AI Training Work Splits the Pay Band

On June 24, 2026, a garment worker in Gurugram became part of the AI labor market without receiving an AI job title.

The Guardian reported that workers in Indian factories were wearing cameras to record ordinary physical work: folding towels, stitching garments, arranging objects, moving through production lines. Those recordings, known as egocentric data, are used to train robotics and AI systems that need first-person footage of how human work actually happens. Several companies told the Guardian that factories were compensated for access, while workers often received no direct payment for producing the footage.

One week earlier, Anthropic had published a very different labor-market signal. Its Claude Corps program would pay early-career fellows a full-time salary of $85,000 plus benefits to help nonprofits use Claude in real work. Fellows would receive training, a token budget, mentoring, and project placement. The first cohort would begin in October 2026.

Both cases sit inside the same category that companies are beginning to call AI training work. One worker produces embodied data that may help robots learn factory movement. Another uses a frontier model inside a nonprofit and turns applied AI fluency into a salaried role. Between them sits a wide market of annotators, evaluators, domain experts, coders, medical reviewers, linguists, model testers, AI operations workers, and contractors who label, rank, correct, evaluate, and explain machine output.

One title is too small for the market.

Deel’s Global Hiring Report 2026 says general AI trainer roles grew 283% in cross-border hiring in 2025. The category now spans more than 70,000 workers across more than 600 organizations, from basic annotators to subject-matter experts in medicine and economics. That is no longer a side task hidden in a vendor invoice. It is a labor category.

But it is not one labor market. It is several markets stacked under the same label.

By 2026, the pay band is splitting before the job architecture has caught up. A company may use “AI trainer” for $15-an-hour annotation, $100-an-hour expert feedback, a salaried AI deployment fellowship, an internal evaluator role, or a startup operations job responsible for turning human review into model quality. HR, finance, procurement, and founders cannot price those roles correctly if they treat the title as the unit of analysis.

They need to price the work.

Deel Found the Job Before HR Found the Band

Deel’s 2026 report gives AI training work a scale that was easy to miss when it was scattered across data vendors, contractor platforms, model labs, and offshore service providers.

Deel says AI trainer roles were the fastest-growing cross-border role on its platform in 2025, rising 283%. The work spans more than 70,000 workers and more than 600 organizations. Deel describes a range from basic annotation to subject-matter expertise in fields such as medicine and economics.

That spread matters more than the growth rate. The market is not simply hiring more of one role. It is using one title to describe different kinds of human judgment.

At the low end, the worker may label images, check whether a prompt output follows instructions, categorize text, transcribe audio, or identify objects in a video. The value comes from volume, consistency, speed, and enough attention to keep the data usable. The work may be hourly, task-based, cross-border, and contractor-heavy.

In the middle, the worker may evaluate model responses against a rubric, explain why one answer is safer or more useful than another, create examples, grade code, identify hallucinations, or test whether a model handles a workflow correctly. The value comes from pattern recognition and quality control. The work may still be contractor-based, but it starts to resemble QA, content operations, knowledge management, or product operations.

At the high end, the worker may bring domain expertise. A physician can judge whether a model’s clinical reasoning is plausible. A lawyer can identify missing legal context. A financial analyst can test whether a generated explanation would mislead a client. A bilingual expert can judge translation quality in a domain where literal accuracy is not enough. The value comes from professional judgment, not only from task completion.

One job label covers all three.

This is why a compensation team cannot benchmark “AI trainer” as a single role. The same label can mean low-margin data production, quality evaluation, regulated-domain review, model behavior testing, or applied AI implementation.

