How to get hired as an AI trainer in 2026-27

What an AI trainer actually does, and the five jobs hiding under one title

"AI trainer" is not one job. It is a label covering at least five kinds of work with different rates, different gates, different employers and sharply different futures. Confusing them is the single biggest reason people take work that pays badly and leads nowhere, then conclude the whole field is a scam. Before you apply to anything, work out which of the five a posting is actually describing, because the posting itself often will not tell you.

The first is generalist rating and annotation. You compare two model responses and pick the better one against a rubric, score a single response on helpfulness, factuality, instruction-following and safety, write short prompts to probe a behaviour, or apply labels from a taxonomy to text. The guidelines document is long, routinely tens of pages, and it is the authority rather than your opinion. This tier has the lowest rates, the most competition and the shortest half-life, because it is exactly the tier that model-generated labelling has been eating.

The second is reference-answer writing, the data behind supervised fine-tuning. Instead of grading the model, you write the response the model should have produced: the correct answer, in the required format, with working shown where working is expected. It is graded on factual accuracy, instruction compliance, formatting discipline and originality, and submissions are screened for having been written by a model rather than by you. Good writers who verify their claims do well here and are genuinely scarce.

The third is domain expert grading and question authoring, and it is where the money went. A licensed physician reviews clinical reasoning in a generated differential. A practising attorney checks whether a cited authority exists and whether it says what the model claims. An accountant grades a consolidation entry. A security researcher writes an exploit chain the model should refuse or should solve. The purpose is no longer to catch obvious errors, because there are few left in these domains. It is to construct problems at the frontier of a specialty and to say precisely why an answer that looks authoritative is wrong.

The fourth is environment and verifier building, which is the newest and the best paid. You author a task, the scaffold it runs in, and a programmatic check that decides pass or fail without a human in the loop: a unit test, a database state assertion, a diff against an expected artefact, a tolerance band on a numerical answer. This sits between annotation and software engineering, and it is the clearest upgrade path available to anyone already doing rating work who can write a little Python.

The fifth is media annotation, a much older lineage with its own tools and employers: bounding boxes, polygons, semantic segmentation and LiDAR point-cloud labelling for autonomous vehicles and robotics; transcription, speaker diarization and pronunciation tagging for speech; contouring and lesion marking for medical imaging, which is credential-gated; document layout, table and handwriting labelling for document understanding; and screen-recording trajectories for computer-use agents. These roles are more often actual employment with a shift pattern, more often on site, and more often measured on throughput than on judgement.

Two further lineages deserve naming because they recruit constantly and are frequently mistaken for the others. Search quality rating predates the current wave by well over a decade, runs through outsourcing vendors, and is rubric-heavy, part-time and tightly audited. Safety and red-team work asks you to attack a model deliberately, and some of it involves reading or producing material you will not enjoy reading or producing. That is a real occupational hazard rather than a footnote, and it is covered later in this guide.

• What a generalist shift actually produces: a run of preference comparisons each with a written rationale, a handful of items flagged because the guidelines cannot resolve them, and a note to the project channel proposing a rule for them.
• What an expert-tier shift produces: a small number of hard problems in your specialty with reference solutions, a rubric for partial credit, and a short memo on the failure pattern you found. Volume is low and the rate per item is high, which inverts the generalist incentive.
• What an environment builder produces: a task definition, a seeded environment, a verifier script that passes on the reference solution and fails on plausible wrong ones, and a README explaining what the task is testing.
• Modalities currently hiring: text, code, voice and speech, images, video, LiDAR and point clouds, documents and OCR, and screen or browser trajectories.
• Task types worth naming by their real names on an application: preference pairs, pairwise ranking, rubric scoring, single-response grading, SFT reference writing, prompt authoring, adversarial probing, trajectory grading, step-level attribution, taxonomy labelling, entity and relation annotation, segmentation, diarization.
• What the job is not: teaching a model anything in the everyday sense of teaching, choosing what the model learns, having any visibility into whether your data was used, or working for the lab whose name is on the model.
• A reliable test of which tier a posting belongs to: look at what it asks you to prove. Availability and typing speed means tier one. A licence number means tier three. A GitHub link means tier four.

Who hires AI trainers, and how the money reaches you

Understanding the money flow explains almost everything else about this work: the rates, the churn, the NDAs, and the fact that the company whose model you improve is not your employer. A model lab or product company buys human data on a contract with a volume and a deadline. A data vendor or expert marketplace wins that contract and takes a margin. You are sourced, screened, onboarded and paid by the vendor, usually as a contractor, and the client's name may never appear in anything you sign. When the client reduces the order, your queue empties the same week, and nobody in the chain owes you an explanation.

Crowd platforms and data vendors are the largest employers by headcount. Outlier, operated by Scale AI, recruits at volume across generalist and expert projects. Surge AI, Invisible Technologies, Toloka, Appen, TELUS Digital, Sama, iMerit, Alignerr (run by Labelbox) and DataAnnotation.tech all run continuous open applications. What they buy is capacity against a guidelines document: the project is defined elsewhere, and your job is to apply it consistently and quickly. Expect long guidelines, frequent calibration tests, and audits.

