By Mufy Pachorawala, founder of getAIwork · Last updated: September 3, 2026
Key facts
- It is vendor work, not direct hiring. The technology client buys the ratings; a staffing vendor recruits, tests and pays the raters
- A qualification exam is standard. Typically a guideline document running to a hundred pages or more, followed by a graded test that can be failed
- Hours are usually capped, frequently in the ten to twenty per week range, and are not guaranteed to be available
- Locale is often the requirement, not a degree. Many roles ask for residency in a specific country and fluency in a specific language variant
- 162 of 1,435 live listings on the getAIwork board are tagged beginner-friendly as of September 3, 2026, and rating work is one of the categories they sit in
- Nobody legitimate charges you to apply. Vendors in this category are paid by the client, never by the rater
What does an AI rater actually do?
You are handed a task and a standard, and you judge one against the other. The oldest version is search quality rating: here is a query, here is a result, does the result serve what the person actually wanted. The newer versions apply the same method to assistant answers, generated summaries, ads, recommendations and safety classifications.
| Rating type | What you judge | What makes it hard |
|---|---|---|
| Search quality | Whether a result meets the intent behind a query | Working out intent from three ambiguous words |
| Ads quality | Whether an ad is relevant, accurate and appropriate for the placement | Policy edge cases, and consistency across hundreds of near-identical items |
| Assistant answer evaluation | Whether a generated answer is accurate, helpful and appropriately hedged | Checking facts quickly without going down research rabbit holes |
| Content and safety classification | Whether an item belongs in a defined policy category | Genuinely unpleasant material on some queues |
| Local and map data | Whether a business listing is correct and current | Local knowledge, and tedium |
The through-line is that you are applying somebody else’s standard, not your own. Raters who substitute personal judgement for the guideline score badly even when their instinct is defensible, which is the single most common reason capable people fail the qualification exam.
The vendor model, and why it matters
Large technology companies almost never hire raters directly. They contract outsourcing and localisation vendors, who recruit the raters, run the exams, manage quality and issue the pay. Names in this space include the big localisation firms and crowd vendors, and the same rater population circulates between them as contracts move.
Three consequences follow, and they explain most of the confusion in community threads about this work.
- Your employer is not the brand on the tasks. You work for the vendor. Support, pay questions and account issues all go through them.
- Contracts move. When a client moves a language or a programme to a different vendor, the work moves with it, and raters are sometimes offboarded en masse through no fault of their own.
- Pools are staffed ahead of demand. Vendors qualify raters so they can promise a client capacity in a given locale. Being qualified is not the same as having work, which is the source of most complaints in this category.
That third point is worth sitting with before you spend a weekend on an exam. It is the same structural pattern we described in the piece on platform silence: qualification is cheap for the vendor to run and expensive for you to complete, so the incentive is to over-qualify. Ask, before any unpaid assessment, whether there is live volume on the programme right now. A vendor with real work answers that.
The qualification exam
Nearly every rating programme gates entry behind an exam on its guideline document, and those documents are long. A hundred pages or more is normal. The exam typically has a theory section on the guidelines and a practical section where you rate real items and are scored against expert consensus.
Most programmes allow a limited number of attempts, sometimes two. Failing can lock you out of that programme for a period, so treating the first attempt casually is expensive. What actually works, from the accounts of people who pass:
- Read the whole guideline document before attempting anything. Skimming and referring back during the practical section is how people run out of time.
- Learn the rating scale definitions verbatim. Most failures are boundary calls between adjacent ratings, not wild errors.
- Study the worked examples hardest. They encode the interpretation the graders actually use, which is not always obvious from the definitions alone.
- Answer as the guideline would, not as you would. If you disagree with a rule, follow it anyway. The exam is testing compliance, not taste.
1,435 live AI listings on the free getAIwork board, 162 of them tagged beginner-friendly. Two minutes matches you to the ones that fit.
Take the 2-minute match quiz →
Pay and hours, as listed
We quote only what postings state. Rating work is generally hourly, generally part-time, and generally capped. Caps in the range of ten to twenty hours a week are common across the category, and they are caps rather than commitments: the hours have to be available before you can work them.
Rates vary by locale and language more than by task difficulty, for the ordinary reason that supply differs. A widely spoken language with many qualified raters lists lower than a scarce one. Nothing about the work itself changes.
| Factor | Effect on what gets listed |
|---|---|
| Language and locale | The largest single driver. Scarce pairs list higher |
| Programme type | Safety and policy queues frequently list above general search rating |
| Vendor | Different vendors list different rates for comparable work on the same client |
| Experience as a rater | Modest effect. This is not a career ladder with steep increments |
| Hours available | No effect on the rate, and the main driver of what you actually take home |
Because rates are set per programme and per locale, any article quoting you one universal AI rater hourly figure is inventing it. Check the specific posting. Our weekly board statistics publish the listed ranges across the whole market, refreshed every week, and the spread there is the honest picture.