A source ledger is already wide enough to force a pay architecture:

Source	2026 signal	Pay-band implication
Deel Global Hiring Report	AI trainer roles grew 283% cross-border in 2025; 70,000+ workers across 600+ organizations	The role is large enough to need a formal job family
Anthropic Claude Corps	Fellows receive $85,000 plus benefits for applied AI work in nonprofits	Applied AI fluency can be a salaried early-career role, not only contractor labeling
Carta AI compensation analysis	Smaller startups raised median AI/ML engineer equity grants sharply from January 2024 to February 2026	AI labor pricing is pulling equity and pay bands upward in startups
PwC 2026 AI Jobs Barometer	AI-skill wage premium reached 62%; AI job postings grew far faster than the overall market	AI capability is now a market-priced skill, but not every AI-adjacent task deserves the same premium
U.S. Department of Labor	2026 proposed rule would revise worker-classification analysis	Contractor-heavy AI training work carries classification risk
Guardian reporting on egocentric data	Workers may generate valuable physical-work datasets without direct compensation	Data contribution and consent need a compensation file, not only a data license
SOMO AI data-worker report	Supply-chain risks include low wages, unpaid labor, unsafe conditions, and weak transparency	Vendor oversight belongs in the AI training budget
Ashby startup hiring report	AI appears in startup job titles and postings; recruiters improve speed in small startups	New AI labor categories enter early-stage companies before process maturity catches up

That is the first warning: AI training work arrived faster than HR naming systems.

Companies have seen this before. Early cloud roles began as “systems administrator with cloud experience” before splitting into platform engineer, site reliability engineer, cloud security architect, FinOps analyst, and developer experience lead. Data roles once collapsed analyst, engineer, scientist, steward, and governance work into one broad label. Over time, the market learned that title inflation hides risk.

AI training work is at that early point now.

Much of the work also sits near the boundaries of employment, contracting, worker data, and product quality. A wrong pay band is not only a retention issue. It can become a wage-and-hour issue, a vendor-risk issue, a worker-consent issue, or a model-quality issue.

Training work teaches the model what good looks like. The company needs to know who taught it, under what terms, and at what price.

Anthropic Put an $85,000 Salary on Applied AI Work

Claude Corps is useful because it pulls AI training out of the contractor-platform stereotype.

Anthropic says each fellowship lasts 12 months. Fellows receive intensive training from Anthropic and CodePath, then spend most of their time embedded with host nonprofits. They receive five hours of ongoing training each week, access to a large Claude token budget, mentorship, and professional guidance. The first cohort of 100 starts in October 2026, with later cohorts planned for January and August 2027. Applicants need to be over 18 and have under two years of full-time work experience; there is no education requirement.

$85,000 is the signal.

Anthropic is not pricing the work like a generic annotation task. It is pricing applied AI deployment, communication, nonprofit workflow understanding, tool-building, and early-career development. The fellow is not training a model in the narrow RLHF sense. The fellow is training an organization to use a model.

That distinction should matter to compensation teams.

The market is beginning to use “AI training” in at least two directions. One direction trains the machine: label this data, rank these answers, record this movement, evaluate this output. Another trains the organization: adapt this tool, build this workflow, help this team use AI without breaking its mission, data practices, or trust with clients.

Both directions need human judgment. They should not sit in the same band.

Claude Corps also shows how a frontier AI company can use labor-market design as distribution strategy. Anthropic is not only funding access to a model. It is funding a cohort of people who learn to apply the model in operational contexts where many nonprofits lack staff capacity. Those fellows become translators between model capability and real organizational work.

That is close to a new role family: AI implementation associate, AI operations fellow, AI workflow analyst, AI enablement specialist, or AI deployment partner.

Titles will vary. The work has recognizable components:

Understand a real organizational process.
Identify where AI can help without creating new harm.
Build or configure a tool around actual users.
Train staff members who do not live inside AI discourse.
Document what worked, what failed, and what should not be automated.
Escalate privacy, accessibility, safety, or data-quality issues.
Convert model capability into repeatable work.

This is not the same as prompt hobbyism. It is not the same as model engineering. It is not ordinary nonprofit program work with an AI tool added to the job description.

It is applied AI operations.

Compensation teams should treat it accordingly. A full-time salaried fellowship with benefits, training, mentorship, and project ownership creates a different market anchor from a task platform that pays by the hour or output. If HR collapses both into “AI trainer,” it will underpay the operational role, overpay some basic tasks, and fail to create a ladder between them.

Early-career design matters too. One day after this site’s article on junior roles, Claude Corps gives a live example of a new bottom rung. Instead of asking an entry-level worker to have years of AI experience, the program defines a structured role where the worker learns by applying AI in a real institution with supervision and training time.

