AI interviews: how they work and how to pass one

The practical guide to the AI-run interview that now stands between you and most expert AI-training work

559 of the 3,684 open roles on this site's job board today sit behind an interview run by an AI. They are the open listings at micro1 ( 249), Mercor ( 232) and Braintrust ( 78), three of the platforms that pay doctors, lawyers, engineers, scientists, linguists and writers to train and test AI models, and they include much of the best-paid work on the board: the median listing states $75 an hour at micro1 and $75 at Mercor. Before any of that money is reachable, an AI asks you questions out loud, listens to your answers, follows up on what you said, and turns the conversation into a score that a human recruiter or a client reads later.

The problem is that almost nobody explains how this works. Candidates walk into an AI interview the way they would walk into a call with a person: warm, a little rambling, saying "we" when they mean "I", assuming they can clarify later. An AI interviewer does not reward any of that. It works from a transcript, it compares what you say against your own resume and the role's requirements, and on most platforms you get a small, fixed number of attempts. People who are perfectly qualified fail it every day for reasons that have nothing to do with their expertise, and people who prepare for twenty minutes pass it comfortably.

This guide explains exactly how AI interviews work at the platforms that use them for AI-training work, what the AI can and cannot judge, how to set up and prepare, how to structure an answer so an AI scores it as the expert answer it is, each platform's rules on retakes and AI tools, your legal rights where you live, and what happens after you click submit. Every platform fact comes from the platform's own documentation or from reporting we could verify, and every count of open roles comes from the live job board on this site.

Contents

1. What an AI interview is, and what it is not
2. Where you will meet one: the platforms that screen with AI
3. How you are scored: what the AI can and cannot see
4. Before the interview: setup, practice and your resume
5. During the interview: how to answer like an expert
6. Platform by platform: Mercor, micro1 and Braintrust
7. Integrity: AI tools, proctoring and fake candidates
8. Accents, accessibility and your legal rights
9. After the interview: results, rejections and retakes
10. Where AI interviews are heading
11. A decision framework: which interview to take first

1. What an AI interview is, and what it is not

An AI interview is a screening conversation in which the interviewer is software. You open a link in your browser, allow your camera and microphone, and an AI interviewer asks you a question, usually out loud and on screen at the same time. You answer by speaking. When you finish, the AI either asks a follow-up that builds on what you just said or moves to the next topic. After roughly twenty minutes it ends the session, and your answers become a transcript that is evaluated against the requirements of the role or the platform's general bar. No human is on the call, but on the platforms covered here people at the platform or at the hiring company read the result before anyone is placed on paid work.

The defining feature is that the conversation adapts to you. micro1 says outright that "no two candidates receive the same questions", because its AI recruiter, Zara, mixes technical and conversational prompts "tailored to your selected area of expertise" - micro1. Mercor describes an AI interviewer that poses role-specific questions, generates a transcript of your responses and evaluates your performance for the listing, in an interview that takes about 20 minutes for most people - Mercor. That is what separates an AI interview from the older one-way video interview, where a fixed list of questions appears on screen and you record an answer to each one with nobody, human or machine, reacting to what you say.

It also helps to be clear about what an AI interview is not. It is not a written test: the platforms that screen with written assessments ( Outlier and DataAnnotation are the best-known examples) ask you to complete tasks in a form, and you never talk to anything. It is not a personality quiz: the platforms in this guide evaluate what you say about your work and your field. And on the platforms covered here it is not the final decision: micro1's own research paper on Zara describes the AI-generated assessments as "subsequently reviewed by human recruiters" - micro1 research team on arXiv. The AI produces evidence and a recommendation; people still decide who works on which project.

The flow below is the shape all three platforms in this guide share, even though each adds its own steps around it. Your profile and resume feed the questions, your spoken answers become a transcript, the transcript is scored against criteria, and a human review turns that score into a place in a talent pool. Paid work comes only after that, when a project needs someone with your profile.

From the profile you submit to paid work on a project

Two things in that diagram matter more than they look. First, the resume is an input, not a formality: micro1's Zara adapts its questions to your background, and Mercor attaches your most recent resume to every application next to your interview, so every line on it is something you may be asked to explain, and something a reviewer will read beside your answers. Second, passing is not the same as being paid. A pass puts you into a pool of approved people, and the money starts only when a project that needs your profile picks you. Many candidates who complain that a platform "never got back to them" passed the interview and then sat in that pool because no matching project opened. Section 9 covers how to stay matchable once you are in.

The rest of this guide follows the diagram from left to right. Section 2 maps where you will meet an AI interviewer and what sits behind each one, sections 3 to 5 cover the interview itself, section 6 gives each platform's specific rules, and sections 7 to 9 deal with integrity, your rights and what happens afterwards.

2. Where you will meet one: the platforms that screen with AI

Most of the companies that hire people to train AI fall into one of two camps when it comes to screening. One camp uses written assessments: you complete sample tasks (rate a response, fix a piece of code, answer domain questions in a form) and a grader scores them. The other camp uses AI interviews: you talk, and an AI evaluates the conversation. The second camp is younger, it is growing quickly, and it is concentrated among the expert networks that place credentialed professionals on projects for AI labs, because a twenty-minute conversation is a better test of whether someone really is a pharmacist or a securities lawyer than a multiple-choice form.

On this site's job board, three platforms run AI interviews as their main screen: Mercor, micro1 and Braintrust. Together they list 559 open roles today, of the 3,684 open AI-training roles on the board. Their work is weighted toward the expert end of the market: micro1 lists 28 evaluation roles, 126 in law, medicine and finance and 32 in coding, while Mercor lists 8 in math and science, 119 in law, medicine and finance, 16 in coding, 12 in evaluation and 6 in red-teaming. Of Braintrust's 78 open roles, 40 are language work.