Expert marketplaces are the fastest-growing buyer and they recruit differently. Mercor, Handshake AI, Micro1, Turing and Snorkel source credentialed professionals directly, often by approaching people on the strength of a degree, a licence, a publication record or a GitHub history, and they frequently screen with a recorded or AI-conducted interview before a human recruiter is involved. The engagement is typically part-time, remote, hourly, and billed against your professional identity rather than a generic rater pool. If you hold a licence or an advanced degree, this is the door to use, and applying to the generalist platforms instead is leaving most of your rate on the table.

In-house teams at labs and product companies are small, salaried, and real. The titles to search are human data specialist, data operations analyst, annotation program manager, AI tutor, model behaviour specialist, evaluation operations, and taxonomy or ontology manager. These roles do less annotation and more design: they write the guidelines the vendor pool applies, decide the sampling, audit the vendor's quality, and sit between the research team and the delivery manager. Most of the people in them came from the vendor side, which is the main reason vendor work is worth doing well rather than quickly.

Companies with their own annotation operations form a separate market: autonomous-vehicle and robotics firms, speech and voice-assistant teams, medical-imaging and digital-pathology companies, document-automation companies, and defence and geospatial contractors. These jobs are more often employment with benefits, more often on site for data-security reasons, and sometimes require a security clearance or health-data training. Be aware that some of this capacity has been cut rather than grown as automatic pre-labelling improved: Tesla's reductions at its Buffalo data-annotation site were publicly reported, and the general pattern is that machines now do the first pass and humans correct it, which needs fewer people with better judgement.

The volatility of this sector is structural, and the right response is to plan around it rather than be surprised by it. A single client contract can support thousands of engagements and end them all at once. Appen's loss of its Google work in 2024 was widely reported and is the clearest public example, and reporting in 2025 on labs moving work away from Scale AI after Meta took a large stake in it showed the same thing from the other direction: the biggest vendor in the market is not a safe harbour either. People who last in this field run two or three platform relationships simultaneously, keep their own records of hours and quality scores because platform dashboards disappear with account access, and treat any one queue as a revenue line rather than a job.

Finally, a warning that belongs in the hiring section rather than a footnote, because the volume of fraud around this specific job title is high. Legitimate AI trainer work never asks you to pay anything: no application fee, no training fee, no "equipment deposit", and no cheque to deposit so you can buy a laptop from a named supplier. Recruiting that starts and stays on Telegram or WhatsApp, pays in cryptocurrency, promises a fixed daily figure for "simple data tasks", or sends an offer before any assessment, is a scam. Apply through the vendor's own domain that you reached yourself, not through a link in a message.

• Crowd platforms and vendors to apply to directly: Outlier, Surge AI, DataAnnotation.tech, Invisible Technologies, Toloka, Appen, TELUS Digital, Sama, iMerit, Alignerr, Prolific (research studies rather than training data).
• Expert marketplaces that recruit credentialed professionals: Mercor, Handshake AI, Micro1, Turing, Snorkel.
• In-house titles worth a saved search: human data specialist, data operations analyst, annotation program manager, AI tutor, model behaviour specialist, evaluation operations specialist, taxonomy manager, linguist, computational linguist.
• Sector employers with their own annotation teams: autonomous-vehicle and robotics companies, speech and voice-assistant teams, medical-imaging and pathology companies, document-automation vendors, geospatial and defence contractors.
• How work is actually found: open application forms on the vendor's site, LinkedIn and GitHub outreach from marketplace sourcers, university and professional-association mailing lists for expert projects, and referrals from people already on the platform.
• What to check before accepting: the published rate, whether it is hourly or per task, whether review and calibration time is paid, the payment processor and cadence, who to contact when a payment is wrong, and whether the project is region-locked.
• Red flags that reliably indicate fraud: any payment from you, cheque deposits, crypto wages, hiring without an assessment, a company name that does not match the email domain, and pressure to move to a chat app.

Contract, staff, or neither: what you are actually signing

Most AI trainer work in the United States is independent contracting, and the practical consequences are worth stating flatly before you start. No employer withholds tax, so you are responsible for income tax and self-employment tax and typically for quarterly estimated payments. There is no unemployment insurance when the queue dries up, no paid sick leave, no notice period, and no guaranteed hours. Access can be withdrawn without a stated reason, and the appeal route is a support ticket rather than a process. Whether this classification is correct for highly supervised, metric-managed work is genuinely contested, and misclassification claims have been brought in this sector, but plan around the status you are offered rather than the status you think is fair.

Pay mechanics vary more than the headline rate and determine what you actually earn. Some projects pay an hourly rate against time you log in a platform timer, with a cap per task and an expectation that you stop when you hit it. Others pay a piece rate per item, which converts to a much lower effective hourly rate on any task that needs real verification, because the time to check a citation is not the time to tick a box. Ask three questions before accepting: is reading the guidelines paid, is calibration and re-training paid, and what happens to pay if a reviewer rejects an item you submitted in good faith. Assume the qualification assessment itself is unpaid unless the posting says otherwise, because it usually is.

