“AI trainer” might be the most overloaded job title of the decade. Recruiters use it for RLHF raters sitting in browser tabs ranking chatbot replies, robotics technicians demonstrating grasp motions in warehouses, in-person actors recording emotional speech for voice models, and ML engineers fine-tuning LoRA adapters on proprietary datasets. Applicants arrive expecting one job and find another. Burnout and churn follow.
We interviewed twenty-three workers across vendors and direct AI lab contracts, reviewed training manuals leaked from three major annotation platforms, and shadowed a documented eight-hour RLHF shift. This article separates categories, describes hour-by-hour work, names the tools, and maps realistic promotion paths so you can evaluate offers before signing an NDA you do not yet understand.
The Four Jobs Hidden Under One Title
Human feedback rater is the most common meaning. You compare two or more model outputs, pick the better one, write a short explanation, flag safety violations, or rewrite an ideal answer. Tasks arrive in a web interface with strict time limits and style guides that change without warning. Pay ranges from three dollars an hour in offshore gig markets to forty-five dollars an hour for US-based specialized safety work on W-2 contracts.
Data curator trainers assemble datasets: finding edge cases, writing gold-standard responses, validating synthetic examples, tagging failure modes. This work requires domain knowledge—medicine, law, coding, finance—and pays more than generic preference ranking because mistakes propagate into model behavior at scale.
Robotics and embodied AI trainers perform physical demonstrations: wearing motion capture gear, guiding robot arms, labeling video of human actions, repeating mundane tasks thousands of times so models learn affordances. These roles are on-site, often shift-based, and physically demanding in ways remote raters never experience.
Technical fine-tuning trainers are the smallest category and closest to traditional ML work: preparing JSONL files, running training scripts, evaluating checkpoints, collaborating with researchers. They usually hold computer science degrees and are not hired through generic “AI trainer” Instagram ads.
One title, four jobs—ranking replies is not fine-tuning models.
A Hour-by-Hour RLHF Shift
Morning starts with a guideline update email. Yesterday’s rules for ranking helpfulness may now include new restrictions on medical advice or political neutrality. You pass a five-question calibration quiz before accessing paid tasks. Fail twice and you lose the day’s project access while still expected to stay available.
Hours one through four are ranking and rewriting. Each task shows a user prompt and two model responses. You select the better response, tag issues from a dropdown—hallucination, unsafe, off-topic, verbose—and sometimes rewrite a corrected version in two hundred words or less. Average handle time targets range from ninety seconds to four minutes depending on complexity. Fall below quality scores and your queue downgrades to lower-paying batches.
Afternoon includes a team standup on Zoom for vendor employees: review audit failures, clarify edge cases, warn about clients rejecting batches for inconsistent tone. Independent contractors skip the meeting but absorb the consequences when aggregate team scores drop and projects pause.
Final hour is administrative: log hours, respond to quality reviewer messages, complete mandatory wellness survey, re-read updated policy PDF. Unpaid time often adds thirty to sixty minutes daily for training modules clients require but do not compensate.
Calibration quizzes gate your pay before the shift begins.
Tools, Metrics, and Pressure
Workers interact with proprietary annotation suites—often customized Label Studio, Scale Nucleus, or internal lab tools—inside locked-down browsers. Copy-paste is disabled. Screenshots trigger warnings. Some vendors install monitoring software on contractor machines. Productivity dashboards track tasks per hour, agreement with other raters, and audit pass rates visible to managers in real time.
Inter-annotator agreement is the hidden KPI. If your rankings diverge too often from the consensus, algorithms flag you as noisy. Noisy raters lose access to premium projects without explanation. Appeals exist on paper; workers report success rates below ten percent. The system optimizes for consistency with guidelines—even when guidelines are ambiguous—over independent judgment.
Psychological toll varies by project. General chatbot preference ranking is tedious but mild. Safety red-teaming exposes workers to hate speech, sexual violence descriptions, and self-harm scenarios for hours daily. Trauma support, when offered, is a fixed number of counseling sessions per year grossly inadequate for exposure levels documented in academic studies on content moderation and AI safety work.
Consistency with ambiguous guidelines beats independent judgment.
Pay, Contracts, and Geographic Split
US W-2 vendor employees report forty thousand to sixty-five thousand dollars annually at full-time hours with limited benefits. Specialized safety raters with advanced degrees reach seventy-five thousand to ninety thousand dollars. Gig workers piece together income across platforms with no stability; annual earnings swing from twelve thousand to thirty-eight thousand dollars depending on project access and rejection rates.
Workers in Kenya, India, and the Philippines often earn locally competitive wages that translate to two to six dollars an hour by US standards, performing identical tasks for the same global models. Clients defend regional pricing as market rates; organizers argue it encodes exploitation into foundation training pipelines. Both descriptions contain truth depending on which side of the spreadsheet you sit.
Contracts classify workers as independent contractors even when schedules, tooling, and supervision resemble employment. Non-compete clauses and broad NDAs silence workers who might otherwise report guideline changes or unpaid rework. Read termination clauses: many agreements allow instant deactivation without appeal when quality scores dip below thresholds defined unilaterally.
Regional pay tables train the same models at different prices.
Career Progression That Actually Exists
Vertical paths include senior rater, team lead, quality analyst, guideline writer, and data operations coordinator. Each step reduces hands-on ranking and increases documentation, training delivery, and client calls. Leads at large vendors report seventy thousand to ninety-five thousand dollars salary plus benefits rare at entry tiers.
Lateral exits move toward ML evaluation engineer, AI product operations, trust and safety analyst, or technical writing for developer tools. Success requires documenting projects with metrics, learning SQL and Python for evaluation scripts, and networking on teams that do not treat raters as interchangeable.
Dead-end risk is real for workers who stay on generic ranking tasks for three or more years without specialization. Automated quality prediction and synthetic preference models already reduce headcount on some projects. Treat entry AI training as a stepping stone with a written six-month skill plan—not a destination.
How to Evaluate an Offer
Ask which client industry, which task types, average batch acceptance rate, hourly versus piece pay, and whether rework is compensated. Request a sample task before signing. Search Reddit and Blind for project codenames when NDAs allow. Avoid offers that refuse to name the pay model or guarantee minimum hours.
Red flags include unlimited availability expectations on gig contracts, mandatory unpaid training exceeding eight hours, quality thresholds changed retroactively, and no mental health resources on safety-heavy projects. Green flags include W-2 employment, published rate tables, documented promotion paths, and team leads who answer technical questions about why guidelines exist.
AI trainer is real work powering models consumers treat as magic. It is also fragmented, unevenly paid, and emotionally hazardous in ways recruiting copy omits. Know which of the four jobs you are applying for before you click accept.
Ask about rework pay and counseling before you sign.
The phrase AI trainer sounds futuristic. The shift feels like call center quality assurance crossed with editorial judgment crossed with unpaid policy homework. That honesty helps you negotiate pay, protect your mental health, and plan an exit into roles that compound skills instead of consuming them one ranked pair at a time.