Blog · Fundamentals

AI annotation jobs and AI data annotation jobs, explained

Published 4 August 2026Updated 26 August 202613 min read

AI annotation jobs and AI data annotation jobs are the phrases people use when they do not want a warehouse scanning gig or a generic “work from home data entry” ad. They want to sit in the loop around models: labels that train the next system, or judgments that tell a lab which answer is less wrong. This post is the hiring-side map: what the work is, which four tracks do the hiring, what a single item actually looks like, what the bands have listed, and why eligibility rejects more people than skill does. The live list is open data annotation jobs.

What AI annotation means on a job card

An AI annotation job is a data annotation job whose customer is a model, not a search index or a human-only analytics dashboard. You still follow a written guideline. You still get graded against other people doing the same items. The difference is the payload: images and point clouds for perception stacks, conversations for assistants, code traces for reasoning models, policy flags for safety systems. That is the whole of the definition, and it is worth saying plainly because the phrase gets used as if it were a separate industry with separate entry requirements.

If you searched AI data annotation jobs, you are in the subset of data annotation jobs that markets itself with the AI prefix. We keep both phrasings because search does. Data annotation ai jobs, data annotation jobs ai, ai data annotator job, and general ai annotator jobs all resolve to the same set of listings on this board. For the concept without any hiring language attached, start at what is data annotation.

What actually changed over the last few years is the mix rather than the definition. Cheap, unambiguous labels are increasingly pre-generated by a model and passed to a human only for adjudication. The human queue has drifted toward the cases a model cannot settle on its own: preference judgments, expert domains, ambiguous safety calls, and evaluation sets that would be circular if a model wrote them. That drift is why so many current cards say trainer or evaluator rather than annotator, and why RLHF shows up in job descriptions written for people with no machine learning background.

One thing an AI annotation job is not: a seat where you chat with an assistant from the couch and get paid for the transcript. There is a genre of ad that implies exactly that. Real programs hand you a guideline, a queue, a clock, and a quality score, and the interesting part of your day is deciding which of two mediocre answers breaks the rubric less.

Not classic labeling, not data entry

Three kinds of work get advertised with overlapping language, and people apply to the wrong one constantly. Data entry transfers known values between fields. Classic labeling applies a closed schema to a fixed input. AI annotation in its current form asks for a judgment about open-ended text or a contested edge case, then asks you to defend the judgment in writing.

Dimension	Data entry	Classic labeling	AI annotation
Input	A form, invoice, or spreadsheet	An image or sentence with a fixed label set	Open-ended model output, long context, or ambiguous media
What correct means	Matches the source document	Matches the schema	Matches a rubric author’s judgment call
How you are scored	Typing accuracy and speed	Agreement with a gold label	Agreement plus the quality of your written reasoning
Throughput	Hundreds of records an hour	Dozens of items an hour	Often a handful of items an hour on writing queues
Failure mode	Typos	Missing an edge case in the guideline	A defensible answer with a reason nobody can learn from

Read the task description on a card rather than the job title. The three columns pay very differently.

The consequence is that speed is not the metric you are optimising. Two raters can look at the same item and disagree, so platforms measure that disagreement with inter-annotator agreement against a gold standard set, and quietly stop sending work to the bottom of the distribution. If you treat the queue like a typing test you will fail hidden gold items without ever seeing which ones.

It is also worth knowing that assessments now screen for pasted model prose in justifications. A fluent, generic paragraph that could have been written about any item reads as a red flag rather than a strong answer. The job is to be the human the model is measured against, not a second and sneakier model. The mechanics of that screen are in how to pass annotation assessments.

The four hiring tracks in AI data annotation

Vendors title the same contractor seat half a dozen ways, so sort cards by track instead. Nearly every AI annotation listing on this board belongs to one of four, and the track predicts the pay, the pace, the tooling, and how much you will write.

Track	What you do all day	Typical listed band
Perception and robotics	Boxes, polygons, keypoints, video tracks, occupancy and sensor data	$15 to $30 an hour, senior on-site QA far higher
Response writing and SFT	Write or heavily edit the answer the model should have given	$20 to $40 an hour
RLHF and preference data	Rank two model answers and justify the choice against a rubric	$25 to $50 an hour
Evaluation and red teaming	Score model output against a rubric, or try to break the model	$25 to $50 an hour, expert domains above that

Bands reflect what cards on this board have listed, not offers or averages. Check the live feed for current numbers.