Payment runs through processors rather than payroll, commonly PayPal, Wise, Tipalti, Airwallex or direct bank transfer depending on your country, on a weekly or fortnightly cycle. Non-US contractors should expect a W-8BEN form, currency conversion costs, and a posted rate that differs for the same task by market. Keep your own ledger of hours and submissions from day one. When a payment is short or an account is suspended, your own records are the only evidence you will have, because the platform dashboard goes away with the account.

Non-disclosure agreements are standard and they shape how you can job-search later. You generally may not name the client, quote prompts or outputs, screenshot the tool, or say which model you worked on. You can describe the work by modality, domain, task type, volume and quality metric, which is actually the more persuasive way to describe it. Observing the NDA carefully is itself a hiring signal, because the people recruiting for in-house human-data roles have all seen candidates leak a client name in an interview and have all declined them for it.

Quality management is continuous and is the main cause of lost access. You will have a rolling quality score from audits, gold-standard items seeded into your queue, and agreement checks against a reviewer. Scores move fast and recover slowly. The practical defences are simple: read the guideline updates when they land rather than when you fail an audit, ask about any item you were marked wrong on, and never increase speed at the expense of the rationale field, because the rationale is what the auditor reads first.

Content exposure deserves an honest paragraph. Safety, red-team and moderation-adjacent projects can require you to write or read material about violence, self-harm, sexual abuse and extremism, and the psychological cost of sustained exposure is documented rather than theoretical: litigation brought in Kenya by workers doing content moderation and data work for outsourcing vendors has been running for years and is a matter of public record. Before accepting this kind of project, ask what content categories are in scope, whether you can opt out of specific categories, what the shift limits are, and what mental-health support is provided and by whom. Declining the project is a legitimate choice and a reputable vendor will not penalise you for asking.

Geography sets your ceiling more than your skill does at the generalist tier, and less than your credential does at the expert tier. The same labelling task is priced differently in different markets, some projects are locked to a country for legal or language reasons, and a few require you to work from a secure facility. The way out of the geographic ceiling is not to argue with it. It is to move up a tier, into work priced on a credential, a language pair or a verifier you can write, where the rate is set by scarcity rather than by local clerical wages.

• Set aside money for tax from the first payment and check your jurisdiction's rules on estimated or advance payments. Contractor income arriving gross is the most common financial surprise in this work.
• Keep a personal ledger: date, project, hours logged, items submitted, rate, amount paid, audit score, and any reviewer feedback. One spreadsheet, maintained weekly.
• Ask whether guideline reading, calibration tests and re-training are paid time. On some projects they are, and on some they quietly are not.
• Expect a per-task time cap and understand what happens if you exceed it. Working unpaid past the cap to protect your quality score is how people end up earning well below the posted rate.
• Keep NDA-safe language ready: "a US model developer, via a data vendor" rather than a client name; "multi-turn dialogue preference data in clinical domains" rather than a product name.
• On safety projects, ask about content categories in scope, opt-out rights, shift caps, and who provides counselling. Get the answer before you accept.
• Never depend on a single queue. Two or three active platform relationships is the difference between a bad month and no income.
• If a platform deactivates you, request the reason in writing once, keep the response, and move on. The time spent fighting a deactivation is almost always worth more applied to the next project.

How hiring works, and what the qualification assessment really tests

For most AI trainer work there is no interview in the ordinary sense, and candidates who prepare as though there is one prepare for the wrong thing. The decision is made by a scored assessment and by your first week of audited output. Nobody asks where you see yourself in five years. The sequence is an application form, sometimes an automated screen, a timed qualification assessment, then onboarding paperwork, and the whole thing can run inside a week. Treat the assessment as the hiring process, because it is.

The platform path, stage by stage, looks like this. You complete a web application with a resume upload, a declaration of domains and languages, and often a short free-text answer. Some marketplaces then run a recorded or AI-conducted interview, usually short. You are given a guidelines document to read and a timed assessment, most often under two hours. If you pass, you sign an NDA and a contractor agreement, submit a tax form, set up payment, complete platform training modules, and are added to a project pool. Then you wait for the queue to open, which can take days or months and is the stage nobody warns you about.

Scoring combines three things and it is worth knowing what each is for. Agreement with gold-labelled items tests whether you can apply a rule. Agreement with a human reviewer across the whole set tests whether you are calibrated to the project rather than to yourself. A human then reads a sample of your rationales, and this is the part that separates candidates, because two people can submit identical scores and only one of them explains the decision in a way a stranger could act on. Timing telemetry is also recorded, and an unusually fast perfect submission is treated as suspicious rather than impressive.

Expert marketplaces increasingly screen with an AI-conducted interview before any human is involved, and Mercor and Micro1 are both open about using them. Treat it as a real interview. Sit somewhere quiet with a working camera, answer in complete spoken sentences, and when asked about your specialty, talk about where current models fail in it rather than reciting what you know. Name a specific case: the drug interaction a model gets wrong, the jurisdictional rule it flattens, the proof step it skips, the API it hallucinates. That answer is the one the project is actually buying, and it is also the answer a generic candidate cannot give.

