AI Training Platform Screening: 5 Gates, and 1 You Only Get Once

Take-home tests stopped working. At Deel, 85% of candidates were using AI on them. After switching to a live AI interviewer, the share of candidates who went on to pass a human interview went from 10% to 50%, and recruitment costs fell by more than 80%. Those numbers come from a case study about micro1's interviewer, and they explain why every platform in this market now screens the way it does.

So the gate moved. AI training platforms no longer ask you to submit work on your own time. They put you in a live conversation, watch your screen while you code, and grade a work sample against a rubric.

Five platforms, five gates, and the difference that should decide your strategy is not difficulty. It is how many attempts you get. DataAnnotation's own FAQ says you can take its Starter Assessment once, full stop. Alignerr's own FAQ says taking more assessments improves your odds of being matched. Those are opposite games, and most people play them in the wrong order.

This guide covers what each gate actually tests, how many attempts each one gives you, the order to apply in, and the single preparation that transfers across all of them.

The three mechanisms, and what each one really tests

Every screen in this market is a combination of three things.

1. The live AI voice interview. Mercor, micro1, and some Alignerr listings. A real-time conversation that reads your resume beforehand and generates follow-ups from your own answers. It is not testing recall. It is testing whether you can reason out loud, in your own field, without notes.

2. The unpaid work-sample assessment. Outlier, DataAnnotation, Alignerr's qualification tasks. A handful of tasks that mirror the real job, graded against a rubric. It is not testing whether you are smart. It is testing whether you follow a stated instruction exactly, which is the actual job.

3. Proctored integrity monitoring. micro1's Ava model, during the coding assessment. Gaze direction, tab switching, external monitors, browser extensions, fused into an integrity score. It is not testing anything about you. It is testing your desk.

The failure modes are completely different, and so are the fixes. People fail the first on silence and padding their resume, the second on skimmed formatting rules, and the third on a copilot extension they forgot to disable.

Attempts: the variable nobody compares

This is the table that should shape your plan.

• Alignerr. Effectively unlimited. It runs as a job board, so a listing that does not convert is followed by another listing tomorrow. Alignerr's FAQ states that taking more assessments and performing well increases your chances of being matched quickly, and that assignments are based on skills, assessment results, availability and past performance. Effort compounds here.
• Outlier. Per subject. Assessments are taken for each subject you want to work in, so failing one does not close the others. Outlier says its onboarding flow runs 30 to 90 minutes end to end.
• Mercor. Three attempts, shared across every application needing that interview, and only your most recent attempt is evaluated. There is also a free, unlimited practice interview that is never shared with companies.
• micro1. No published retake or cooldown policy at all. It does ship a free Interview Prep simulator. Plan on one attempt.
• DataAnnotation. Exactly one, ever. Its FAQ: you can only take the Starter Assessment once, so review carefully before submitting. No score, no feedback, and duplicate accounts to retry get banned.

Two platforms give you room to be bad at this while you learn. Two give you a little. One gives you nothing.

Apply in this order

The strategy follows directly from the table above: spend your reversible attempts first, and arrive at the unrepeatable ones already practised.

1. Alignerr. Lightest onboarding in the market, and the only place where trying more times measurably improves your odds. You also meet the AI interview format here at zero risk.
2. Outlier. Per-subject assessments mean a bad one costs you a subject, not the platform. Qualify in two or three while you are already in assessment mode, which is also the fix for quiet weeks later.
3. Mercor. Take the free practice interview first, then the real one. Three attempts, most recent scored, so do not retake a good result chasing a marginal improvement.
4. micro1. Run the free Interview Prep simulator, clear your desk properly for the proctored half, then take it once as though it is the only attempt, because it may be.
5. DataAnnotation. Last. One hour, one attempt, no feedback, and by this point you will have been through four screens and know exactly how these rubrics read.

The counterargument is that DataAnnotation has the lowest entry bar and the fastest path to actual money, so why wait. Because the bar being low does not make the attempt repeatable. A wasted hour on DataAnnotation is permanent in a way a wasted hour anywhere else is not.

Alignerr and micro1 may be running the same interviewer

A detail worth knowing before you prepare twice.

Alignerr's own FAQ says some jobs may require additional steps like a Zara interview or a skills assessment. Zara is micro1's AI interviewer. micro1 licenses it to outside companies, and its own case studies name Deel and Legal Soft as customers, including the Deel figures at the top of this article. Alignerr does not publicly name its vendor, but the trademarked name belongs to micro1's product.

The practical consequence: the free Interview Prep tool micro1 ships is rehearsal for two platforms, not one. And if you follow the order above, you will have already sat this exact format on an Alignerr listing before it counts at micro1.

Full mechanics of that interview, including micro1's published research on how it scores you, are in the micro1 AI interview guide.

The Outlier assessment

Outlier's onboarding is an account, an expertise selection, skill screening assessments and identity verification, which Outlier says takes 30 to 90 minutes in total. You need a valid ID and a mobile phone from your country of residence, a current resume and a LinkedIn profile. Most tracks expect at least an associate degree; specialist tracks in law, medicine, math and science expect a master's, a PhD, or equivalent professional standing.

The assessments are unpaid, one per subject, and they grade rubric-following at least as heavily as subject knowledge.

