Scale AI annotation jobs: Scale AI data annotation, AV, and polygons

What Scale AI annotation jobs usually are

Scale is the large data platform: perception for vehicles and robotics, government programs, and language data for models. That breadth is the single most useful fact about applying there. A Scale AI data annotation job might be remote vision-and-language labeling at a mid hourly band, or senior on-site autonomous-vehicle quality assurance at a much higher listed rate, and those two cards have almost nothing in common except the company name. The title and the task description on the card matter more than the brand.

That is different from a specialist vendor where every queue looks roughly the same. It means the question “what is it like to work for Scale” has no single answer, and it means two people can both be correct while describing completely different experiences: one drawing cuboids in a point cloud with a stopwatch running, the other reading two model answers and writing a justification. Browse the current rows on the jobs feed and read them as separate jobs that happen to share an employer.

Worker-facing brands under the Scale umbrella have changed names over the years. If an old forum post says Remotasks or Outlier, that is history, not a second careers site you must find. Apply through the URL we publish for that opening, and be suspicious of any site claiming to be a Scale worker portal that wants a fee, a deposit, or your banking details before there is a signed agreement with a named entity. Our side of the arrangement is explained in how our listings work.

The program mix and where you fit

The tracks below are the shapes we see, drawn from cards this board has carried plus what Scale describes publicly. Not all of them appear as open contractor listings at any given moment, and the mix moves with client demand.

Picking a track honestly is the highest-leverage thing you do before applying. Perception and language assessments test unrelated skills, and a strong writer who applies to a cuboid track fails on tooling precision while concluding, wrongly, that Scale rejected them as a candidate. If you are unsure which side of that line you sit on, the free skills assessment gives you a rough read in a few minutes, and AI annotation jobs explains the four hiring tracks in the market generally.

Automotive annotation and a polygon annotation career

Searches for an AI annotator for automotive jobs and for a polygon annotation career are pointing at the same thing: perception data. Draw a tight polygon around a pedestrian, cuboid a vehicle, track it through occlusion, mark weather and lighting. That is slower, more spatial, and closer to bounding box work than to writing a chatbot critique, and it rewards a completely different temperament.

The craft is worth describing concretely, because job ads never do. Tight does not mean small: it means the boundary follows the silhouette rather than a convenient hull, so a cyclist gets the gap between the arm and the torso, and a parked car gets the wheel wells rather than a rectangle glued to the tarmac. The shadow is not part of the object. The wing mirror usually is. A pole crossing a pedestrian splits the visible region into two pieces, and the guideline, not your instinct, decides whether that is one instance or two.

What tight polygon and cuboid work demands

• Occlusion and extent. When the back half of a truck is hidden, you still estimate the full cuboid. Reviewers check whether your estimate is plausible against the visible wheels and the road plane, and a cuboid that shrinks as the object hides is a classic flag.
• Orientation. A cuboid with the right footprint and the wrong heading is wrong. Yaw errors on parked vehicles sitting at an angle to the curb are one of the most common corrections in review.
• Track identity. An object that leaves the frame and comes back should keep the same identity if the guideline says so. Identity swaps between two similar cars passing each other are expensive downstream and easy to create by accident.
• Truncation at the frame edge. Half a bus at the image border has its own rule, and it is usually different from the rule for an occluded object in the middle of the scene.
• Attributes. Night, rain, glare, wet road, and heavy-versus-light occlusion tags exist so a model can be evaluated by condition. They feel like busywork and they are graded like labels.
• Ambiguous classes. Is a delivery robot a vehicle. Is a trailer separate from the tractor. Is a group of twelve people one crowd region or twelve instances. The document decided months ago, and consistency with that decision is the whole job.

What the annotation tooling feels like at hour four

The tooling shapes the day as much as the rules do. You work in a browser with hotkeys you will eventually use without thinking, interpolation between keyframes that saves an hour and quietly introduces drift in the middle frames if you never check them, and point-cloud panes beside camera views where a sparse object at forty metres is a dozen returns you have to interpret. Six hours of that without your boundaries loosening is the actual skill. People who cannot sit with it discover the fact in week two, not in the assessment.

Some of that queue is remote. The highest listed specialist quality assurance role we have carried for Scale was on-site in San Francisco, not a worldwide laptop gig. If you specifically need data annotation jobs worldwide, do not assume every Scale AI annotation job is global. Autonomous vehicle annotation jobs also tend to sit at the senior end, so the realistic route in is production labeling first and review later.

Language, RLHF, and general AI annotator work at Scale

Scale also runs language and evaluation projects. Those overlap what people call general AI annotator jobs or AI data annotator jobs: rate answers, follow a rubric, survive QA. The daily texture is the opposite of perception work. Nothing is pixel-precise and everything is arguable, so the graded artefact becomes your written justification rather than your boundary.

Switching between perception and language work

If you have only ever done vision work, the adjustment is real. You will be asked to say why one answer beats another in a way a stranger could audit, and “it reads better” is not an answer. If you have only ever done language work, the adjustment in the other direction is also real, because a guideline that felt like a suggestion in a preference task becomes a measurement standard when the output is a polygon. Both jobs are scored with inter-annotator agreement against a gold standard set, which is the thread connecting them.

For a writing-heavy preference queue specifically, Surge AI annotation jobs may fit better. For a mixed platform where vision and language sit side by side, Scale is the default search. If you want the concept before the hiring language, start at what RLHF is and the RLHF jobs feed, or the narrower LLM evaluator jobs list.

