"Is data annotation legit" is one of the most searched questions in this field in Canada, and it deserves a straight answer rather than a listicle. The short version is that the established platforms are real companies paying real people, that the field also attracts fraud, and that one test separates them reliably.
I run the contributor side of a speech data company, so I compete with several of those platforms and I am going to be plain about them anyway. This page covers the parts a Canadian page should cover and a generic one cannot: the legitimacy test, how the CRA treats the money, and why a bilingual Canadian is worth more to a speech project than almost anyone assumes.
What the work involves
Data annotation is adding the labels a model learns from, and it covers more ground than most listings suggest: classifying text, marking entities, ranking model answers, transcribing audio, checking transcripts against recordings, marking speaker turns, tagging accent and noise, drawing boxes around objects, and recording prompts in your own voice.
What reaches a human queue is by definition the part a script could not handle, which is why judgement matters more than speed, and why a submission that fails review is the most expensive mistake available to you.
Is data annotation legitimate work?
Yes, as a field. DataAnnotation.tech, Outlier, Appen, Mindrift, Toloka and Prolific are real companies that pay real contributors, and the complaints you will read about them are mostly about task volume and slow support rather than about work going unpaid. Speech-focused crowds operate the same way.
The reason the question gets asked so often is that the field also attracts a lot of fraud, and the fraud imitates the real thing closely. The test that separates them does not depend on brand reputation, which is mostly forum anecdote from people on different projects. It depends on the direction the money moves, which is covered below.
What the work pays
Tasks are priced individually and shown before you accept. Across current projects that works out to roughly 12 to 50 USD per hour, depending on the task type and how scarce your language is. It is an observed effective rate on per-task work rather than a wage, and it is not promised for any particular week, because projects open and pause according to what clients have commissioned.
Payouts are set per project and usually in US dollars, so factor the conversion and any transfer fee into what actually lands in a Canadian account. Ask which payment methods reach Canada and what the minimum payout threshold is before you spend unpaid time on a qualification task, not after.
Canadian French is the strongest card in this market
Speech datasets are chronically short of Canadian French. European French is far better represented, and the two are not interchangeable for training purposes: the vowels, the vocabulary, and the everyday register differ enough that a model trained on Parisian French handles Québécois poorly. That gap is a commercial opportunity for the people who can fill it.
The same logic extends further. Acadian and other regional varieties, English spoken with a French first language, natural code-switching between the two in one sentence, and the many Canadian households speaking Punjabi, Mandarin, Cantonese, Tagalog, Arabic or Spanish alongside English or French are all harder to source than standard North American English. If any of that describes you, it belongs on your profile at the level you genuinely have it.
How the CRA treats this income
This is self-employment income rather than employment income. Nothing is withheld at source, so keep your own records of what you were paid and when, and set money aside as you go rather than discovering the position at filing time.
In practice self-employment income is reported with your personal return, and GST/HST registration only becomes a question above the small-supplier threshold, which most people doing this part-time will not reach. Thresholds and forms change, so confirm your position with the Canada Revenue Agency rather than with a platform. No data company, this one included, is a source of tax advice, and you should treat any that presents itself as one with suspicion.
Checking a platform before you give it your time
One test catches almost every scheme in this field: which way does the money move? A real platform pays you for approved work and never asks you for a registration fee, paid training, a starter kit, an equipment charge, or a deposit to unlock a withdrawal. If money is asked to flow from you to them, it is not work, however polished the site looks.
Before you invest unpaid qualification time, get three things in writing: that the payout is visible before you accept a task, what the review and approval process is, and what the payment cycle and minimum threshold are. A platform that answers all three plainly is a reasonable bet even if its rates are unremarkable. One that answers none is a bad bet at any advertised rate.
How to start from Canada
Build the profile while it is quiet rather than when a project opens. Projects are matched against details already on file, so list both official languages if you have them, name the regional variety rather than just "French", say what you can record with, and describe the room you can actually use. A recent phone with a clean microphone in a quiet room beats good equipment in a room with hard floors and a furnace cycling in the background.