Credential verification is real at the expert tier and is where misrepresentation ends an application permanently. Expect licence-number lookups against public registers, diploma or transcript upload, employer verification, identity documents matched against a live selfie, and in technical domains a short screen with a human expert. Account farming and credential sharing are common enough that platforms have become strict about this. Declare licence status precisely, including state or jurisdiction and whether it is active, inactive or restricted, and never let anybody else use your account.

In-house roles run a conventional loop, and it is writing-heavy because the output of the job is documents. Expect a recruiter screen about domains and volumes you have handled, then a take-home with a predictable shape: here is a batch of messy model outputs and a vague quality goal, write the rubric, grade the set, report what you found, and state what you would change about the task. Then a hiring-manager conversation that is mostly about judgement calls on ambiguous cases, a writing round or a review of your take-home memo, and sometimes a light sampling and statistics round. A few weeks end to end is normal.

A note on silence and reapplication. Platform applications often go unanswered for months and then activate, so do not read silence as rejection. Most platforms allow reapplication after a cooling-off period if you fail an assessment, and some allow you to attempt a different domain immediately. The highest-yield response to a failed assessment is to apply to a different vendor the same week rather than to wait, because the assessments are similar enough that the practice transfers and different projects weight the rubric differently.

The assessment itself looks like a knowledge test and is not one. It tests four things: whether you can apply somebody else's rule exactly, whether you can tell a wrong answer from an answer you merely dislike, whether you verify or assume, and whether you can write a sentence a stranger can act on. Smart, articulate people fail it routinely, and they fail in the same handful of ways. All of them are avoidable if you know they are coming.

Guideline fidelity is the first and the one most people underestimate. The guidelines will contain at least one rule you disagree with, and at least one item constructed so that your instinct and the rule point in opposite directions. The rule wins. The reason is not bureaucracy: a dataset where every annotator applies their own taste is noise, and the whole purpose of a guideline is to make many people produce the same label. If you think a rule is wrong, apply it and flag it, with the item id and a proposed alternative. That behaviour is rewarded. Quietly overriding it is marked wrong even when your judgement was better than the rule.

Confusing style with correctness is the most common failure and the fastest to diagnose. A candidate prefers the response that is better organised, more fluent, longer, or formatted with headings, when the other response is the one that is actually right. Most rubrics are explicitly ordered, with factual accuracy and instruction-following above tone and presentation, and the assessment will contain an item where a polished answer contains an invented citation and a plain answer does not. Pick the plain one. Then say why in the rationale, which is the only way the reviewer can tell you did it deliberately.

Rationale writing is where the human reviewer forms their opinion of you, and it takes about thirty seconds longer than the version that fails. "Response B is better" scores nothing. "Response B is correct: Response A cites a clinical guideline that does not exist, and the dose it gives is above the licensed maximum for this indication, which fails factual accuracy regardless of A's clearer structure" scores. The test is simple and you can apply it yourself. Could somebody who cannot see the two responses act on what you wrote. If not, rewrite it.

Verification discipline is seeded deliberately. Assessments contain claims that are plausible, specific and false: a statute section that does not exist, a function that was never in the library, a study with real authors and a fabricated finding, an arithmetic result that is close but wrong. Candidates who read for fluency miss all of them, and candidates who check catch them. In your own specialty, check the things that are checkable, and in the rationale say what you checked, because "the cited section does not exist in the current code" is evidence that you looked.

Calibration and scale use decide whether your labels are usable. People cluster their scores in the middle of the scale, never award the bottom score, and avoid declaring ties because a tie feels like a non-answer. All three make a dataset less useful. Use the whole range, award the bottom score when the rubric says the item earns it, and mark a genuine tie as a tie with a sentence explaining what would have broken it. Where the interface forces a choice, say in the rationale that the difference was marginal and name the marginal factor.

Escalation is the behaviour that separates a competent rater from a future lead, and almost nobody does it on an assessment. Some items cannot be labelled correctly under the guidelines as written, because the guidelines did not anticipate them. The weak move is to guess and move on. The strong move is to label it as best you can, then write one sentence: this item is not covered, here is the ambiguity, here is the rule I would add, and here is how I would label it under that rule. Project leads keep the people who do this, because the hardest part of their own job is finding the holes in their guidelines.

Finally, the behaviour that ends engagements fastest: using a model to produce your submissions. It is detected, routinely, through timing telemetry, paste patterns, stylometric consistency, honeypot prompts designed to produce a recognisable model answer, and the simple fact that model-written rationales are generically fluent and specifically empty. It also destroys the value of the work, because a dataset of model output graded by model output teaches the model nothing it does not already believe. You are being paid for human judgement. Using tools to look something up is fine and often expected. Having the tool write the answer or the rationale is fraud, and it gets accounts terminated with payment withheld.