Those counts move every day as projects open and close, and each one opens its current list: micro1's open roles, Mercor's and Braintrust's. What matters for this guide is the shape: one interview unlocks many roles. Mercor keeps your result and reuses it for every later application that needs the same interview - Mercor, micro1 tells you in real time whether you meet its certification criteria - micro1, and Braintrust says its AI skills interview "evaluates your expertise and matches you to appropriate tasks" - Braintrust. An hour of preparation is spread across every role you later apply to.

The pay behind the gate is the other reason to take it seriously. The chart below shows the stated hourly pay on each platform's open listings: the middle of the range (the median listing) and the band that holds the middle half of all listings. The middle half of micro1's stated rates runs from $50 to $270 an hour, and Mercor's from $55 to $150. Braintrust's median, $25, depends on its mix of roles: its language rater roles pay less than its expert roles.

The table below summarizes how the three interviews differ. Each one is covered in detail in section 6, and the full platform reviews (pay, payment schedule, ratings and what has gone wrong) are on the Mercor page, the micro1 page and the Braintrust page.

The differences are smaller than they look. All three want the same evidence: that you have done the work you claim, that you can explain your reasoning clearly, and that you will be reliable on a project. The rest of this guide is about producing that evidence on demand, in a format built for software to read.

If your field is not on these platforms at all, or you prefer written tests to talking, the platforms that screen in writing are covered on the Outlier vs Mercor comparison and in the kinds of AI-training work on this site.

3. How you are scored: what the AI can and cannot see

None of the three platforms publishes its scoring rubric, and anyone who claims to know the exact weights is guessing. What the platforms do publish is enough to reason about the rest. Mercor says it keeps your video and audio recordings and an AI-generated transcript of official interviews "to evaluate your candidacy" - Mercor. micro1 says Zara "evaluates how you think, solve problems, and communicate" - micro1. micro1's published research on Zara describes the interviewer "evaluating their technical competencies and conversational abilities" and dynamically adjusting its questions to the candidate's responses - micro1 research team on arXiv. In every case the raw material is the same: what you said, turned into text.

Start from that fact and most of the advice in this guide follows on its own. A scorer that works from a transcript cannot check your claims against the outside world in the middle of the interview. It cannot call your former employer, read your patents or watch you operate. What it can do is judge the internal evidence in your words: whether you use the terms of your field correctly, whether your explanation of a mechanism holds together, whether you name trade-offs and edge cases, whether the numbers you mention are plausible, and whether what you say matches the experience your resume claims. Vague statements carry almost no evidence, because a bluffer can produce them as easily as an expert. Specific statements carry a lot, because they are expensive to fake. Specificity is the currency of an AI interview.

micro1's paper shows what the output of that judgment looks like. The sample report below is the "vetting results" view of a candidate profile: each skill the interview covered gets a level (Senior, Experienced), the coding exercise and the "human data exercise" get their own levels, communication gets a separate rating, and skills listed on the resume that the interview did not test are marked as not vetted by the interviewer. That last line is worth noticing. Claims you make on paper but never demonstrate in the conversation do not count the same way as the ones you prove.

What an AI interview produces: a per-skill report

Reading the report from the reviewer's side explains a lot of candidate frustration. The person who decides whether you join a project usually sees a summary like this, plus your resume and sometimes the recording, not a transcript of every word. A rambling twenty minutes produces a thin summary even if the expertise was there, and an interview that never touches your strongest skill leaves that skill marked as unproven. The practical goal is therefore not to "sound good" but to put verifiable evidence of your best skills into the transcript, so the summary a reviewer reads is strong.

What the AI is likely weighing

Because the rubrics are private, the safest way to think about scoring is in terms of what a transcript can show. A scorer reading text can count specifics, follow an argument, compare a claim against the resume it was given and notice when an answer drifts away from the question. It cannot see effort, enthusiasm or the expertise you meant to show but never put into words.

The platforms' own descriptions point the same way. micro1 talks about how you "think, solve problems, and communicate", Mercor's glossary describes an interview that assesses "your domain knowledge and communication skills" - Mercor glossary, and Braintrust describes role-specific scoring. The short list below names the dimensions that every one of these platforms describes in some form, in the order that tends to decide outcomes for experts.

• Depth in your field: correct terminology, mechanisms explained, trade-offs named
• Concrete experience: specific projects, numbers, tools and outcomes you can own
• Reasoning out loud: how you approach a problem you have not seen before
• Clear communication: answers that are organized, complete and easy to follow
• Consistency: what you say matches your resume and your earlier answers

The first two carry the most weight for expert roles, because they are what a client is paying for: a pharmacist who explains why two drugs interact, an attorney who names the precedent and its limits, an engineer who explains why a system failed under load. Communication matters because nearly every AI-training task ends with you writing an explanation that a lab will read, and a clear speaker is usually a clear writer. Consistency is the quiet one. An AI that has your resume in context can notice when you describe five years of experience with a tool and then cannot answer a basic question about it, and that mismatch costs more than simply admitting you are less experienced with it.

What the AI is not looking at

Candidates spend a surprising amount of energy on things that transcript-based scoring barely registers. Your background, your clothes, your accent within reason, and whether you smile on camera are not what these systems are built to measure, and Braintrust's AIR product page describes its interviews as conversational voice interviews with role-specific scoring - Braintrust. The camera requirement on Mercor exists mainly to confirm that the person answering is you: Mercor asks that your full face is visible throughout - Mercor. Treat the camera as an identity check, not a stage.

