The annotation descriptions on micro1’s portal ask for work done against a project’s guidelines and scoring guides. This page reports what those descriptions say, and then works through four ways a batch of labels stops being consistent — this guide’s own practice material — and what an annotator does about each.
It is written for people who can follow a detailed specification exactly, over a long session, and who would rather ask a clarifying question than guess. Several descriptions ask for consistency through repetitive tasks in those terms; what else a given listing requires is in its own description.
The mix of listings changes; the portal is where to see what is open. What follows first is what the full descriptions said on the date in Sources. The four failures after that are the blog’s own practice material.
Opens micro1. You apply and are vetted by micro1, not by this site.
What the listings describe
Four different kinds of work sat under annotation and data collection, and the title can fail to tell you which one you are looking at.
Labeling video of robots. Several descriptions are about footage of robotic arms performing tasks: watch each session, tag actions, events and errors against the customer’s grading guidelines, score clips with a predefined scoring guide, write up observations, and review other annotators’ work. One description goes further: decide whether a clip of a robot or a person matches a written instruction, compare two clips, and correct AI-generated annotations, including where an action starts and ends. Listings titled Photographer and Filmmaker are this same work, with photography or filmmaking experience asked for.
Checking what a model said about a clip. Two descriptions ask the annotator to label short video and audio clips, judge whether AI-generated descriptions of them are accurate, write up what is wrong, and flag material against moderation standards. A music and sound listing asks trained listeners to annotate short field recordings — instruments, tempo, key, time-stamped background events — and, in later phases, to grade a model’s description of the same clip.
Contributing photographs of yourself. The photo contributor descriptions ask participants to take and submit photographs of themselves following instructions and to fill in a short questionnaire, from their own smartphone. They say the dataset is meant to represent people across a broad range of racial and ethnic backgrounds and appearances; the twins and siblings versions say the data helps AI systems understand visual similarity between related people, and that each relative takes part. This is data about you, not labeling.
Capturing gameplay. One description asks for playing current big-budget games to capture and annotate gameplay data with the project’s tools. It requires a high-end gaming PC and access to the specific titles.
Not what the title says. The Data Entry Keyer description is neither data entry nor annotation: it asks for designing evaluation tasks for AI agents and writing long grading rubrics. It is outside this page. Read the description before deciding a title is or is not for you.
The examples below are illustrative exercises created for this guide, not published micro1 assignments or interview questions. They show the kind of judgment the work needs; they are not a preview of what a project will hand you.
The four failures are this guide’s grouping, not a taxonomy micro1 publishes. Each one names the kind of work, from the section above, that it is practice for. All four are practice for the two labeling kinds — video of robots, and checking what a model said about a clip. Nothing here practices photo contribution or gameplay capture: those are about supplying data, and the descriptions are the place to read what they ask. Every clip, instruction and guideline in the examples is invented.
When the guideline is silent. Three of the exercises below meet a case no rule covers. In them, keep four things apart. A missing rule is something to report, not permission to decide. Your proposed label or convention is a proposal. It becomes the rule only when the guideline’s owner accepts it — in these exercises, the owner answers every flag before the batch is handed in. Only then is the batch final. On a real project, who answers a flag, and how, is for the project to tell you; this page does not know.
Failure: guessing at a category boundary
Practice for: labeling clips and flagging material against a project’s standards. When labels come from a fixed list, the difficult items sit on the line between two neighboring labels. The failure is picking one quickly and moving on.
Worked example. An invented guideline has the event labels “grasp” and “push”. In the clip the arm closes its gripper on a block without lifting it and slides it across the table. Apply the written rule if the guideline has one for this case. If it has none, flag the item with a note giving the label you propose and why, and hold it out of the batch until the owner answers; then label it by the answer. The flag is part of the work.
Failure: grading the instruction instead of the clip
Practice for: judging whether a clip matches a written directive, which one video annotation description lists. The failure is marking what the instruction says should have happened, not what the clip shows.
Worked example. The invented directive reads “pick up the red block and place it in the tray”. In the clip the arm picks up the red block and releases it beside the tray. Whether that counts as a match depends on how the guideline defines completion; under a guideline that defines it as the block resting in the tray, this clip is not a match, and the note says which part of the directive was not met.
Failure: a different start frame on every clip
Practice for: marking action boundaries, which the same description lists when it asks for AI-generated annotations to be corrected. The failure is deciding the start and end of an action afresh on each clip.
Worked example. “Arm picks up cup” — does the action start when the arm begins to move, when the gripper touches the cup, or when the cup leaves the surface? Apply the guideline’s definition. Where it is silent, write down the convention you propose, send it to the owner, and apply the answer — yours if accepted, theirs if not — to the whole batch. If the answer changes a convention you had already used, re-assess the clips it touched.