Perception and robotics

Boxes, tracks, occupancy grids, and weather edge cases. Half the guideline is occlusion rules: what to do when a pedestrian is two-thirds behind a van, whether a reflection in a shop window counts as an object, how to label a cyclist who is also a person. The skill that separates people here is spatial patience, meaning tight polygons at pixel level for six hours without drifting. Programs of this kind still hire, including senior on-site quality assurance that looks nothing like a laptop gig, and the detail sits in Scale AI annotation jobs.

Response writing and SFT

You write the response the assistant should have produced, or you take a mediocre one and rebuild it until it would pass as a gold example. This is supervised fine-tuning adjacent work and a large slice of AI trainer jobs. Throughput is low and the bar is style as well as substance: the rubric will usually specify structure, hedging, refusal behaviour, and how much the answer is allowed to assume about the user. Bilingual candidates are worth more here than in the crowded English-only pool, so look at multilingual and localization evaluation if you have a strong second project language.

RLHF and preference data

Compare two outputs, pick a winner, explain the choice, and sometimes produce a third gold response. This is the cluster most people mean when they type AI annotation jobs, and it is where the top of the general band lives. Practice matters more than credentials here, which is why the RLHF and model evaluation course exists, and the live filter is RLHF jobs.

Evaluation and red teaming

Score a model against a rubric across dimensions such as factuality, instruction following, and harm, or deliberately probe for failures. LLM evaluator jobs and red teaming sit here. This is not beginner volume work: the rubrics are long, the items are adversarial by design, and safety queues carry real exposure to material you cannot unsee. Read the wellness and break rules before accepting a policy queue and treat it as a genuine tradeoff rather than an easy door in.

Inside one RLHF item, step by step

Job descriptions describe this work in the abstract, so here is the concrete shape of a single item on a preference queue. You are shown a user prompt, for example a request for advice about stopping a prescribed medication early. You are shown response A and response B. The rubric lists dimensions in priority order: safety first, then instruction following, then factual accuracy, then tone and length. You have a comment box with a minimum character count and a timer that the platform is watching.

Response A is warm, thorough, and quietly recommends a taper schedule. Response B is shorter, refuses the specific dosing advice, and tells the user to contact the prescriber. Most newcomers pick A because it is more helpful and better written. The rubric almost certainly picks B, because safety outranks helpfulness and dosing guidance is a restricted category. The item is not testing your taste. It is testing whether you can find the rule and apply it when the wrong answer feels nicer.

Justification style	What it reads like	How a reviewer scores it
Empty	B is better. It is safer and more accurate.	Fails. Nothing here tells a model what to change.
Model-sounding	Both responses have merit, but B demonstrates a more responsible approach overall.	Flagged. Generic phrasing that would fit any item, often read as pasted output.
Rubric-anchored	B wins on the safety dimension: A gives a specific taper schedule, which the guideline lists as restricted medical advice. A is stronger on tone, but tone ranks below safety.	Passes. Names the dimension, quotes the rule, concedes the tradeoff.

On language tracks the written reason is the product, not a formality attached to the click.

Notice what the strong version does. It names the dimension that decided the call, points at the specific sentence that triggered it, and admits the losing response was better on a lower-priority dimension. That last part is what reviewers look for, because a rater who never concedes anything is usually pattern-matching rather than reading. Assume some fraction of the items in any batch are hidden gold questions with a known answer, and that your agreement with those decides how much volume you see next week.

The same structure shows up on evaluation queues with a score instead of a winner, and on prompt and response writing queues with a rewrite instead of a score. Learn to write three tight sentences that reference a rule and you are employable on all three.

Skills that show up in real screens

Nothing on this list is about machine learning. All of it shows up in the first hour of a real assessment.

Native or near-native writing in the project language. Not conversational. You are the quality bar the model is measured against, so subtle awkwardness in your text becomes a defect in the dataset.
The ability to quote a rubric instead of a feeling. A justification that names the dimension and the rule survives review. One built on instinct does not, even when the verdict was right.
Tolerance for long, boring documents. A 40-page guideline with a mid-project addendum is normal, and the addendum is usually where the gold items come from.
Genuine reasoning on specialist tracks. On code queues you have to actually read the snippet and predict what it does. The code and reasoning data course covers what that grading looks like.
Comfort being wrong and changing after QA notes. Feedback arrives as a score change, not a conversation. People who argue with the reviewer instead of adjusting lose the queue.
Stable equipment and predictable hours. Vendor tools are browser-heavy, video and 3D tasks need a real machine, and flaky login patterns get people dropped from projects.