There are two important exceptions. Audio quality matters a great deal, because a transcript built from a muffled microphone is a transcript full of errors, and the scorer judges the transcript, not what you meant. And delivery speed matters indirectly: very long answers bury your best evidence, and very short ones do not give the AI enough material to score. Section 5 deals with both.

The human on the other side

It is easy to imagine an AI interview as a pass/fail machine, but the result is read by people as well. micro1's paper describes AI assessments that are "subsequently reviewed by human recruiters" - micro1 research team on arXiv. Mercor keeps your recordings and transcript to evaluate your candidacy and to recommend your profile to potential employers, and its submitted applications automatically display your most recent resume, interview and form - Mercor. That human reads quickly and in volume, so the most useful mental model is this: you are writing the evidence file a busy reviewer will skim, and the AI is the clerk who files it. Make the clerk's job easy and the file will be good.

micro1's paper also shows that the platform treats feedback seriously. Over a three-day evaluation it looked at 4,820 unsuccessful interviews, of which 10.7% of candidates requested detailed feedback; the feedback system was built to name two or three specific strengths and two or three areas to improve, and 400 candidates rated it 4.37 out of 5 - micro1 research team on arXiv. If a platform offers feedback after a failed interview, ask for it. It is the only window you will get into how your transcript was read.

4. Before the interview: setup, practice and your resume

Most failed AI interviews are lost before the first question. The causes are boring and fixable: a browser the platform does not support, a laptop microphone that turns every answer into mush, a resume that claims skills the candidate cannot talk about for two minutes, or a first attempt spent discovering the format instead of using it. Because the number of attempts is limited (Mercor allows up to three retakes per interview across all the applications that use it - Mercor), every one of these mistakes costs a real chance at the work, not just a few minutes.

The good news is that preparation is cheap and it compounds. A pass on Mercor is reused for every later application that needs the same interview, and a micro1 certification carries across matching projects, so an hour spent getting ready is spread across every role you apply to afterwards. Think of the preparation in three layers: the equipment that produces a clean recording, the practice that removes surprises, and the resume that the interview will be built around.

Get the equipment right

The interview is evaluated from a recording, so the recording is the product. Mercor is unusually explicit about what it needs: the latest Google Chrome is recommended, Microsoft Edge and Safari are supported, and Firefox and mobile browsers are not; you need a computer with a working camera and microphone, your full face visible, and Mercor recommends headphones and an external microphone - Mercor. micro1 runs its interview in the browser on demand, so the same basics apply there - micro1.

The reason headphones matter is less obvious than it seems. Without them, the AI interviewer's own voice comes out of your speakers and back into your microphone, and some setups then cut your audio while the "other speaker" is talking, or mix its words into your transcript. A cheap wired headset solves that and gives the speech-to-text system a cleaner signal at the same time. The checklist below covers the setup that removes almost every technical failure.

• Supported desktop browser, updated, with other tabs closed
• Wired headset or external mic, tested in a short recording first
• Quiet room, door closed, notifications off on every device
• Stable connection, ideally wired or close to the router
• Face lit from the front, camera at eye level, nothing on screen to read

Walk through that list the day before, not ten minutes before, because the fixes (buying a headset, booking a quiet room, finding a laptop that runs Chrome) take time. The last item deserves emphasis: having notes or a script visible is not only a bad habit, on several platforms it is a reason to fail the integrity checks covered in section 7. If you need a reminder of the projects you want to mention, memorize three words for each, not a paragraph.

Practice the format, not the answers

The single best preparation step is free. Mercor offers a practice interview you can take "as many times as you like"; practice interviews are not shared with any companies and do not affect future opportunities - Mercor. micro1 offers an interview prep tool alongside its official guide - micro1. Take the practice version until the format bores you. The goal is not to rehearse answers to specific questions (the questions adapt, so a script will not survive the first follow-up) but to stop being surprised by the rhythm: how long you can speak, how the AI signals it is ready, how it follows up, and what it feels like to talk to a screen without nodding faces.

Practice also lets you hear yourself. Record two or three answers to questions from your field on your phone and play them back. Most people discover the same three habits within a minute: they start with background instead of the answer, they say "we" through the whole story, and they trail off at the end instead of landing a result. Fixing those three habits in practice is worth more than any amount of reading about the interview. Our free AI interview practice helps with the substance: it asks questions in each platform's format, one at a time, and gives feedback on every typed answer.

Make your resume interview-proof

On micro1, Zara adapts its questions to your background - micro1, and on Mercor your most recent resume is shown next to your interview on every application - Mercor. Either way, your resume sets the agenda. Read it as an interviewer would and mark every claim that invites a question: each tool, each result, each title, each number. For every mark, make sure you can talk for two minutes with specifics. If you cannot, either prepare that story or cut the claim; a skill you list and then fail to discuss is worse than a skill you leave off.

Choose your lane as well. Expert networks match people to projects by specialty, and a profile that claims eight fields reads as a generalist with no proven depth in any of them. Lead with the one or two domains where you are strongest and most current, and make sure your resume's first lines say so plainly. If you are a nurse who also codes a little, apply as the nurse first: the healthcare roles on this site pay far more than entry-level coding work, and the interview will be about the field you lead with.

Finally, prepare the boring facts the interview may ask at the end: the hours per week you can genuinely commit, your time zone, when you could start, and your expected rate if the platform asks. Overstating availability is one of the fastest ways to lose a place on a project later, so answer honestly. A modest, reliable commitment is more valuable to a project lead than an enthusiastic promise that disappears in week two.