That is a more serious answer than telling graduates to “learn AI.”

It also creates a benchmark. If a lab can fund applied AI work at $85,000 plus benefits in the nonprofit sector, companies that ask internal coordinators, analysts, HR operations employees, or customer support workers to become AI deployment staff will need a pay story. They cannot call the work strategic in the transformation plan and invisible in the compensation plan.

Same Title, Different Labor Market

AI trainer now hides four different labor markets.

Commodity annotation is global, price-sensitive, and exposed to automation pressure. Tasks include labeling, tagging, transcription, simple preference ranking, and data cleanup. The worker may bring care and consistency, but the buyer often prices the work as substitutable capacity.

Evaluation and QA sits closer to product quality. Workers test model outputs, grade responses against rubrics, identify failures, evaluate tool use, compare answers, reproduce errors, and improve prompts or datasets. The buyer needs judgment and reliability, but not always formal domain credentials.

Expert training uses professional knowledge. Medicine, law, finance, engineering, language, education, cybersecurity, and scientific work all require reviewers who know what a good answer means inside the domain. The buyer is no longer paying only for time. It is paying for the right to turn scarce human expertise into model behavior.

Applied AI operations helps an organization use AI in work. It includes workflow design, staff enablement, policy translation, data stewardship, model-output review, exception handling, and adoption measurement. The worker is not simply producing data for the model; the worker is making the model useful in context.

Those four markets have different pay logic:

Labor segment	Main output	Quality risk	Likely employment model	Pay logic
Commodity annotation	Labeled or ranked data at scale	Noise, inconsistency, speed-quality tradeoff	Contractor, vendor, BPO, cross-border platform	Volume rate with minimum labor safeguards
Evaluation and QA	Tested model behavior and failure notes	Hidden model defects, weak rubrics, false confidence	Contractor, internal QA, product ops	Hourly or salaried quality role tied to defect prevention
Expert training	Domain judgment embedded in model behavior	Professional error, liability, bad advice, loss of trust	Specialist contractor, consultant, part-time expert, internal SME	Premium rate tied to credential, domain scarcity, and risk
Applied AI operations	Workflow adoption and human review in context	Bad automation design, privacy failure, mission drift, employee resistance	Salaried role, fellowship, internal mobility path	Job family with level, scope, training, and outcome evidence

A company fails when it pays for the label instead of the layer.

A founder may hear that AI trainers are booming and hire a cheap contractor to build evaluation data for a product that actually needs expert review. A procurement team may buy vendor capacity without understanding whether annotators, domain experts, or internal operators are doing the work. A compensation team may create one AI trainer band and discover that managers are using it for tasks that range from data labeling to regulated-domain signoff.

Budget leakage can run the other direction too. A company may pay a large premium for anyone with “AI evaluator” in the title when the work is actually basic rubric application. That creates internal resentment. Employees doing higher-liability review in adjacent functions may ask why the AI-labeled role receives a premium while their own review work remains inside an old band.

A workable pay-band split needs two dimensions: task complexity and responsibility for consequences.

Task complexity asks what the worker actually does. Responsibility asks what happens if the work is wrong.

If a worker labels images for a low-risk consumer model, the pay question is fairness, minimum wage, working conditions, data rights, and volume. If a worker evaluates whether an AI model gives safe health advice, the pay question includes professional liability, credentials, documentation, auditability, and the cost of a false answer. If a worker helps a nonprofit automate case intake, the pay question includes client trust, privacy, staff adoption, accessibility, and workflow fallback.

One title cannot carry all of that.

Startup Compensation Pulls the Ceiling Higher

AI training work does not sit apart from the broader AI labor market. It is pulled upward by startup compensation pressure.

Carta’s April 2026 analysis of VC-backed startups found that for startups valued between $1 million and $10 million, the median equity grant for AI/ML engineers rose 59% from January 2024 to February 2026. For startups valued between $25 million and $50 million, the increase was 30%. Carta also reported that in 2025, about 40% of every dollar invested in startups on its platform went to an AI company; in early 2026, that share rose to 54%.

This matters for AI training roles even when the workers are not engineers.

