There is a category of remote work that has appeared in the last two years and that almost nobody here is talking about properly: training and evaluating AI models. The rates look made up. They are not — but the entry bar is real too, and most of what gets shared about this work online skips that half.
We keep a job board, InformMe, and this is what is currently on it. Here is an honest read of it.
What the work actually is
It is not "using AI". It is being the expert a model is measured against.
You are given model output in your own field and asked whether it is right, where it is wrong, and what the correct answer would have been. Sometimes you are writing the problems the model is tested on. Sometimes you are labelling source material so it can be trained in the first place. The common thread is that the value is your subject expertise, not your typing.
That is why the rates are what they are, and also why the screening is harder than a normal freelance gig.
What the roles pay
Straight from the current listings, all remote, all contract:
- Actuarial & Insurance AI Partner — $70 to $150 an hour. Reviewing AI pricing and reserving models, annotating claims documents.
- Financial Analyst AI Specialist — $70 to $150 an hour.
- Function-based Code Annotator — $120 to $150 a day. Designing algorithm problems and the test data that grades them.
- Repo-based Code Annotator — $80 to $120 a day. Building reproducible Docker environments that replicate known bugs.
- Generalist AI Trainer — $30 to $65 an hour. Writing and ranking model responses across general subjects.
- AI Trainer, Image/Video — $25 to $45 an hour.
- Photoshop AI Specialist — $45 to $65, by project and skill level.
At roughly 130 shillings to the dollar, the generalist trainer band alone is about 3,900 to 8,400 shillings an hour. That is the number that makes people assume it is a scam. It is simply what specialist annotation costs in the market that buys it.
Treat the ranges as ceilings you work up to, not starting salaries. What you actually earn in a month is the rate multiplied by the work you are offered, and that second number is the one nobody advertises.
Who actually gets accepted
This is the part usually left out. The requirements on these listings are specific:
- The actuarial role wants a degree in actuarial science, statistics or mathematics, and names ASA or FSA credentials as a major plus. Pricing, underwriting or reserving experience is expected.
- The function-based code role asks for competitive programming background — ACM/ICPC, Codeforces — plus real depth in algorithms and Python or C++.
- The repo-based code role wants Docker, pytest, Git workflows and comfort in a Linux shell with grep, sed, awk and jq.
- Even the generalist trainer asks for near-native written English and prior annotation or LLM experience.
None of that is unreachable. All of it is checkable, and the screening does check. Applying without it wastes the attempt.
The realistic way in
If you already have the actuarial or competitive-programming background, apply to those directly — the rates justify the effort of a proper application.
For everyone else the generalist trainer role is the door. It is the lowest bar on the list that still pays properly, the experience it asks for is the experience it gives you, and once you have six months of annotation work behind you the specialist roles stop screening you out on the first filter.
What actually gets you through:
- Written English that is genuinely clean. This is assessed, usually with a writing test, and it is the most common rejection.
- Evidence of the domain you are claiming. A degree, a GitHub profile, a portfolio — something checkable.
- Following instructions exactly. Every one of these roles says so in a different way, because guideline adherence is the product.
Before you apply
The practical groundwork is the same as for any remote contract, and we wrote it up separately in how to get a remote job from Kenya — payment rails verified in advance, evidence someone can look at, and a connection that survives a screening call.
That last one matters more here than usual. Annotation work is paid by the hour or the day and assessed on consistency. A week of dropped sessions does not read as a bad line at the other end; it reads as an unreliable contractor, and the work quietly stops arriving.
Where the listings are
All seven roles above are live on InformMe, with the full descriptions, the requirement lists and direct application links. It is our own board, so the usual disclosure applies: we built it, and some listings carry referral links. We have kept the rates and requirements exactly as the hiring platforms publish them, because a job board that flatters the listings is worth nothing to the person reading it.