Remote versus on-site, and what that means for eligibility

Searchers looking for Scale AI jobs remote are asking a fair question with an annoying answer: it depends on the data. Language and evaluation projects travel well, so those cards are usually remote. Some perception and public-sector data does not leave a controlled environment, which is why a senior quality assurance role can be on-site even at a company built on distributed work.

Read the location line before the pay line, every time. A US-locked program carrying data residency or public-sector requirements often lists higher than a global sibling running a similar rubric, and that gap is compliance pricing rather than a comment on your ability. There is no version of a strong application that overcomes a geography mismatch, so filter first and apply second. The narrower feed is remote data annotation jobs.

How hiring works, step by step

Assessment first, hours later. That order is the thing most people get backwards, and it explains why passing feels anticlimactic: you have joined a qualified pool rather than started a job.

1. Read the location line before the title. Scale cards on this board have ranged from remote labeling to senior on-site quality assurance in San Francisco. Decide whether you are eligible for the location before you get attached to the rate, because no writing sample fixes a geography mismatch.
2. Match the track to a skill you can prove. Perception work rewards spatial patience with polygons and cuboids. Language and evaluation work rewards precise writing. Applying to the wrong track is the quiet reason a strong candidate fails an assessment that had nothing to do with their actual strength.
3. Apply through the employer link on the card. Open the role page on DataAnnotationJobs.org for the stable URL, the company name, and the listed band, then follow the apply link into Scale’s own hiring or contractor flow. Old worker-facing brand names from forum threads are not a second door to hunt for.
4. Treat the assessment or calibration task as production. Read the whole guideline before your first item and assume gold questions are mixed into the set. On perception tasks the graded thing is boundary precision and label consistency, not how many frames you got through.
5. Protect your quality score through the first batches. Early samples decide how much volume you see later. Redo flagged items exactly as the reviewer described, and raise ambiguous cases in the project channel rather than guessing twice and dragging your agreement rate down.
6. Hold a second vendor qualification. Perception and language programs both end when a client changes direction. Pass an assessment somewhere else while your Scale queue is healthy, so a paused project turns a zero week into a slow week.

Step four deserves extra attention on perception tracks, because the graded qualities are unintuitive. Reviewers look at boundary precision, label consistency across similar objects, correct use of attributes, and whether you followed the occlusion rule rather than your eye. Throughput is a later conversation. On language tracks the same step means quoting the guideline in your reasoning instead of asserting a preference. The longer version is how to pass annotation assessments, with timed practice in the assessment course and the fundamentals in the free AI data annotation foundations course.

Pay bands by track

Pay bands for Scale cards on this board have included roughly $18 to $30 an hour for remote vision and language work, and much higher for senior autonomous-vehicle quality assurance. Always trust the live card over any number in an article, including this one.

Three things move a rate more than seniority does: scarcity of the skill, location restrictions, and how long a single item takes. On per-item perception projects that last point is the one that bites, because a rate that looks generous per frame can be poor once you include the frames where interpolation drifted and you fixed twelve of them by hand. Track your effective hourly number through the first week. The cross-vendor picture is in the data annotator salary guide, and hours stay variable on contractor work whatever the band says.

Scale AI versus Surge AI, and when to pick which

The short version: Scale is breadth, Surge is a writing filter. Scale gives you more ways in, including tracks where spatial skill counts for more than prose. Surge concentrates on language, reasoning, code, and preference quality, so its assessment leans hard on how you justify a judgment.

The side-by-side detail lives in Scale AI versus Surge AI, and the current roster is the companies index with hubs for Surge AI, Mercor, Micro1, and HireCade. If you searched Meta AI annotator roles, Wolfestone, Prodigy, or other names we do not currently list, we will not invent those requisitions. Use the companies index for who is on the board today.

What people get wrong about applying to Scale

The recurring mistakes are specific, and most of them happen before anyone grades your work.

• Treating the brand as one job. Scale contractor jobs span perception, language, and evaluation. Applying to whatever is newest instead of whatever matches your skill is how good candidates collect unrelated rejections.
• Assuming every role is remote. The senior quality assurance card we carried was on-site. Reading the pay line first and the location line second wastes more time than any other single error here.
• Hunting for a legacy worker portal. Old brand names in forum threads send people looking for a door that is not there, and straight into whatever lookalike site ranks for it.
• Racing the clock on a perception assessment. Boundary precision and consistent labels are what get scored. Frames per hour matters once you are in production, not before.
• Trusting interpolation. Keyframing an object across forty frames and never checking the middle is the fastest way to deliver drift that a reviewer will catch and you will redo unpaid.
• Skimming the guideline. Occlusion, truncation, and class edge cases are all decided in the document. Guessing them consistently is still worse than reading them once.
• Expecting steady hours. Programs pause when a client changes direction, so a single vendor is a single point of failure for your income rather than a job.

If you have applied and heard nothing, the diagnosis is usually one of a short list of fixable things in how to get data annotation jobs. If the specialist perception cards are out of reach for now, start with beginner and entry-level roles, build a quality record, and read how to become a data annotator for the full on-ramp. For the market view across every vendor, the long read is data annotation jobs in 2026.

FAQ: Scale AI annotation jobs, automotive work, and polygon careers

The official Scale site is scale.com, the roles we track sit on the Scale AI company page, and the vocabulary you will meet in a perception guideline is in the glossary. A Scale AI annotation job is filled by Scale, not by us.