As venture-backed AI companies compete for model quality, product reliability, domain data, and deployment speed, they need humans who can improve model behavior. Engineers build the system. Evaluators, domain reviewers, AI operations employees, and deployment specialists decide whether the system works in the world.

That creates a hierarchy problem inside small companies.

A founder may pay a premium for AI/ML engineers because the market forces it. Then the same company may underprice the human evaluation work that determines whether the product can be trusted. In an AI product company, the evaluation layer is not a support function. It is product infrastructure.

Ashby’s State of Startup Hiring report adds a hiring-process view. The report covers more than 1,200 venture-backed startups, 32,000 hires, and 11 million applications. It found that AI appears in roughly one-third of startup job postings and that the percentage of jobs with “AI” in the title doubled from 2% to 4%. It also found that startups with fewer than 25 employees cut time to hire by almost 30% when a recruiter was involved.

Ashby’s recruiter finding is not about AI training directly. It is about maturity. Small startups often create new roles before they have compensation, recruiting, leveling, and interview infrastructure to support them. If a startup has not learned when to involve a recruiter, it is unlikely to have a clean pay architecture for AI evaluator, data-quality specialist, domain reviewer, AI trainer, and AI operations associate.

The first hiring conversation may sound simple:

“We need people to review outputs.”

Harder questions arrive later:

Are these workers employees, contractors, vendors, or experts?
Are they reviewing quality, safety, compliance, tone, factuality, domain reasoning, or customer workflow?
Who writes the rubric?
Who has authority to reject model output?
Who signs off before the work affects a customer, candidate, patient, employee, student, or nonprofit beneficiary?
Does the work create product IP, training data, worker data, or regulated records?
Is pay benchmarked against annotation, QA, domain consulting, product operations, or AI engineering adjacency?

A startup can answer those questions late. It will be cheaper to answer them early.

Compensation pressure also reaches internal workers. A customer success manager may become the person who teaches the AI where customers get stuck. A support lead may build the failure taxonomy. A product operations employee may turn user feedback into evaluation sets. A data quality analyst may become a model behavior reviewer. None of them may receive a new title at first.

That is where salary compression begins.

A startup hires an external AI evaluation lead with a premium because investors expect model quality. The internal employee who built the early evaluation workflow remains in a legacy support or operations band. The founder sees a practical distinction: one person was hired for the new role; the other grew into it. The employee sees a simpler distinction: external AI-labeled work is paid, internal AI work is absorbed.

This is how new job categories create trust problems before they create org charts.

Contractor Status Turns Pay Into a Legal File

AI training work is contractor-heavy enough that pay architecture cannot be separated from worker classification.

On February 26, 2026, the U.S. Department of Labor announced a proposed rule to revise its analysis for distinguishing employees from independent contractors under the Fair Labor Standards Act. The department proposed replacing the 2024 rule with a streamlined analysis designed to provide greater clarity for workers and employers. The proposal also addressed the Family and Medical Leave Act and the Migrant and Seasonal Agricultural Worker Protection Act, both of which incorporate FLSA definitions.

AI training buyers should pay attention because the operating model often depends on flexible labor. Contractors log into platforms, complete tasks, qualify for projects, lose access, retake assessments, submit work, challenge payment decisions, or move across vendors. For companies buying AI training services, that flexibility can make labor feel like cloud capacity. Turn it on, scale it up, change the rubric, shut it down.

People do not become infrastructure because a dashboard treats them that way.

Legal risk is not only whether a worker is called a contractor. It is how the work is controlled, paid, supervised, evaluated, and integrated into the company’s business. AI training work often has tight instructions, quality thresholds, platform monitoring, task assignment, rework requirements, and access controls. Those features may be necessary for model quality. They also make the employment analysis more sensitive.

Procurement teams need to stop treating classification as the vendor’s problem alone. If a company buys training data, evaluation work, expert review, or contractor capacity at scale, the buyer should know:

Which workers are employees of a vendor, independent contractors, agency workers, or platform contributors.
How workers are screened, trained, and paid for assessments.
Whether unpaid qualification work is used.
Whether task rejection can erase completed work.
Which jurisdictions govern the contract.
Whether minimum wage, overtime, benefits, leave, or local labor rules are implicated.
How worker complaints, payment disputes, and unsafe content exposure are handled.
Whether the buyer’s instructions create additional control over the worker.