If you want a rough read on where you stand before you spend an attempt on a real test, run the skills assessment, then work through the free AI data annotation foundations course. If you have never done the work at all, the on-ramp is how to become a data annotator.

Pay by track, and why it differs

Listed bands for AI annotation work sit in the same range as the rest of the board, with RLHF and specialist evaluation toward the top. Entry content and policy review has listed from about $12 an hour, general remote labeling clusters around $18 to $30 an hour, and specialist RLHF and evaluation reach roughly $50 an hour and above, with expert domains higher again. The full breakdown, including how per-task rates convert into an effective hourly number, is in the data annotator salary guide.

Three things move a rate more than seniority does. Scarcity of the skill comes first: a working developer or a licensed clinician is rare in this labor pool, so expert queues list multiples of general review. Location restrictions come second, because a US-only program carrying residency or compliance conditions lists higher than its global sibling running the same rubric. Format comes third: writing a gold response takes far longer per item than picking a category, so writing queues carry higher rates and lower throughput.

Why the hours are the real risk

The listed rate is the ceiling of a good week, not the base of a predictable month. AI annotation is contractor work: projects end, you buy your own idle time, you handle your own taxes, and there is no notice period when a queue dries up. Passing an assessment and then waiting is a normal outcome rather than a sign something went wrong.

Assume a ramp. Most vendors trickle a small calibration batch before they open real volume, so week one earnings say nothing about week five.
Track your effective rate on per-task projects. Two dollars an item is excellent at three minutes and terrible at twelve.
Read the payment terms before you accept. Weekly platform payouts and net-30 invoices feel very different when rent is due.
Keep a second qualification live, because the fix for a paused project is another project rather than a support ticket.

Eligibility, location, and US-only programs

Worldwide eligibility is not automatic, and this is where most AI annotation applications actually die. A large share of AI programs list as US-only. That is rarely a judgment about your skill. Frontier labs push data residency, background check, and tax and identity requirements down to their vendors, and the cheapest way for a vendor to satisfy a clause about where data is processed is to lock the project to one country.

Three patterns cover almost every card. Remote global is the widest door outside the United States and usually lists lower bands for the same rubric. Remote US-only carries the compliance premium and often asks for a tax form and identity verification during onboarding. On-site is a small slice, mainly perception quality assurance where the data cannot leave the building. The full explanation of location patterns, including how payment works from abroad, is in data annotation jobs worldwide.

Two practical rules follow. Read the location string on every card before you invest an hour in the application, because a Remote (US) card is closed to you if you are not physically in the United States no matter how good your writing is. And if you are outside the US, prioritise multilingual and global queues where your language is the scarce input rather than competing in the deepest English-only pool on the market. Beginners who keep landing on senior or locked cards should start at beginner and entry-level roles instead.

Ready to apply?

AI annotation openings change weekly, and the location line matters as much as the pay line. Filter the live feed, then apply on the employer site from the role page.

Browse open roles Take the skills assessment Free foundations course

Job titles decoded

Titles in this market are close to meaningless on their own. The same seat gets four names depending on who wrote the requisition, and a fancier noun does not reliably mean a higher band. Use the table below to translate a card into the work, then check the task description to confirm.

Listed title	What the work usually is	What to check on the card
AI data annotator	Generalist labeling or review, whatever queue you qualify for	Which modality, and whether the band is per hour or per task
AI trainer or LLM trainer	Response writing and supervised fine-tuning style work	Whether you write from scratch or edit existing answers
RLHF rater or preference annotator	Ranking pairs of model answers with written justifications	Minimum comment length and whether gold items are mentioned
LLM evaluator or model evaluator	Scoring output against rubric dimensions, sometimes adversarial probing	How many dimensions, and whether a domain credential is required
Content or policy reviewer	Trust and safety queue work on harm, scams, and disallowed advice	Exposure level, wellness rules, and the break policy
AI annotation specialist	Usually an evaluator or domain expert seat with a marketing label	The actual domain, because the title alone tells you nothing

Titles vary more between vendors than the underlying work does. Sort by task description and band.

The searches follow the same pattern. Data annotation ai jobs, data annotation jobs ai, data annotation jobs for ai, ai data annotator jobs, ai data annotator job, and general ai annotator jobs all land on the same pool of listings. AI annotation specialist and AI data annotation specialist are usually evaluator or trainer titles with a fancier sticker. Use data annotation evaluator jobs when the card is scoring model output, and keep the glossary open for the vocabulary that shows up in the rubric.