5. During the interview: how to answer like an expert

An AI interviewer asks two broad kinds of question, and they need different answers. Experience questions ask you to describe something you did ("tell me about a project you are proud of", "walk me through a difficult case"). Expertise questions ask you to explain something in your field ("explain the trade-offs between two architectures", "discuss the significance of Bell's inequalities", to take two of the examples micro1 publishes for its software and physics roles - micro1). Both are scored on the evidence your words contain, but the structure that produces the most evidence is different for each.

The underlying rule is the same for both: lead with the answer, then prove it. A person interviewing you can stop you, nod, or ask you to get to the point. An AI interviewer usually waits until you stop talking, and the scorer reads whatever you said. If the first forty seconds are background, the evidence you meant to give may never arrive before you run out of time or energy. Answer first, in one sentence, and spend the rest of the answer on support.

Experience questions: a tighter version of STAR

The classic structure for experience questions is STAR: situation, task, action, result. It works for AI interviews with one adjustment: most candidates spend far too long on the situation. Give the situation and task in one or two sentences, spend most of the answer on your actions and the reasoning behind them, and finish with a result you can quantify. Say "I" when you mean what you did. "We" is honest when describing a team, but an interviewer, human or AI, is trying to measure you, and "we" makes your part invisible.

A strong experience answer for an AI-training interview has the following shape. It is short enough to hold in your head, and every part of it produces something a transcript scorer can use.

1. The headline: what the project was and why it mattered, in one sentence
2. Your role: what you were responsible for, specifically
3. Two or three actions: what you did and why you chose it over the alternatives
4. The result: a number, a decision, an outcome someone else would notice
5. The lesson: one sentence on what you would do differently

That structure takes about ninety seconds to two minutes to deliver at a normal speaking pace, which is a good length: long enough to carry real evidence, short enough that the AI has time to follow up and the reviewer can absorb the summary. The "actions and why" part is where experts separate themselves. Anyone can say they "improved the process"; an expert says which bottleneck they measured, which fix they rejected and why, and what changed afterwards.

Expertise questions: explain it to a smart colleague

Expertise questions are closer to a short oral exam. The trap is to answer as if writing a textbook definition. What scores well is the reasoning of a practitioner: the core idea stated correctly, the mechanism explained, the conditions under which it breaks, and an example from practice. Picture explaining it to a smart colleague from a neighbouring field, someone who will notice if you skip a step but does not need the basics spelled out.

For AI-training roles in particular, interviewers probe whether you can evaluate work in your field, because that is the job: rating a model's answer, finding the error in a solution, writing the rubric a model is graded against. So when a question allows it, show judgment, not only knowledge. Mention the common mistake people make on this topic, how you would check an answer for it, and what a good answer would need to include. That is exactly the skill a lab is paying for, and an interviewer built by a company that sells that skill is built to look for it.

Follow-ups, pauses and "I don't know"

The follow-up is where AI interviews differ most from recorded video interviews. micro1 says Zara can ask real-time follow-up questions based on your answers and answer clarification questions mid-interview - micro1 research team on arXiv. Treat a follow-up as a good sign: it usually means your answer gave the AI something worth probing. Answer the follow-up directly, at the same level of specificity, and do not repeat your previous answer.

Pauses feel long when you are talking to software, but a few seconds of thought before a hard question costs nothing, and saying "Let me think about that for a moment" is a perfectly normal thing to say to an AI interviewer. When you genuinely do not know something, say so plainly and then show how you would find out: "I haven't worked with that directly. My approach would be to..." An honest gap with a sound method is evidence of the judgment experts have; a confident invention is exactly the kind of answer an expert reviewer will catch later, and on an AI-training project, confidently wrong answers are the problem the lab is paying to remove.

Pace, length and language

Speak at a normal conversational pace, slightly slower than usual if English is not your first language, and finish your sentences. The transcript is the record, and a speech-to-text system handles complete, evenly paced sentences much better than fast fragments. Keep most answers in the range of one to two minutes. If you notice you have been speaking for a long time, land the point: "So the result was..." is always a good way to close.

Avoid filling silence with filler. "Um" and false starts are normal and are not what anyone is scoring, but long, circling answers bury your evidence. If you lose your thread, stop, take a breath, and restate the point in one sentence; that reads far better in a transcript than three minutes of searching. The interview is not a performance. It is twenty minutes in which you put the best, most specific evidence of your expertise on the record, one answer at a time.

6. Platform by platform: Mercor, micro1 and Braintrust

The general advice in sections 3 to 5 applies everywhere, but each platform runs its interview differently, and the differences decide how you should spend your attempts. This section collects what each platform documents about its own interview, what happens after it, and the specific traps people fall into. Where a detail comes only from worker reports rather than the platform itself, it is labeled that way.

A note on scale before the details. These are not small experiments. Mercor alone says someone starts an interview with its AI interviewer every nine seconds, and that around 10,000 of these conversations happen every day - Mercor engineering blog. At that volume, every rule below exists because thousands of people have already tested its edges.

Mercor: one interview, many roles

Mercor's AI interviewer is called Monty. Mercor describes the interview in its glossary as "an automated, on-camera interview used to assess your domain knowledge and communication skills" - Mercor glossary. Its data policy explains where the questions come from: the interviewer "generates questions based on your résumé and the job description, transcribes your responses, and evaluates your performance" - Mercor data and AI usage policy. Most interviews take about 20 minutes, some longer, and the interview is conducted in English unless the role says otherwise - Mercor.