SOMO’s April 2026 report on AI data workers argues that supply-chain transparency remains weak and that data workers face risks including low wages, unpaid labor, unsafe conditions, and limited remedy channels. The report’s policy recommendation is not subtle: companies using AI data work should take responsibility across the supply chain rather than hiding behind intermediaries.

That principle should enter the pay band. A low hourly rate that ignores unpaid assessment time, payment rejection risk, content exposure, surveillance, and unstable access is not a true cost. It is a shifted cost.

Guardian reporting on egocentric data adds another layer: the worker may not even be hired as an AI trainer. A factory worker can become a data producer because a camera records their movements. The work may be paid as ordinary factory labor while the resulting footage becomes valuable training data for robotics systems.

That breaks ordinary compensation categories. The worker is paid for time on the factory floor. The dataset buyer pays for captured know-how. The factory receives a side payment. The data company sells a training asset. The worker may not receive a separate share, may not understand downstream use, and may not have practical power to refuse.

The people named in the Guardian’s reporting show the conflict from different angles. Puneet Jindal of Labellerr AI described India’s scale and density of physical labor as useful to robotics data collection. Geeta Thatra of Work Fair and Free Foundation focused on whether insecure workers can meaningfully refuse cameras. Madhumita Dutta at Ohio State framed the recorded routines as a valuable digital asset, not just a byproduct of a workday. Sarayu Natarajan of Aapti Institute pushed the issue toward ownership of bodily knowledge.

Those are not the same objection. One is about market supply. One is about consent. One is about compensation. One is about rights over embodied skill.

This is not only a labor-law question. It is a pricing question.

If human skill creates training data that can be licensed and used to automate similar work, the compensation model cannot be limited to hourly labor. At minimum, companies need consent, direct payment rules, data-use disclosure, safety controls, and restrictions on secondary use. In higher-risk settings, they may need royalty-like mechanisms, worker councils, collective bargaining, or procurement standards that require proof of worker compensation.

Every data file needs a wage file.

A Compensation Map for AI Training Work

Companies do not need perfect job titles before they act. They need a map that prevents one vague label from hiding different labor costs.

The following framework separates task type, quality liability, domain expertise, data sensitivity, employment model, pay signal, and governance owner. It is meant for founders, HR leaders, procurement teams, and compensation teams that need to price AI training work before the market settles.

Work layer	Typical tasks	Pay anchor	Premium trigger	Main risk	Governance owner
Data labeling and annotation	Tag images, transcribe audio, categorize text, mark objects, rank simple outputs	Local wage floor plus fair task-rate design	Scarce language, difficult modality, harmful-content exposure, high accuracy requirement	Low wages, unpaid qualification work, unstable access, vendor opacity	Procurement + legal + vendor manager
Model response evaluation	Compare answers, apply rubrics, identify hallucinations, test tool use, log failure modes	QA, content operations, or product operations band	Complex rubric, high defect cost, production-release impact	Weak evaluation data creates false product confidence	Product + AI quality lead
Domain expert review	Review medical, legal, financial, coding, scientific, or regulated-domain outputs	Professional consulting or specialist contractor rate	Credential, liability exposure, scarce expertise, named signoff	Expert judgment is underpriced as generic annotation	Function leader + legal + model risk
AI data-quality operations	Maintain datasets, reconcile labels, audit drift, manage evaluator consistency	Data operations or analytics operations band	Dataset ownership, release gating, customer impact	Poor lineage, hidden label drift, unreproducible model behavior	Data lead + AI governance
Applied AI operations	Configure workflows, train users, review AI-assisted outputs, measure adoption	Salaried AI operations / enablement role	Workflow ownership, human review authority, measurable productivity or trust impact	AI adoption fails or creates hidden work for teams	Business owner + HR + IT
Worker-generated physical data	Record human movement, workplace routines, tool handling, embodied skills	Direct worker payment plus data-use consent standard	Commercial reuse, biometric or safety sensitivity, automation of similar work	Consent failure, surveillance, uncompensated data extraction	Procurement + labor relations + privacy
Internal AI evaluation lead	Build rubrics, manage evaluators, report quality, connect product and risk	Product operations / model risk / quality management band	Release authority, external audit support, customer trust impact	Evaluation becomes informal and unaccountable	Product + risk + compensation

This map should produce a decision file, not only a table.