How and where to apply

The recurring names on this board are Scale AI, Surge AI, Mercor, Micro1, and HireCade. None of them are this website. We collect apply links; they run the tests, the contracts, and the payments. The full list of who has been listing is the companies index. If you want Surge specifically, read Surge AI annotation jobs, and for a head-to-head use Scale AI versus Surge AI.

Pick a track before you pick a listing. Decide whether you are applying as a perception annotator, a response writer, an RLHF rater, or an evaluator. The four tracks test different things, and a scattergun application to all of them usually produces four mediocre assessments instead of one pass.
Read the location line before the pay line. A card that says Remote (US) is closed to you if you are not physically in the United States. Many AI programs carry residency or compliance conditions, so eligibility filters you out long before skill does.
Apply on the employer domain from the role page. Open the listing on DataAnnotationJobs.org for the stable URL, company name, and listed band, then use Apply to reach the employer form. The vendor owns the test, the contract, and the payment. This board does not hire.
Treat the assessment as production work. Read the whole rubric before item one, assume hidden gold items are mixed in, and justify every judgment with a phrase you could point to in the guideline. Do not paste chatbot prose, because screens now flag fluent generic paragraphs.
Protect the first two weeks of QA, then stack a second vendor. Early samples decide how much volume you see. Read reviewer feedback, redo flagged items the way it was asked, and pass a second vendor assessment while your first queue is still healthy so a paused project does not become a zero week.

Step four decides more outcomes than the other four combined, so it is worth over-preparing. Timed practice on rubric-anchored justifications sits in the assessment course, and if you have applied everywhere and heard nothing, the diagnosis is usually one of a handful of specific mistakes listed in how to get data annotation jobs. One last filter worth applying yourself: any ad promising a high rate with no rubric and no test, asking for an upfront fee, or hiring entirely inside a chat app is not an AI annotation job. Compare it against the live remote data annotation jobs feed and how our listings work before you hand over anything.

FAQ: AI annotation jobs and AI data annotation jobs

What are AI annotation jobs?

AI annotation jobs are data annotation roles whose output feeds a model: labels that train it, or judgments that score it. In practice that means tagging images and video for perception systems, writing the answer an assistant should have given, ranking two model responses against a rubric, or grading output for accuracy and safety. People search AI data annotation jobs for the same openings.

Are AI annotation jobs just data entry?

No. Data entry moves known values between fields and is scored on accuracy against a source document. AI annotation asks you to make a judgment under incomplete instructions, then defend it in writing against a rubric. Throughput is lower, quality assurance is stricter, and two reasonable people can disagree on the same item.

What do AI data annotation jobs pay?

Cards on this board have listed roughly $12 an hour at the entry end of content and policy review, about $18 to $30 an hour for general remote labeling, and up to about $50 an hour and above for specialist RLHF and evaluation work. Expert domains such as code, medicine, and law list higher. Those are listed bands, not offers, and hours are rarely guaranteed.

Do you need a degree for an AI data annotator job?

No degree is required for general annotation, response writing, and review queues, and most vendors screen with a test rather than a resume. Specialist tracks are different: code, math, medical, and legal projects genuinely check the skill or the credential. Claiming a background you do not have tends to end an account rather than upgrade it.

Are AI annotation jobs remote and open worldwide?

Most listings are remote, but remote splits into two very different things: remote global and remote US-only. A large share of AI programs carry US-only conditions, and a small number of perception roles are on-site because the data cannot leave the building. Read the location string on every card rather than assuming remote means anywhere.

What is an AI annotation specialist?

AI annotation specialist is usually a marketing label rather than a distinct occupation. Behind the title you normally find an evaluator, an RLHF rater, or a domain expert grading model output in one narrow area. Read the task description and the pay band on the card, because the noun in the job title varies more between vendors than the work does.

How long does it take to get hired?

Plan for weeks rather than days. A typical sequence is application, an unpaid or lightly paid assessment, identity and tax onboarding, then a small calibration batch before real volume opens. Passing the test is not the same as having work, because queues open when a project needs raters. Silence after a pass is common rather than a rejection.

Where do I find AI data annotation jobs right now?

Start with the live jobs feed on DataAnnotationJobs.org, then narrow with the RLHF, AI trainer, and LLM evaluator pages if you already know your track. Company pages for Scale AI, Surge AI, Mercor, Micro1, and HireCade show who has been listing. We are an independent directory, so you apply on the employer site and we never charge to unlock a listing.

For the wider market including non-AI-branded titles, read data annotation jobs in 2026. For the definition without hiring language, read what is data annotation. For live openings, stay on the jobs feed.