Mercor's engineers have written unusually openly about how the interview works. The dominant type is the Domain Expert interview, which accounts for over 70% of all sessions, and together with code and language interviews it covers about 90% of the volume; for coding interviews, the problem "is generated fresh from that profile and the live conversation" - Mercor engineering blog. The same post explains that Mercor processes your resume before the session begins, and that this context "shapes what gets asked, how deeply to probe, and what to skip." In other words, the interview is built to go deep exactly where your resume makes claims.

The screenshot below, from Mercor's own documentation, shows the interview screen: the AI interviewer in the center, your camera in the corner, a timer with the interview type at the bottom left, and a Report issue button at the bottom right. Some interviews, like the code review interview shown here, run well past twenty minutes and require screen sharing; Mercor notes that only full-screen sharing is supported - Mercor.

Mercor's AI interview screen

The rules that matter most are about attempts. You can retake an interview up to three times across all the applications that use it, so retaking it twice for one job leaves one attempt for every other job that needs the same interview - Mercor new application steps. A retake replaces your previous submission rather than keeping the better one - Mercor how to apply, and an interview that does not meet minimum completion criteria such as duration is not counted as completed - Mercor assessments. If your attempts are used up and a technical failure ruined one of them, you can email support for a one-time extra retake, with a screenshot or recording of the problem, but approval is not guaranteed - Mercor. Mercor's own advice on its assessments page is blunt: "Always aim to do your best on the first attempt, even if retakes are allowed."

Mercor also publishes a handful of practical tips that are easy to miss. Have pen and paper ready (and a calculator if your field needs one); if the AI cuts you off, tell it you were not finished; and ask it to repeat or clarify a question when you need to - Mercor. The interviewer also decides when you have finished speaking from the length of your silence, and extended silences may trigger it to interject - Mercor. If you need thinking time, say so out loud rather than going quiet.

After the interview, the result works in your favor for a long time. Each unique interview only needs to be completed once, and Mercor's example is that an interview completed in January still counts as valid in December - Mercor new application steps. Your recordings and transcript are kept to evaluate your candidacy and to recommend your profile to potential employers - Mercor, and Mercor's engineers write that more than half of its offers now go out proactively, without the expert applying for that specific role - Mercor engineering blog. It usually takes two to four weeks to hear back - Mercor, and Mercor does not share specific feedback - Mercor how to apply.

Two later steps surprise people who passed. Mercor verifies your identity once, through Persona, and that verification is required before you can be onboarded to a project - Mercor ID verification. And nearly every Mercor contract requires the Insightful time-tracking app - Mercor Insightful guide, which may take a few screenshots an hour while you work - Mercor time tracking and pay. Neither is part of the interview, but both are part of the deal, and it is better to know before you invest twenty minutes.

Mercor's interview has also been at the center of its biggest controversy. In March 2026 Mercor said it was hit by a cyberattack tied to a compromised open-source project, and a leaked sample reportedly included videos of conversations between Mercor's AI systems and contractors - TechCrunch. Class actions followed, alleging exposure of interview recordings and facial biometric data - PYMNTS, while Mercor's own update on the incident says that of its nearly five million experts, only a very limited subset had sensitive information affected - Mercor. The full review of Mercor, including its pay and what has gone wrong, is on the Mercor page, and every open Mercor role is on Mercor's list of open roles.

micro1: a conversation with Zara, then certification

micro1's interviewer is an AI recruiter called Zara, and micro1 describes the interview as "a conversational interview with our AI recruiter, Zara," available on demand so you can complete it "anytime, from anywhere" - micro1. The content depends on the role: open-ended technical questions for your chosen field, scenario-based questions about how you think and communicate, a coding challenge for technical roles and a human data exercise for annotator roles. At the end you get a real-time result telling you whether you meet micro1's certification criteria, and if you do not, you can request feedback.

micro1 publishes example questions for some of its most common roles, and they show the level it expects. For a physics AI trainer, one example asks you to "discuss the significance and derivation of Bell's inequalities"; for a full-stack developer, one asks how you would diagnose "a sporadic, high-latency issue in a production system where multiple services and databases are involved" - micro1. These are not trivia questions. They are invitations to reason out loud like a senior practitioner, which is exactly the skill section 5 describes.

The screenshot below comes from micro1's own research paper on Zara and shows the interview from the candidate's side: Zara's question in text beside a recording indicator, a "Done answering? Continue" button that you press to end your answer, your camera in the corner and a countdown timer. That button is worth knowing about in advance. It means you, not a silence threshold, decide when an answer is finished, so you can take a breath and add the result you forgot without being cut off.

The candidate's view of a Zara interview

Passing is the start of a short sequence rather than the end. After you meet the certification criteria, micro1 asks you to complete your profile and verify your identity with a passport, driver's license or national ID: you scan the original physical document with your camera and take a selfie, and IDs in English or Arabic are typically verified within minutes while other languages can take up to 72 hours of manual review - micro1 library. micro1 says the verification is handled entirely by a third-party provider and that it never sees or stores your documents. Once that is done you receive your micro1 certificate and join the certified talent pool.

Certification means you are "eligible for pipeline opportunities": you receive invitations as relevant roles open, and you can apply directly to open roles in micro1's Opportunities Portal that match your skills - micro1 library. Payments are processed in US dollars twice a month - micro1 library. The current list of open micro1 roles, each with the pay it states, is micro1's list of open roles, and the review of micro1 as a platform, including what its workers praise and complain about, is on the micro1 page.