• Platform timeline: application, an optional recorded interview a few days later, an assessment you sit in one session, onboarding paperwork over a few days, then queue access that may take days or months.
• In-house timeline: recruiter screen, a take-home of a few hours, hiring manager interview, writing round, decision. A few weeks.
• Before the assessment: read the entire guidelines document once, then re-read the scoring section and the edge-case section. Most failures are guideline failures rather than knowledge failures, and the scoring hierarchy is the part to keep visible while you work.
• Rank on correctness first, then instruction-following, then safety and tone, unless the rubric says otherwise in writing.
• Write every rationale so it is intelligible without the item: name the specific error, name the rule it breaks, and name why the competing virtue does not outweigh it.
• Check what is checkable: citations, section numbers, drug doses, API signatures, arithmetic, dates, units.
• Use the whole scale. Award the lowest score when earned and mark true ties as ties with a reason.
• Flag items the guidelines cannot cover with a proposed rule rather than guessing silently. One line is enough, and it is the behaviour that gets you noticed for a lead role.
• Stay inside the time cap and note where the cap forced a shallower check, rather than working unpaid to hide it.
• Have ready for an expert interview: three examples of where current models fail in your specialty, and how you would tell a wrong answer from a right one in each case.
• Have ready for an in-house take-home: a rubric with explicit tie-breaking rules, a breakdown of results by failure category rather than one average, and a named limitation of your own method.
• Declare credentials exactly: "Registered nurse, active licence, Ohio" or "Admitted to the New York bar, active". Vague claims get checked and fail.
• If you fail, apply to two other vendors within the week rather than waiting out a cooldown with one.
• Never let a model write your submission or your rationale. Detection is routine and the penalty is termination with pay withheld.

The resume, the profile and the portfolio: what lands and what is ignored

Be clear about where a resume matters in this field, because effort spent in the wrong place is wasted. For generalist platform work the resume barely matters: the assessment decides, and the form exists mostly to route you to a domain. For expert marketplaces it matters a great deal, because a human or automated sourcer is matching your credential against a project and will never see you work. For in-house human-data roles it matters in the ordinary way, and the portfolio matters more. Write one resume that serves the second and third cases, and do not agonise over the first.

Put the credential first, stated precisely. Profession, qualification, jurisdiction, status, and years in practice, in the top third of the page. "Registered nurse, active licence, Ohio, eleven years, medical-surgical and critical care" is worth more than any paragraph about enthusiasm for artificial intelligence. For languages, use a real scale rather than adjectives: CEFR levels or ILR ratings, with the dialect or variety where it matters, because "fluent Spanish" tells a localisation project nothing and "Spanish, native, Rioplatense; CEFR C1 Portuguese" tells it everything.

Describe the work in the field's own vocabulary, because both human and automated screening match on it. "Reviewed AI answers" is invisible. "Graded multi-turn dialogue against a nine-point rubric; authored preference pairs for supervised fine-tuning; red-teamed refusal behaviour in a clinical domain; annotated agent trajectories with step-level attribution for tool-call errors" is legible to anybody who buys this work. Name the modality, the task type, the domain, and the volume.

Use numbers you can actually defend. Items completed per hour and total volume on a project. Your audit or quality score and the threshold it was measured against. An inter-annotator agreement figure you helped move, with the metric named (Cohen's kappa for two raters, Fleiss' kappa or Krippendorff's alpha for more, or percentage agreement if that is genuinely what was measured). The number of guideline revisions you authored or proposed. The number of reviewers you trained or whose work you audited. If you do not have these figures, start recording them this week, because they are the only quantitative evidence this job generates.

List tools, because projects are staffed partly on tool familiarity. Annotation and data platforms: Label Studio, CVAT, Labelbox, SuperAnnotate, V7, Encord, Prodigy, Doccano, Argilla, Dataloop, Snorkel. For environment and verifier work, the list that matters is different: Python, pytest, git, JSON and YAML, regex, SQL, Docker basics, and whichever harness the project uses. A trainer who can write a passing and a failing test case for their own task is in a different and better market than one who cannot.

Respect the NDA in the resume itself, and make a virtue of it. "Expert rater, clinical domain, via a US data vendor for an undisclosed model developer" is both compliant and credible. Recruiters in this field read it as professionalism, whereas a client logo you were not permitted to name reads as a liability. If you are asked directly in an interview, say what your agreement allows and stop there. Nobody good will push.

Delete the following, all of which actively hurt. Generic enthusiasm ("passionate about shaping the future of AI"). Any claim that you trained a named model, because you cannot know that and the agreement probably bars the claim. A list of eight platforms with no outcome attached, which reads as churn rather than experience. Prompt-engineering certificates and AI course-mill credentials, which carry no weight with the people who hire for this work and in some cases count against you. And a resume that was obviously written by a model, which in a field whose entire business is detecting model output is the worst possible opening argument.

The portfolio is the thing that moves you off piecework, and almost nobody building one is competing with you. Three artefacts are enough, all built on public data so no agreement is breached. First, an annotation guideline you wrote for a small, genuinely ambiguous task, including the edge cases and the tie-breaking rules. Second, a small labelled set, a hundred items or so, labelled by you and one other person, with the agreement figure and an honest account of where you disagreed and what you changed. Third, a verifier: a task with a reference solution and a test that passes it and rejects three plausible wrong answers. Put them in a public repository with a short README. That package is the single most effective application material in this field, and it is also precisely what an AI evaluation hiring manager at a product company is looking for.

