results were widely reported as evidence that AI would displace lawyers and doctors. What actually followed was different: law firms began hiring Legal AI Specialists, hospitals began creating Clinical AI Coordinator roles, and financial services firms started posting for AI Compliance Analysts. The tests measured knowledge retrieval. Professional work requires something harder to automate — contextual judgment, fiduciary responsibility, and institutional trust.
AI's impact on medicine has been most dramatic in diagnostic imaging. FDA-cleared AI diagnostic tools now assist radiologists in reading mammograms, chest X-rays, and CT scans. Companies including Aidoc, Viz.ai, and iCAD have deployed systems used in hundreds of hospitals. These systems do not replace radiologists — they prioritize the queue, flagging scans that show potential abnormalities for immediate human review.
The result has been a shift in radiologist workflow rather than radiologist displacement. Hospitals have created new roles: AI Imaging Coordinators who manage the integration of AI tools into radiology workflows, Clinical Data Scientists who validate AI diagnostic accuracy against local patient populations, and AI Ethics Consultants who evaluate whether AI tools perform equitably across race and gender groups.
A 2022 study in Nature Medicine found that AI-assisted radiologists detected 11.5% more cancers than unassisted radiologists, while also reducing false positives — suggesting a genuine capability augmentation rather than mere automation. This result increased, rather than decreased, the value of the radiologist-AI pairing.
In 2023, Epic Systems (which manages electronic health records for over 250 million patients) integrated ambient AI documentation tools into clinical workflows. Physicians speak during patient encounters; AI transcribes, structures, and drafts clinical notes. Epic reported that early adopters reduced documentation time by 25–35 minutes per day. This did not reduce nursing or physician headcount — it shifted physician time toward patient interaction, increasing capacity per provider and reducing clinician burnout. New "AI Clinical Implementation Specialist" roles emerged to train and support providers.
The legal profession's relationship with AI is more complex because law firms are partnerships that bill by the hour — a structure fundamentally challenged by AI tools that compress the time required for document review, contract drafting, and legal research.
In 2023, major firms including Allen & Overy, Freshfields, and Linklaters announced partnerships with AI companies (Harvey, a legal AI startup, raised $21 million in 2023 with A&O as a key partner). These deployments shifted junior associate work toward AI supervision and quality control rather than raw document review. New roles emerged: Legal AI Trainers who calibrate AI tools for specific practice areas, Legal Technology Directors who manage the firm's AI stack, and AI Audit Associates who verify AI-generated legal research before it reaches clients.
The pattern is consistent: AI compresses the time required for information-intensive tasks; professionals redirect that time toward judgment-intensive tasks; new hybrid roles emerge to manage the AI-professional interface.
Financial services has absorbed AI more rapidly than almost any other credentialed profession — in part because the industry already operated with large quantitative teams and algorithmic trading infrastructure. The transition to LLM-era AI has primarily affected research, compliance, and client-facing roles.
Bloomberg launched BloombergGPT in March 2023 — a 50-billion parameter language model trained on financial data. It outperformed general-purpose models on financial natural language tasks including sentiment analysis, named entity recognition in financial texts, and earnings call summarization. Bloomberg analysts began using it to compress research preparation time and to generate initial drafts of market summaries.
JPMorgan Chase filed a trademark application for "IndexGPT" in May 2023 — an AI-powered financial advisory tool. Goldman Sachs reported internally that AI tools had reduced the time required to prepare IPO prospectuses by 60–70%. In both cases, the firms did not eliminate the analysts who performed these tasks. They redeployed them toward client advisory work, relationship management, and the oversight of AI outputs — tasks where human judgment and institutional accountability remain legally and commercially necessary.
In medicine, law, and finance, AI is compressing time spent on information retrieval and document processing — tasks that historically required professional-level credentials but not professional-level judgment. The result is not displacement but restructuring: professionals spend more time on judgment, relationships, and AI oversight. New hybrid roles — coordinators, trainers, compliance analysts — fill the AI-professional interface. Credential requirements for these roles are evolving faster than degree programs can track.
Using what you've learned about AI-native roles, safety work, human-in-the-loop labor, and augmented professions, design a realistic hybrid role in a field you care about. The assistant will push you to be specific about what AI handles, what the human handles, and what new skills the role requires.