The practical lesson from micro1's design is that the interview checks the field you select, so select carefully. Choose the area where you can answer a senior-level question without notes, not the one with the most open roles, and if you are a specialist in a licensed field, make that the field you interview for: micro1's clinician, legal and finance roles are some of the best-paid listings on this site, and the roles in law, medicine and finance page shows how many are open today.

Braintrust: an AI skills interview that sets your level

Braintrust runs its interviews on its own AI recruiter product, AIR, which it sells to employers as "AI interview software for enterprise screening" with "conversational voice interviews, role-specific scoring, and ATS-synced rank lists" in 16+ languages - Braintrust. For people joining its AI-training work, Braintrust describes a three-step path: create or complete your profile (about 20 minutes for a new member), complete the AI skills interview, in which "our AI evaluates your expertise and matches you to appropriate tasks," and then browse and accept tasks - Braintrust.

The detail in that description that matters most is the matching. On Braintrust the interview does not only decide whether you are in; it decides which tasks, and which pay grade, you are offered. That changes the preparation. A Braintrust interview is the moment to make your most advanced expertise visible, because an answer pitched at a generalist level can place you in generalist work. If you are a credentialed specialist, say so early and keep returning to the specialist depth in every answer.

Braintrust's open roles on this site include language work ( 40 of its 78 open roles today), alongside evaluation, audio and expert roles, and its language rater roles pay less than its expert roles. Check Braintrust's open roles for the roles in your field before you interview, and read the Braintrust page for how and when it pays, since it is the platform where the gap between "approved" and "earning" depends most on applying to specific jobs after you pass.

What the three have in common

Put the three platforms side by side and the shared logic is clear. Each one uses a short spoken conversation to collect evidence of expertise at a scale no team of human recruiters could match, each one ties the result to a verified identity, and each one reuses the result to match you to many future projects. The differences are in the mechanics: Mercor's strict attempt limits and long-lived results, micro1's certification and on-the-spot feedback, Braintrust's use of the interview to set your level.

For you, that means one preparation routine serves all three, with a small adjustment for each. On Mercor, protect your attempts and do your best on the first one. On micro1, choose your field deliberately and ask for feedback if you miss. On Braintrust, aim every answer at the most advanced level of work you can genuinely do. In every case, the work you put into your resume and your stories pays off again and again.

7. Integrity: AI tools, proctoring and fake candidates

Every candidate thinks about it at some point: the interview is run by an AI, so why not have another AI help with the answers? The short answer is that the platforms forbid it, they are building systems specifically to catch it, and getting caught costs far more than a failed attempt. The longer answer is more interesting, because it explains why these platforms care so much about integrity in the first place.

Their entire business is selling verified human expertise to AI labs. A lab that pays an expert network for a physician's judgment is paying precisely for the part a model cannot provide; if the "physician" is reading answers from a chatbot, the network is selling the lab its own model's output back at a premium. So integrity is not a side rule for these companies, it is the product. Mercor's rules say you may run grammar and wording checks but must not use large language models "to compose your responses, rate model responses, generate rationales or explanations, or evaluate code" - Mercor, and its policy on AI use says that submissions "evidently produced by artificial intelligence will be identified and flagged accordingly" - Mercor LLM usage policy.

How platforms look for it

It also helps to understand what "cheating" means from the platform's side, because it is broader than most candidates assume. Reading a prepared script, having someone else in the room prompt you, pasting a question into a chatbot on a second device and relaying the answer, or letting someone else take the interview under your name all fall on the same side of the line. Using a calculator or a notepad where the platform allows it does not: Mercor's own tips suggest having pen and paper ready.

Detection is mostly automated, and it gets better with volume. Mercor's chief executive has written that its fraud team "doesn't manually catch people cheating in interviews" but builds evaluations that teach an AI agent to spot the patterns - Mercor. Mercor also blocks sign-ins through VPNs, proxy services and other anonymizing networks since July 17, 2026 - Mercor account security and conduct. And both Mercor and micro1 tie every result to a verified government ID and, at micro1, a selfie matched to the document, which makes the classic fraud (someone else taking the interview for you) very hard to sustain.

The patterns these systems look for are the ones you would expect from someone reading or relaying answers. The list below is not a platform's published rulebook, which none of them provides; it is the set of signals that follow from how a cheater has to behave, and every one of them is also simply bad interview technique.

• Reading cadence: even, fluent delivery that never pauses to think
• Answer latency: long silences before every answer, then polished prose
• Off-screen glances: eyes repeatedly moving to another screen or a phone
• Generic depth: textbook-correct answers with no personal detail
• Inconsistency: answers that do not match your resume or earlier replies

The last two are the most reliable for a scorer, because they do not depend on camera tricks at all. A chatbot can produce a correct explanation of a concept, but it cannot produce your own project, your own mistakes and your own numbers, and that is exactly what a good AI interviewer probes for with follow-ups. That is also why the honest strategy and the winning strategy are the same: the evidence an AI interviewer rewards is the evidence only you can provide.

Fraud is real, and it is why identity checks keep tightening

It is easy to read strict identity checks as distrust, but the numbers explain them. In a Gartner survey of 3,000 job candidates, 6% admitted to taking part in interview fraud, either posing as someone else or having someone else pose as them, and Gartner predicts that by 2028, 1 in 4 candidate profiles worldwide could be fake - HR Dive. Platforms that pay strangers on the internet $70 or more an hour for expert work are the most attractive target for that kind of fraud, so expect identity checks to become stricter, not looser.

Fraud also explains a few rules that otherwise look petty. Mercor asks candidates not to share specific interview questions or screenshots - Mercor, because a question bank that circulates online stops measuring anything. Interviews adapt to each candidate partly for the same reason: a conversation built from your own resume cannot be memorized from a forum post.