Who gets hired
This is one of the genuinely accessible corners of AI work, and it is worth being precise about why. Vendors are not buying a credential. They are buying coverage of a locale, so the requirements tend to be residency in a specific country, fluency in a specific language variant, and enough cultural familiarity to judge what a local user meant.
Some programmes ask for a degree, many do not. What they all effectively require is careful reading, consistency across repetitive work, and the self-discipline to follow a rule you find slightly wrong for four hours. That is a real skill and not everybody has it.
If you are starting from no experience at all, this category and platform task work are the two realistic entry points, and we cover both in AI training jobs with no experience. On the board today, 162 of 1,435 live listings are tagged beginner-friendly, which is worth knowing before anyone tells you the market is wide open.
How to apply well
Apply directly on vendor career sites rather than through aggregators, because rater postings are among the most impersonated listings in remote work. Verify the vendor exists independently, and remember the fixed rule: legitimate vendors are paid by their client, so nobody should ever ask you for money to apply, to train or to be assigned work.
Apply to two or three vendors rather than one, because contracts move between them and a single relationship is fragile. Then block out real time for the exam. Treating the guideline document as bedtime reading and the exam as a formality is the most common self-inflicted failure in this category.
What the job listings leave out
Qualification is not work. Passing the exam puts you in a pool. Volume depends on the client’s contract, not on your score.
The exam time is unpaid. Reading a hundred-page guideline and sitting a two-part exam is a real weekend, and nobody pays for it. Budget it as a cost of entry rather than expecting it back.
Quality monitoring is continuous. Your ratings are audited against expert consensus throughout, and sustained low agreement ends the assignment. This is normal in the category and is not personal.
Some queues are grim. Policy and safety rating means reading things you would not choose to read. Reputable programmes disclose it and pay accordingly. Read the description before accepting the queue.
It is self-employment in most arrangements. Nothing is withheld, and the tax is yours to handle. In the US that means the rules we set out in the guide to taxes on AI task work.
1,435 live AI listings as of September 3, 2026, screened and human-approved, filled listings deleted daily. No signup.
Frequently asked questions
What is an AI rater job?
Work judging whether a machine got something right: whether a search result matches a query, whether an ad is appropriate, whether an assistant’s answer is accurate and helpful. You apply a published guideline standard rather than your own judgement, and the ratings feed back into improving the system.
Do AI rater jobs require a degree?
Often not. The usual requirements are residency in a specific country, fluency in a specific language variant, and passing a qualification exam on the guideline document. Careful reading and consistency matter more than credentials in most programmes.
Who hires AI raters?
Outsourcing and localisation vendors, on behalf of large technology clients. The vendor recruits, tests, manages quality and pays you. The technology brand whose tasks you rate is not your employer, which is why support and pay questions go to the vendor.
How much do AI rater jobs pay?
Rates are set per programme and per locale rather than by a single market figure, and are driven mainly by language scarcity. Any source quoting one universal AI rater rate is inventing it. Check the specific posting, where pay is stated as listed.
How many hours a week can you work as an AI rater?
Usually capped, commonly in the ten to twenty hour range, and the cap is a ceiling rather than a promise. Hours have to be available before you can work them, and availability moves with the client’s contract.
Is the AI rater exam hard?
It is passable but not casual. Expect a guideline document of a hundred pages or more, a theory section and a practical section scored against expert consensus. Attempts are usually limited, so read the full document and learn the rating scale definitions before starting.
Why did my rater work dry up?
Most often because the client’s contract changed or moved to a different vendor, which offboards raters in groups regardless of performance. It can also mean your locale currently has more qualified raters than volume. Neither is a judgement on your work.
Are AI rater jobs legit?
The established vendors are, and the work has existed for well over a decade under the search-quality label. The category is also heavily impersonated, so apply on vendor career sites directly and never pay anyone to apply, train or be assigned work.
Mufy Pachorawala · Founder, getAIwork
My AI scans thousands of AI-job posts (40,786 screened so far) and I personally approve every listing before it reaches the board. I write these articles by the same rules: pay quoted only as listed, and no income promises. Read our editorial rules.
Related: AI training jobs with no experience · Get paid to train AI · AI jobs with no experience · AI training platforms hiring now · AI job market statistics
