Before accepting AI-training work, check the rate, paid hours and terms

Read the pay figures as advertised ceilings

Kei Watanabe and Kazuki Nakayashiki's Glasp report recorded 117 pay observations across 19 platforms on October 15, 2026. Its pay medians use 103 confirmed hourly observations in US dollars. The reported median advertised ceiling was USD 150 an hour for coding and USD 20 for voice work. These describe the upper rates in the sampled advertisements. They do not measure money received by contributors or hours offered to them.^[1]

The platforms also have uneven weight in that sample: Glasp counted Mercor listings individually, while DataAnnotation published a rate per tier. Glasp says it found no study of realized hourly earnings on the expert platforms of 2025 and 2026. That is a limit of the evidence the authors found. It does not justify using an advertised maximum as a typical rate, a Canadian pay benchmark or a guaranteed offer.^[1]

Glasp's wider company research uses a purposive, mostly English-language sample and model-coded company profiles. The authors report additional model checks without human review of the labels. Those company checks do not turn the pay observations into a representative earnings survey. Glasp also discloses that it is studying possible future partnerships to help people find paid AI-training work. Read the report as attributed research with stated limits.^[1]

Confirm your rate and country eligibility

Start an offer record with the project name, contracting company, task type, rate, currency and date confirmed. Ask whether the amount applies to your actual assignment, rather than to another skill tier or a recruitment advertisement. Record any lower rate for training, review or other activities. Glasp identifies rate changes between projects as one reason a posted rate can differ from a contributor's eventual pay.^[1]

Match the assignment to work you can do. Glasp describes engineers writing code, doctors grading medical answers and speakers recording conversations. Ask what you will produce, which skills the task requires and how the work will be assessed. For career planning, record the task you completed and the skill you used, without retaining restricted material. This is a suggested way to describe relevant work truthfully, not evidence that an assignment improves your chances of an engineering job.^[1][2]

For work from Canada, ask whether the specific project accepts contributors in your province, what work authorization it requires and which payment methods you can use. This is a project check, not a finding that the report audited Canadian access. Keep USD and CAD separate. Ask about conversion charges and withdrawal fees before comparing the offer with your budget. Outlier's terms, for example, require an approved payment method in US dollars or the specified currency and place foreign-exchange fluctuation risk outside the company's responsibility.^[1][2]

Separate a paid hour from an available hour

Ask which activities are paid: the qualification test, general onboarding, project training, reading instructions, revisions and completed tasks. Request the expected time for each unpaid step and whether test work can be used if you are not accepted. Then ask about task availability for your field and country, any weekly cap and whether hours are guaranteed. An unanswered availability question leaves expected weekly income unknown.^[1]

Before starting, set your own limit on unpaid qualification time. If you proceed, keep a private record of time spent, accepted paid work and money actually received. For a personal review, divide receipts after platform and payment fees by all the time you spent on that work, including unpaid tests and task searches. Keep the currency and period consistent, and track taxes separately. This is an editorial record-keeping suggestion, not a result measured by Glasp or a platform earnings estimate.^[1]

Read payment and rejection rules before doing tasks

Get the payment schedule, minimum withdrawal, task-acceptance rules and dispute route. Ask when a task becomes payable, what can cause rejection, whether rejected work may still be used and how to challenge a decision. Save the version of the terms and rate confirmation that apply when you start. These are practical checks based on the gaps Glasp identifies between advertised rates and payment.^[1]

Outlier's Terms of Use, effective May 11, 2026, give one concrete example of why this matters. Section 3(b) links payment to completion and verification that tasks meet requirements. Subject to applicable law, the company reserves the right to delay or suspend payment during a compliance investigation. The terms say a violation may lead to non-payment of some or all accrued amounts. This is the wording of one provider's agreement, not evidence that every contributor experiences withheld pay.^[2]

Ask what happens if a project ends: how much notice you receive, what happens to submitted work, when final payment is due and whether another project has a different rate. Outlier's section 3(a) describes an independent-contractor arrangement without employment benefits or a right to perform tasks for a particular period. That contract description does not settle your legal employment status in Canada or another jurisdiction. Seek local advice if classification or a dispute affects your decision.^[1][2]

Check rights before sharing work or recordings

Before providing ID, voice, video or recordings from your home, read what is collected, who receives it, how long it is kept and what withdrawal or deletion options exist. Do not record other people without their consent. Ask whether qualification material is covered by the same rights terms as paid tasks. Glasp raises these questions because human-data work can involve both the material you create and information about you.^[1]

Read ownership and confidentiality terms before using existing code, client documents or examples from your job. Outlier's section 4(b), for example, assigns rights in defined work product to the company; its terms also restrict confidential information. Ask for written permission before retaining task material for a portfolio. Do not supply employer or client material that you lack permission to share. A high advertised rate is no reason to accept unclear rights obligations.^[2]

Send one request, then decide what remains unknown

You can adapt this original request: Please confirm the project and task type, my rate and currency, country eligibility, which tests and activities are paid, expected task availability and any hour cap. Please also send the payment and rejection rules, appeal contact, project-end terms, and the policies for my data and work. If a detail is unknown, please identify it before I begin. This template is editorial guidance based on the report's checks, not a tested method for obtaining work.^[1][2]

Put the response beside your offer record. If the rate is clear but paid hours are unknown, keep income expectations open. If payment or rights terms remain unclear, pause before supplying more work or sensitive data. Keep AI-training projects separate from your salaried-job search and describe any completed work truthfully. The useful decision is whether the specific assignment fits your time, payment needs and obligations, with unresolved terms recorded rather than assumed.^[1][2]