For an honest candidate, the practical consequence is to keep your identity consistent everywhere: the same legal name on your profile, your resume, your ID and your payment account, a stable location, and no VPN while you use these platforms. Most "my account was flagged" stories in worker forums trace back to one of those mismatches rather than to anything said in the interview. If you share a computer or move between countries, tell the platform's support team before it becomes a problem, not after.

Integrity rules protect honest candidates more than they constrain them. Every fake profile that slips through takes a project slot from a real expert and makes clients trust the whole pool less, which lowers the rates everyone in it can charge. The next section turns to the opposite risk: honest candidates who are misjudged because the technology hears them less well than it hears others.

8. Accents, accessibility and your legal rights

AI interviews are built around speech, and speech technology is not equally good at hearing everyone. In research published in 2025, Dr Natalie Sheard of the University of Melbourne reported that HireVue, one of the largest interview-software vendors, cites transcription error rates of less than 10% for US English speakers but as high as 22% for non-native speakers with accents from other countries - ACS Information Age. A transcript with one word in five wrong is a weaker record of what you said, and the scorer judges the record.

Candidates are skeptical for good reason: in Gartner's survey, only about a quarter said they trust AI to evaluate them fairly - HR Dive. You cannot fix the technology, but you can reduce the gap it creates. Speak at an even pace, finish sentences, use the precise technical terms of your field (which speech systems often transcribe more reliably than casual phrasing), and use a good microphone close to your mouth. If a question is unclear, ask the AI to repeat it; Mercor explicitly encourages candidates to ask for repetition or clarification - Mercor.

Language options

Language support differs by platform, and it is worth checking before you start. Mercor says its interviews are conducted in English unless the role specifies otherwise - Mercor, which means a language-specific role may interview you in that language. Braintrust's AIR product supports interviews in 16+ languages - Braintrust. If you are applying for a role that depends on your native language, the interview for that role is likely to test it, and the multilingual roles on this site show which languages have open work today.

Accessibility and accommodations

If you have a disability that affects speaking, hearing, reading or processing time, contact the platform's support team before you start the official interview and ask what accommodations are available. Do it in writing so there is a record, and do it before the attempt, because an attempt that has started usually counts. The same research that found accent gaps also warned that systems built around an "ideal" candidate can disadvantage people with disabilities - ACS Information Age, which is exactly why asking early matters.

What the law says, by place

The rules that protect candidates depend on where the job is and where you live, and they are changing quickly. None of the laws below bans AI interviews; they require notice, consent, audits or human oversight. The summary that follows covers the four regimes most readers of this site fall under, and it is general information, not legal advice.

New York City's rules have been enforced since July 5, 2023 - NYC Department of Consumer and Worker Protection. Illinois' law requires employers who use AI to analyze video interviews to notify applicants, explain how the AI works and what general types of characteristics it evaluates, obtain consent, and delete the interview within 30 days of a request - Frost Brown Todd. The EU lists AI tools for employment and recruitment among its high-risk uses - European Commission, and since the AI Omnibus entered into force on 27 July 2026 those rules apply from 2 December 2027 - European Commission. GDPR's Article 22 gives people in the EU and UK the right not to be subject to a decision based solely on automated processing that significantly affects them, with exceptions - GDPR text.

What does this mean in practice for someone applying to an AI-training platform? Mostly that you can ask questions and expect answers: what data the interview collects, how long recordings are kept, whether a person reviews the result, and how to request deletion. Mercor, for example, states that it does not sell interview data - Mercor. Read each platform's privacy terms before you start, because platforms that work with contractors in many countries do not all apply the strictest rule everywhere.

9. After the interview: results, rejections and retakes

What happens after you click the final button depends on the platform, and knowing the timeline saves a lot of anxiety. micro1 gives a real-time result at the end of the interview and lets you request feedback if you do not meet the certification bar - micro1. Mercor usually takes two to four weeks to respond - Mercor and does not share specific feedback - Mercor how to apply. Braintrust uses the interview to match you to tasks at a level, so the result shows up in the work you are offered rather than as a score.

Silence after a pass is the most common source of confusion. On every one of these platforms, passing puts you in a pool, and paid work starts only when a project that needs your profile opens. Mercor's engineers say more than half of its offers now go out proactively to experts in its pool - Mercor engineering blog, and micro1 sends invitations to certified experts as relevant opportunities become available - micro1 library. Passing and then hearing nothing for a few weeks usually means no matching project has opened yet, not that you did something wrong.

Staying matchable after you pass

The work you do after the interview decides how quickly the pool turns into income. Keep your profile current, because matching runs on it; apply to specific open roles rather than waiting for invitations, since every platform lets certified or approved people apply directly; and keep your availability honest and up to date. If your situation changes (more hours, a new certification, a new language), update the profile the same day.

It also helps to watch the job lists for your field. When a project in your specialty opens in Mercor's open roles or micro1's, apply promptly, because Mercor warns that listings can close at any time - Mercor how to apply. A free weekly email of new roles in your kind of work is available on the job alerts page, which saves checking several job boards every day.

When you fail: what to change before you retake

A failed attempt is information, and the worst response is an immediate retake with the same approach. Mercor caps retakes at three per interview across all applications - Mercor, so a rushed second attempt burns a scarce resource. If the platform offers feedback, as micro1 does, read it first. If it does not, review your own attempt honestly against the structure in section 5: did you lead with answers, stay specific, say "I", and give results with numbers?