• Top of resume: credential, jurisdiction, licence status, years in practice, specialty, languages with CEFR or ILR levels.
• Task vocabulary that screens well: preference pairs, pairwise ranking, rubric scoring, SFT reference answers, RLHF annotation, red-teaming, adversarial prompting, trajectory and tool-call grading, taxonomy and ontology labelling, segmentation, diarization, search relevance rating.
• Defensible numbers: items per hour, total volume, audit score against threshold, agreement metric and movement, guideline revisions authored, reviewers trained or audited.
• Tooling to list if true: Label Studio, CVAT, Labelbox, SuperAnnotate, V7, Encord, Prodigy, Doccano, Argilla, Dataloop, Snorkel, plus Python, pytest, git, regex, SQL and JSON for verifier work.
• NDA-safe client phrasing: "a US model developer, engaged through a data vendor", "an enterprise search provider", "a healthcare AI company".
• Delete: passion statements, claims about named models, bare platform lists, prompt-engineering certificates, and anything a model wrote for you.
• Portfolio, three items: an annotation guideline for an ambiguous task, a small double-labelled set with an agreement figure, and a verifier with a reference solution and failing cases.
• One line that works in a cover note: "I can write the rubric, measure whether two people apply it the same way, and tell you which slice of your data is broken."

What this work pays, and how to get numbers you can trust

There is no honest single answer to what an AI trainer earns, and any article that gives you one number is guessing. The same job title covers piece-rate labelling priced against local clerical wages and expert grading priced against a specialist's consulting rate, in markets that are not within shouting distance of each other. What is worth knowing is the structure of the pay, which determines what you can do about it, and the sources that will give you a real figure for your own situation today rather than a stale band from an article.

The generalist tier is priced as flexible remote piecework. Platforms publish a rate on the project before you accept it, which is the single most reliable number available to you, and in the United States it tends to sit in the range of other flexible remote clerical work rather than in technology salaries. In lower-cost markets the posted rate for the identical task is substantially lower, which is a deliberate feature of how this labour is sourced rather than an error. Two things quietly reduce what you keep: piece rates on tasks that require real verification, and unpaid time.

The expert tier is priced on scarcity and is benchmarked against your profession. A platform buying a licensed physician's judgement is competing with what that physician could earn clinically in the same hour, and the rate reflects it, which is why expert-tier hourly rates are a multiple of generalist rates rather than a premium on them. What moves your rate inside this tier is the depth and verifiability of the credential, the scarcity of the specialty, your willingness to take the hardest frontier tasks rather than the easy volume, and whether you can author problems rather than only grade them.

Environment and verifier building, and the lead and reviewer roles above the annotation pool, form a third tier. These pay more than generalist annotation because fewer people can do them and because the output has longer-lived value: a verifier keeps working after you stop, and a guideline you wrote governs thousands of items labelled by other people. If you are currently in the generalist tier and want a raise, the shortest route is not to label faster. It is to become the person who writes the rule or the test.

Use these sources, in this order, for an actual number. The project rate published on the platform before you accept. Salaried postings for in-house roles in pay-transparency jurisdictions such as California, Colorado, New York, Washington and Illinois, which must state a range and which also tell you what the work is called internally. Rates other contractors report on the platform's own community forums and in the public subreddits for each platform, which are current in a way no published band is. The BLS OES codes as a floor reference only, with their caveat: no code covers AI trainer work, and 43-9021 Data Entry Keyers and 43-9111 Statistical Assistants describe adjacent clerical and research-support work rather than this. For expert work, the OES code of your own profession is the more meaningful benchmark.

Track your effective rate rather than your posted rate, because they differ and the gap is where people lose money without noticing. Count every hour you spent: reading the guidelines, sitting the calibration test, re-reading an updated rubric, waiting for a queue, and redoing rejected items. Divide actual payments received by total hours spent, including the unpaid ones, over a month. That number is your real wage, and it is the only number on which to decide whether a project is worth continuing.

Negotiation exists at the expert tier and essentially does not at the generalist tier. Platform rates for pooled work are set by the project and the support agent cannot change them. Expert engagements, especially those sourced by a human recruiter against a named project, are negotiable on rate, on hours, and sometimes on scope. Your leverage is a verified scarce credential, demonstrated ability to author hard problems rather than grade easy ones, and a track record of high audit scores you can state with numbers. Ask for the band in writing before the assessment, and ask what the path to a higher tier on the same project looks like.