Failure: a gap in the spec nobody reports
Practice for: both labeling kinds — the robot-video descriptions list writing up observations and reviewing other annotators’ work. A guideline that does not cover a recurring case leaves each annotator to decide alone.
Worked example. Partway through a batch you notice the guideline does not cover a case you have now seen several times. You write it up with two invented examples and the two labels you could defend, and send it to whoever owns the guideline.
Checking your own batch
This self-check is ours, not micro1’s, and it follows the stipulation above: the owner answers every flag before hand-in. Each row has two columns that exclude each other, and each row is read on its own. A row you have not checked is at Fail until you check it. A batch with any row at Fail is not ready to hand in; the Fail cell says what to do, item by item — no row asks you to relabel the whole batch.
| Check | Fail | Good |
|---|---|---|
| Covered items | A label differs from what the written rule, or the owner’s answer, says for its case. Correct that label | Every label matches the written rule, or the owner’s answer, for its case |
| Uncovered items | An item no rule covers is unflagged, or is in the batch before the owner has answered. Clarify: flag it and hold it out | Every item no rule covers was flagged and held out until the owner answered |
| Consistency | At least two items of the same kind carry different labels. Re-assess both against the rule, then correct the one that is wrong | Every two items of the same kind carry the same label |
| Boundaries | Start or end frames follow more than one convention, or a convention of yours the owner has not accepted. Correct the frames; Clarify a pending convention with the owner | One convention across the batch: the guideline’s, or one the owner accepted |
How quality is measured is set by each project. The descriptions mention peer review, scoring guides and calibration; they do not say how those are applied. Ask.
Before you accept a project
- Photographs of you are different in kind. The photo listings collect images of your face and body and answers about you, and the family versions need a relative who also agrees. The descriptions we read do not say how long the images are kept or what else they may be used for. Ask, and read the project’s terms on retention and use before you submit anything. Declining this category while accepting others is a perfectly coherent position.
- Content exposure. One description includes flagging material against moderation standards. Ask what a project involves before accepting, and what support exists.
- It is repetitive by design. Several descriptions say so in their own words, asking for consistency through repetitive tasks. Decide whether that suits you before you apply.
- Ask how quality is checked. The descriptions mention peer review, scoring guides and calibration, but not what happens when your work falls short. That is a question for the project.
- Treat a fee to start as a warning sign. That is this guide’s advice about any request for a “training fee” or an “equipment kit”. Apply through official listings.
A short word on the interview
An AI interview — step two of the four micro1 lists, between applying and onboarding — plus, per micro1, human recruiters reviewing profiles against project requirements.
Suggestions from this guide:
- Have a true example of following a detailed specification from a job you held — inventory, lab work, accounting, warehouse — told at the level you may share.
- Describe a time you flagged an ambiguity instead of guessing. If you have none, describe one of the practice exercises above and say it is practice.
- State your equipment and connection plainly if the listing involves video or recording; the photo descriptions ask for your own smartphone.
- State a pace you can keep. The descriptions ask for consistency; they give no target volume.
Pay, hours and countries
The unit of payment changes from listing to listing, and the card does not tell you which one applies. The cards on the opportunities page show an hourly range. The descriptions say:
- photo contribution: a fixed amount per approved task, per participant — approved, not submitted;
- gameplay capture: a contract labeled pay per task;
- music and sound annotation: a stated weekly availability for the engagement and a paid onboarding session, with no pay structure spelled out;
- most of the video annotation descriptions: nothing about how pay is calculated.
So within this page’s listings alone, an hourly range on a card sits on top of pay per approved task in one case, a per-task contract in another, and no stated structure in the rest — where the unit of payment is unknown. Where pay is per approved task, a submission that is sent back may lower what the work pays per hour in effect; the contract is where to confirm how that is handled. Before accepting, ask what the unit of payment is, who approves the work and on what criteria, whether rejected work is paid or can be redone, and whether there is a weekly minimum — and have the answers in writing, not from a call. Do not count on being paid for each hour you spend until the contract says so.
Amounts are left off this page by the blog’s own decision. Independently of that, micro1’s referral terms forbid anyone promoting the program from making promises about compensation, placement or acceptance rates.
micro1 states it does not hire residents of: Afghanistan, Belarus, China, Cuba, Democratic Republic of the Congo, Hong Kong, Iran, Iraq, Libya, Macao, Myanmar, North Korea, the Russian Federation, Somalia, South Sudan, Sudan, Syria, Venezuela, Ukraine and Yemen. Snapshot from the date in Sources. Several descriptions add their own geography — regions for the robot-video and photo listings, an age minimum for the photo listings — so read the location and eligibility lines of each one.
Is a referred annotator assessed differently? We found nothing saying so. This site does not take part in the vetting, and micro1’s referral terms forbid anyone promoting the program from promising placement, compensation or acceptance. If a page tells you a referral link improves your chances, treat that as a warning sign.