Then fix the cause, not the symptom. If your answers were vague, prepare five stories with numbers before the next attempt. If a technical problem broke the session, fix the setup first; on Mercor, a one-time extra retake for technical failures can be requested once your three attempts are used up, and it is not guaranteed - Mercor. If the interview exposed a gap in the field you selected, consider interviewing for an adjacent field where you are stronger. And take the free practice interview again before the official retake; it costs nothing and does not count.

Beware of fake "interviews"

The popularity of AI interviews has created a new kind of scam. Impostors pose as recruiters for well-known platforms, send candidates to fake "interview" or "onboarding" pages, and then ask for a fee, banking details or a copy of an ID outside the platform's own verification flow. Real platforms run their interviews inside their own logged-in sites, and none of the three in this guide charges you to interview. The Mercor review on this site describes how impostors have used its name, and the same rules apply everywhere: never pay to apply, never move the process to a messaging app, and never hand over documents outside the platform's own verification step. If an invitation looks off, paste it into our scam checker, which checks its links and sender against each platform's real addresses.

10. Where AI interviews are heading

The economics explain why AI interviews spread so fast, and why they will not go away. A human screening interview costs a recruiter's time, which puts a hard ceiling on how many candidates a company can talk to. An AI interview costs a few minutes of computing, so the marginal cost of one more conversation is close to zero. That changes who gets interviewed at all: instead of a recruiter skimming résumés and calling the ten that look best, a platform can talk to everyone who applies and judge each one on what they actually say. Mercor's growth shows the effect: it went from a few hundred interviews a week to over ten thousand a day during 2025 - Mercor engineering blog.

The second force is demand. AI labs need large numbers of people with genuine expertise to write, judge and correct the data their models learn from, and they need them fast and in narrow specialties. That is the market these platforms serve, and it is large: Mercor raised money at a $10 billion valuation in October 2025 - CNBC and was reported in July 2026 to be in talks at $20 billion - TechCrunch. The chart below, from Mercor's own engineering blog, shows what its daily interviews are mostly about: expertise in a domain, far ahead of language and coding.

Three changes follow from those forces, and each one matters to candidates. First, interviews will go deeper into your field, because the cost of a longer, more specialized conversation is small and the value of a correctly matched expert is large; expect more domain-specific interviews like Mercor's code review interview, and more interviews that include a live task. Second, identity and integrity checks will tighten, as the fraud figures in section 7 make inevitable, so a consistent, verified identity becomes part of your professional profile on these platforms. Third, regulation will catch up: New York City and Illinois already set rules, and the EU's high-risk rules for employment AI apply from December 2027, so expect clearer notices about how interviews are scored, longer data-retention disclosures and, in some places, a right to ask for human review.

What does not change is what the interview is for. Platforms use AI interviews to find people whose judgment in a field is real and whose explanations are clear, because that is what they sell. A candidate who brings verifiable expertise and explains it well will do better as interviews get more sophisticated, not worse. The preparation in this guide is not a trick for beating a machine; it is a way of making real expertise legible to one.

It is also worth being clear-eyed about the limits. A transcript-based interview will always favor people who explain well under time pressure, and it can disadvantage strong experts who do not, including many non-native speakers and people with some disabilities. The platforms have every commercial reason to reduce those errors, because a missed expert is lost revenue, but they have not published audits showing how well they do. Until they do, the best protection is the one this guide keeps returning to: a clean recording, a clear structure and specific evidence that leaves a scorer little room to misread you.

11. A decision framework: which interview to take first

If you qualify for more than one of these platforms, the order you take the interviews in matters, because each one is practice for the next and because your first result shapes how quickly you reach paid work. The right first interview is the one where your strongest, most current expertise matches the most open roles, so that a pass turns into a match quickly and you build momentum and income before attempting the others.

Use the live job lists rather than reputation to decide. On the day this guide was published, micro1's open roles leaned toward evaluation work, law, medicine and finance, and coding, Mercor's toward math and science, law, medicine and finance, coding and red-teaming, and Braintrust's toward language work. Those mixes change every week, so check micro1's open roles, Mercor's and Braintrust's for roles in your own field before you choose, and look at the pay each listing states on the live pay table.

Start where your strongest expertise meets the most open roles

Whichever you choose, the sequence that works is the one this guide has built up: fix the equipment, take the practice interview until the format is familiar, make the resume interview-proof, answer with the headline first and specifics after, and treat a follow-up as an invitation. Do the first official interview only when a practice attempt feels routine, use the feedback or your own notes to adjust, and take the next platform's interview within a week or two while the habits are fresh.

For people with licensed expertise, it is worth knowing how much those fields pay before you decide where to start. The attorney roles, healthcare roles and finance roles pages on this site list every open role in those professions with the pay each one states, which is the clearest way to see whether an AI-interview platform or a written-assessment platform has more of the work you want this month.

How this guide is researched and kept current

Every count of open roles and every pay figure in this guide comes from the live job board on this site, which re-reads each platform's public job feed every morning, and is read again each time this page loads: each one opens the roles or the pay it counts. The platform rules (interview length, browsers, retakes, rules on AI tools) were read from each platform's own documentation on the same day, and anything we could only find in worker reports is labeled as such. AITraining.jobs is built and run on Founden, an autonomous software engine where one AI runs a whole company, from the website and its live data to the back office; Founden is run by the same founder as this site. If a platform changes its interview, the numbers and rules in this guide are what change with it, so check the date at the bottom before relying on a specific rule.

This guide reflects the AI interviews used by Mercor, micro1 and Braintrust as of October 2026. Platforms change their interview formats, retake rules and pay often, so confirm current details on each platform's own pages before you start an official interview.