• Most reliable number available: the per-project rate the platform publishes before you accept the task. Read it every time, because it changes between projects.
• Second source: salaried in-house postings in pay-transparency jurisdictions, which must publish a range and reveal the internal job title.
• Third source: current rates reported by contractors in the platform's own forums and community subreddits.
• Official reference points, with caveats: BLS OES 43-9021 Data Entry Keyers and 43-9111 Statistical Assistants as adjacent floors; your own profession's OES code for expert work.
• Pay structures to distinguish: hourly against a logged timer, per-item piece rate, per-item with a time cap, and project bonus for throughput or quality.
• Hidden unpaid time to count: the qualification assessment, guideline reading, calibration tests, re-training after a rubric change, queue waiting, and rework on rejected items.
• What actually raises your rate: a verified scarce credential, authoring rather than grading, writing verifiers, and moving into reviewer or lead work.

Does AI trainer work lead anywhere? Five real exits and one that is not

The honest answer has two halves and most discussion of this job only reports one of them. Logged annotation hours, by themselves, lead nowhere. There is no seniority ladder inside a pooled queue, the work is deliberately interchangeable, and five years of it looks on paper almost the same as six months of it. But the skills the work can build, if you deliberately build them, are in demand and are genuinely scarce. The difference between the two outcomes is whether you leave with artefacts and measurements or with a timesheet.

The first real exit is up the vendor ladder, and it is the most common. The progression runs rater, reviewer, quality analyst, guideline or taxonomy author, project lead, delivery manager, program manager. Vendors promote from their own pool because the pool is where the people who understand the work are, and the selection criterion is visible: who flags the ambiguity, who proposes the rule, who writes a clear escalation, who other raters ask questions. Make yourself that person deliberately and say so when a lead role opens.

The second exit is AI evaluation and AI quality work at product companies, and this is the highest-value jump available to a strong generalist. What transfers directly is the thing those teams are short of: rubric writing, measuring whether two people apply a rubric the same way, sampling, and the willingness to read several hundred real outputs by hand and name a failure pattern. What you must add is enough Python to own a harness, basic statistics for sampling and uncertainty, and the habit of writing a memo a decision maker can act on. The portfolio described earlier is the bridge.

The third exit is into the expert and environment tiers of the same market, at a multiple of the generalist rate. For a credentialed professional this is often the best move available and it requires no career change at all: you stay in your profession and sell your judgement part-time at a scarcity price. For a technically inclined generalist it means learning to write verifiers and author tasks, which is a focused few weeks rather than a retraining programme, and which puts you in the part of this market that is growing.

The fourth exit is a human-data or model-behaviour role inside a lab or large product company: data operations, annotation program management, model behaviour and policy, evaluation operations. The number of these seats is small and they are gated on writing, because the job is producing the documents that govern how thousands of other people label. People get them by being visibly excellent on the vendor side, by having authored guidelines somebody can read, and by writing clearly in public or in the application itself.

The fifth exit is the one people overlook and it is frequently the most valuable: back into your own profession with a credible, specific AI specialism. The clinician who has graded a few thousand model outputs in their specialty knows exactly how these systems fail in it, which is precisely what a hospital evaluating a documentation or triage tool needs and cannot buy. The same holds for the lawyer setting review standards for AI-assisted drafting, the accountant assessing an audit tool, and the teacher advising on assessment integrity. That expertise is rare, it is paid as professional work rather than as data work, and the trainer work is how you acquired it.

The exit that is not real, stated plainly because it is sold constantly: AI trainer work does not convert into a machine learning engineer or research scientist job. Grading model outputs teaches you nothing about training models, the hiring bars for those roles are software engineering and mathematics, and no volume of annotation hours substitutes. If that is your goal, the honest route is the ordinary one: programming ability, systems, statistics, and shipped projects, with the trainer work funding the study rather than replacing it. Be equally sceptical of "prompt engineer" as a destination title, which has largely been absorbed into ordinary product and engineering work.

• Vendor ladder: rater, reviewer, quality analyst, guideline and taxonomy author, project lead, delivery or program manager.
• Product-company exit: AI evaluations engineer, AI quality analyst, evaluation operations, trust and safety analyst. Transfers rubric and agreement skills; requires Python and basic statistics.
• Same-market upgrade: expert-tier grading, problem authoring, environment and verifier building, at a multiple of generalist rates.
• Lab-side roles: human data specialist, annotation program manager, model behaviour and policy, data operations. Few seats, writing-gated.
• Return to profession: the clinician, lawyer, accountant or teacher who now owns AI tool evaluation and standards where they work.
• Not a path: machine learning engineer or research scientist, which require software engineering and mathematics regardless of annotation experience.

Your first sixty days, in order

Everything above is advice. This is the sequence, because the common failure is not ignorance of what to do but doing it in an order that wastes the first two months. Treat it as a checklist with dates against it.

Week one is applications, and breadth matters more than polish because the assessment decides. Write one resume with the credential and jurisdiction in the top third and the task vocabulary in the body. Then apply to three generalist platforms and, if you hold a licence or an advanced degree, to all of the expert marketplaces, because they are not substitutes for each other and an expert application to a generalist pool is priced as generalist work. Use the vendor's own domain each time. Expect most applications to go quiet.