• Treat the rubric as the entire job description. The grader is checking whether you applied it exactly as written, not whether you would have written it better.
• Justify every rating. A rating without a specific, concrete reason scores as incomplete.
• Read the formatting instructions twice. They are stated once, and skimming them is the most common failure.
• Qualify in more than one subject. This is the fix for the quiet weeks that follow, and it costs least while you are already in assessment mode.
• Expect silence. Outlier does not always send a rejection. No task access about two weeks later usually means no.

Why Outlier's volume matters less than its screen right now is covered in the Outlier AI review.

The Alignerr interview and qualification tasks

Alignerr's application is the fastest here: sign in with Google, upload a resume that auto-fills your profile in about 30 seconds, set your preferences, verify identity through Persona, then browse listings and apply. Billing and contract setup happen after approval.

Screening is per listing rather than per platform. Some listings add a Zara interview, some add a skills assessment, some add qualification tasks that are short samples of the real work graded against the project rubric.

Two failure modes, both fixable. Speaking in generalities during the interview instead of reasoning through a specific case. And rushing qualification tasks that are graded on exactness rather than speed.

The structural advantage is the one Alignerr states itself: more assessments, better matching. On a platform of more than 100,000 experts, that is the closest thing to a lever you get. What to watch instead is how you are paid once you are in, which the Alignerr review covers, because two different payment models are in play and only one pays for your time.

The DataAnnotation Starter Assessment

One hour. One attempt. No exceptions.

DataAnnotation's FAQ is unambiguous: you can only take the Starter Assessment once, so review carefully before submitting. Specialist assessments in coding, math, chemistry and similar run one to two hours under the same rule. Approval notification arrives within a few days, and DataAnnotation says it cannot respond individually to every applicant, so there is no dashboard, ticket or escalation path.

The stated bar is a bachelor's degree or equivalent real-world experience, English fluency and reliable internet. There is no credential check at the base tier, which is exactly why the assessment carries all the weight.

How to spend the single attempt:

• Read every instruction twice and treat it as a contract. Most failures are formatting rules stated once and ignored once.
• Answer completely. If a task asks for a rating and a justification, both are scored. One without the other fails.
• Write plainly. Padding a short answer to look thorough reads as noise.
• Do it rested, in one sitting, on a real computer. There is no pause and no way to explain a technical failure afterwards.
• Do not run answers through an LLM. The entire product is human judgment models lack, and generated prose is the specific thing this assessment is built to catch.

Rates, payment mechanics and what the work is actually like are in the DataAnnotation review.

Mercor and micro1: the two hardest gates

Both screen with a live AI interview, and both have deep-dive guides because the mechanics matter more than the questions.

Mercor publishes its rules. Three attempts with only the most recent scored, an unlimited free practice interview, an automated Application Fit check that can block a submission outright with no manual override, and a paid work trial of four to six hours before some contracts. You get one paid work trial per person ever, and accepting one invalidates the rest. See the Mercor AI interview guide.

micro1 publishes research instead. Its own study of 800 candidates found the score built from your live interview predicted actual hiring far better than a resume score, and the two correlate at r = 0.19, meaning they measure nearly unrelated things. It also documented 4,820 unsuccessful interviews in a three-day window, of which only 10.7% requested the free feedback report. See the micro1 AI interview guide.

The reviews behind both, covering pay and what happens after you pass, are the Mercor review and the micro1 review. Every platform in the market is ranked side by side in the best AI training jobs roundup, and the lighter end of the market is covered in the Remotasks review.

What every gate is actually scoring

Strip away the formats and the same two things are being measured everywhere.

Can you follow a rubric exactly? Every work-sample assessment in this market grades instruction-following before subject knowledge, because the job is applying someone else's quality standard consistently across thousands of items. Improving your answer beyond what the rubric asked for is a failure, not a flourish.

Can you explain your reasoning out loud? Every AI interview scores how you think rather than what you recall. The pattern that works is the same in all of them: state your position in one sentence, explain what you considered and ruled out, anchor it with a real number or constraint, then stop cleanly.

Two things follow. Do not name a technology, method or subject you cannot defend for two minutes, because adaptive questioning turns every mention into a probe. And never think in silence, because these systems treat a pause as the end of your turn.

The preparation that transfers

The formats differ. The skill does not. Five platforms, one gate: a scored conversation where you explain your reasoning to something that asks follow-ups.

Most people have never practised that. Their recent experience is written applications and task queues, and the first time they hear an AI interviewer ask why they ruled something out is the attempt that counts.

The free practice tools help with nerves and setup. Mercor's practice interview is unlimited and never shared with companies; micro1's Interview Prep simulates the format. Neither tells you whether your reasoning holds up, because neither is there to coach you.

That is the gap worth closing before the attempt you cannot repeat. A scored mock interview puts you in the same conditions with feedback attached: follow-ups built from what you just said, and a transcript you can read afterwards to find where you went vague. One evening, and it applies to all five gates rather than one.

When you are ready to apply, the AI training jobs board collects live listings from Mercor, micro1, Alignerr, Outlier, DataAnnotation and Terac into one feed with published rates, refreshed daily.

Verdict

The screening in this market has converged, and it converged because take-home tests broke. What is left is a live conversation and a rubric, at five companies with wildly different tolerance for you getting it wrong.

Apply where mistakes are cheap first. Alignerr rewards volume outright. Outlier isolates failure to one subject. Mercor gives three attempts and an unlimited rehearsal. micro1 gives a simulator and no stated second chance. DataAnnotation gives one hour, once, and never explains.

Do them in that order and the only unrepeatable attempt in this market is the one you take last, with four screens behind you.