For each role or vendor contract, the file should answer:

What human judgment is being purchased?
Does the work train a model, evaluate a model, deploy a model, or generate physical-world data?
What happens if the work is wrong?
Does the worker bring domain expertise, lived context, operational knowledge, or only task capacity?
Is the worker an employee, contractor, vendor employee, platform worker, consultant, fellow, or ordinary worker generating data as a byproduct?
Which wage floor, market benchmark, or professional rate applies?
Which benefits, overtime, payment-dispute, safety, and content-exposure protections apply?
Which data rights and downstream-use disclosures apply?
Which internal employee could qualify for the role, and what pay path would they receive?
Which budget owns the work: product, research, HR, procurement, operations, risk, or the business unit?

Question ten is the one many companies avoid. AI training work often sits between budgets. The model lab wants data. Product wants quality. HR wants workforce strategy. Legal wants risk records. Procurement wants rates. Finance wants control. Business units want useful automation. Workers want predictable pay and respect for the expertise being captured.

When everyone benefits from the work, no one wants to own the full cost.

That is why the pay band splits. Cheap annotation looks like a procurement expense. Expert review looks like a consulting expense. Applied AI operations looks like headcount. Worker-generated physical data looks like a data acquisition expense. Internal evaluation looks like product quality. If those costs remain scattered, the company never sees the true price of making AI useful.

Finance should want the consolidated view. Underpricing training work can make an AI project look cheaper than it is. Overpaying titles can make the project look uneconomic even when the right mix of labor would work. A clean map lets finance decide which human inputs actually change product quality, liability, adoption, or revenue.

Employees should want it too. A map creates a path from low-end task work into higher-skill evaluation, domain review, AI operations, or internal mobility. Without it, AI training becomes a temporary side hustle for some workers and a premium career path for others, with little explanation of how to move between them.

Dataset Work Leaves a Human Price Tag

Most AI training work carries one uncomfortable risk: the worker may disappear from the product story.

The model improves. The benchmark rises. The demo works. The robot handles the towel. The chatbot answers the case. The nonprofit builds the workflow. The startup wins the renewal. The dataset is licensed, cleaned, labeled, evaluated, and absorbed into a system that buyers call intelligent.

Human work remains in the trace, but often not in the title.

That is why this topic belongs with compensation, not only AI ethics. Companies can issue principles about responsible AI while still buying data work through opaque supply chains. They can promise workers future skills while paying them as temporary task labor. They can celebrate applied AI fellows while ignoring the annotators, evaluators, and recorded workers who made the model usable. They can offer an external premium for AI-labeled roles while asking internal employees to absorb evaluation and deployment work inside old bands.

Pay is where the claim becomes real.

Over the next 12 months, AI training work will likely split into clearer segments. Commodity annotation will keep facing price pressure and automation pressure. Expert evaluation will become more valuable as models enter higher-risk domains. Applied AI operations will turn into a salaried job family inside companies that need adoption, review, and workflow redesign. Worker-generated physical data will become a labor and consent fight as robotics companies search for embodied datasets. Internal AI quality roles will move closer to product release, risk, and customer trust.

That split is healthy if companies name it.

It is dangerous if they hide it behind one title.

A fair AI training pay system does not require every task to become a premium job. It does require companies to say what kind of human value they are buying. Volume is one price. Judgment is another. Domain expertise is another. Liability is another. Consent to capture bodily knowledge in a dataset is another. Turning AI into real work inside an organization is another.

Traditional compensation priced jobs by title and market survey. The new AI labor market asks for a more exact file: task, judgment, risk, data rights, worker status, pay floor, premium trigger, and path upward.

That file will not slow AI down. It will show what AI has been running on.

This article provides a deep analysis of AI training work, evaluator roles, and emerging AI compensation bands. Published June 24, 2026.