Week two is the assessment, and it is worth preparing for exactly once because the preparation transfers to every other vendor. Read a published rater guideline end to end so the genre is familiar before you are timed on it: Google's Search Quality Rater Guidelines are public, long, and structurally the same animal as a preference-rating guideline. Then sit whichever assessment lands first, with the scoring hierarchy open beside you, and write full rationales even when the field looks optional.

Weeks three to six are the queue, which may be empty, and the right response to an empty queue is to build the portfolio rather than refresh the dashboard. Pick a small ambiguous labelling task on public data: whether a product review is sarcastic, whether a news headline is an opinion, whether a support ticket is a bug or a request. Write the guideline, with definitions, a decision procedure, worked near-misses and tie-breaking rules. Label a hundred items. Get one other person to label the same hundred. Compute Cohen's kappa, see where you disagreed, revise the guideline, relabel the disputed items and report both figures.

Weeks six to eight are the verifier, which is the highest-return week in this plan and the one most people skip because it sounds like programming. It is less than that. Take a task with a checkable answer, write the reference solution, then write a pytest file that passes on the reference and fails on three wrong answers you construct deliberately: one wrong value, one right value in the wrong format, one that hard-codes the expected output. If you can explain in a README what the third test catches and why it matters, you are describing reward hacking, and you are now qualified to talk to the part of this market that is hiring.

At day sixty you should have, regardless of whether any queue ever opened: a resume written in the field's vocabulary, applications live at several vendors, one or two assessments sat, a published guideline, a double-labelled set with a named agreement metric and a revision history, and a verifier in a public repository. That is a portfolio, and it is the thing that converts platform work into a reviewer role, an expert engagement or an AI evaluation job. Hours logged convert into nothing, which is why the plan starts producing artefacts in week three rather than waiting for permission.

• Week 1: one resume, credential and jurisdiction in the top third; apply to three generalist platforms and every expert marketplace you qualify for, through their own domains.
• Week 2: read one long public rater guideline end to end, then sit the first assessment with the scoring hierarchy open and write full rationales.
• Weeks 3 to 6: write a guideline for an ambiguous public task, label a hundred items, have one other person label the same hundred, compute Cohen's kappa, revise, relabel, report both numbers.
• Weeks 6 to 8: write one verifier with a reference solution and three deliberate failing cases, and a README explaining what it cannot catch.
• Throughout: keep the ledger (date, project, hours, items, rate, paid, audit score) and a running list of every ambiguity you hit with the rule you would add.
• Do not: wait for a queue before building anything, pay for a certificate, or let a model write a submission, a rationale or your resume.

What a screen is looking for

These are the terms that a resume screen, human or automated, is matching against for this role. Use the ones that are true of you, in the words the posting uses.

• AI trainer
• AI tutor
• AI coach
• Data annotator
• Data labeller
• Data labeling specialist
• Expert rater
• Search quality rater
• Human data specialist
• Human data operations
• RLHF annotator
• Subject matter expert contributor
• AI data trainer
• Model trainer
• Annotation specialist
• Annotation analyst
• Linguistic annotator
• Content rater
• Quality rater
• Supervised fine-tuning
• SFT data
• Reference answer writing
• Preference pairs
• Pairwise ranking
• Preference data
• Reward model
• RLHF
• RLAIF
• Direct preference optimisation
• DPO
• Reinforcement learning from verifiable rewards
• RLVR
• Rubric-based reward
• Verifier
• Reward hacking
• Rubric design
• Rubric scoring
• Annotation guidelines
• Guideline authoring
• Taxonomy design
• Ontology labelling
• Gold set
• Held-out set
• Ground truth labelling
• Labelled dataset
• Inter-annotator agreement
• Inter-rater reliability
• Cohen's kappa
• Fleiss' kappa
• Krippendorff's alpha
• Calibration
• Audit sampling
• Quality assurance review
• Model judge
• LLM-as-a-judge
• Judge validation
• Model evaluation
• LLM evaluation
• Eval set
• Benchmark contamination
• Red teaming
• Adversarial prompting
• Jailbreak testing
• Refusal calibration
• Safety annotation
• Policy enforcement
• Trust and safety
• Content moderation
• Hallucination detection
• Confabulation
• Sycophancy
• Factual accuracy review
• Instruction following
• Agent trajectory annotation
• Tool call grading
• Step-level attribution
• Computer use annotation
• Screen trajectory labelling
• RL environment design
• Prompt authoring
• Prompt engineering
• Multilingual annotation
• Localisation quality
• Transcription
• Speaker diarization
• Speech annotation
• Image annotation
• Bounding boxes
• Semantic segmentation
• LiDAR point cloud annotation
• Medical image annotation
• Document layout annotation
• OCR annotation
• Named entity annotation
• Relation extraction
• Label Studio
• CVAT
• Labelbox
• SuperAnnotate
• Encord
• V7
• Prodigy
• Doccano
• Argilla
• Snorkel
• Dataloop
• Python
• pytest
• SQL
• Regex
• JSON
• Git
• Independent contractor
• Remote contract work
• Freelance AI work
• Clinical subject matter expert
• Legal subject matter expert
• CEFR language proficiency

Mistakes that cost people this job

Questions people ask