Get paid to train AI: RLHF, evaluation and coding tasks from the UK
AI labs pay people to rate chatbot answers, write ideal responses, catch reasoning errors and review code, and the platforms that run this work recruit in the UK. Pick the real platforms, pass the one-shot assessments, track your true hourly rate with unpaid time included, and move into the coding, maths, language or professional projects that pay several times the generalist rate.
- A scam checklist that spots the fake "AI training job" messages before they cost you anything
- A specialisation plan, profile lines and CV bullets for the higher-paid expert projects
- A laptop and a reliable internet connection
- Excellent written English (or native-level fluency in another language), and patience with long guidelines
- Python 3.9+ for the work log script. New to the terminal? Terminal survival kit
- The right to work in the UK, and ID for platform verification
What AI training work actually is
Know the tasks, why labs pay for them, and what it takes to earn well.
Every chatbot you use was shaped by people judging its answersLabs need humans to compare two responses and say which is better and why, to grade answers against rubrics, to write the ideal answer to hard prompts, to try to make models fail, and to check code, maths and expert content. This is often called RLHF (reinforcement learning from human feedback) and evaluation work. The better the human judgement, the better the model, so the pay rises with the skill.
The common task types:
- Preference ranking: two answers to one prompt; pick the better, score both, justify in a few sentences.
- Rubric grading: score an answer against a checklist of criteria.
- Ideal responses: write the answer a model should have given.
- Red teaming: write prompts that make the model go wrong, within the project's rules.
- Coding: review, test and fix model-written code, or write tasks and unit tests.
- Expert tasks: maths, science, law, finance, medicine: write and check problems at professional level.
This is labour, not a business you own: you're paid per hour or per task as an independent contractor. If you'd rather build something you can sell repeatedly, see Build and sell licensed datasets.
Careful- What it takes. One-shot assessments: many platforms give you a single attempt, unpaid.
- Long guidelines, read carefully and re-read when they change.
- Uneven work: projects open and pause without much warning, so income goes up and down.
- Accuracy over speed: quality scores decide whether you keep getting work.
- Who this suits: you're precise, you write clearly, you enjoy spotting what's wrong with an answer, and you want flexible hours. A degree, a profession, coding skill or a second language moves you into the projects that pay most.
The kitread the kit overview or download everything: the platform guide and scam checklist, practice tasks, the work log script and the application kit.