Coding tasks on AI training platforms: what you actually do
How-to · Today
Coding work on AI training platforms means reviewing what a model wrote, building tasks and tests for it, or auditing its agent runs. What each kind of task asks of you, which languages come up most, and what the 86 software listings on this site pay.
What the work is
AI coding training tasks pay programmers to judge and shape code that models write. DataAnnotation's software engineer page describes its version in three parts. You "vibe-code with frontier models" on real tasks such as fixing a bug or extending an API. Then you analyze the result "against production standards: find the bugs, explain the failure, and rate the model's reasoning". And you red-team the model "to expose unsafe code, faked test passes, and confidently wrong solutions".
For the full map of engineering roles and their pay, read AI training work for software engineers. This page is about what a working session looks like.
The four tasks you will meet
The first is reviewing model code. You get a prompt and one or more model solutions, run them, and write up what is wrong. A rationale like "fails on an empty list because line 12 indexes [0] without a check" is the kind of evidence a reviewer can verify.
Writing tasks and tests comes next. You write a problem, a reference solution and a test that accepts correct answers and rejects wrong ones. micro1's Senior Software Engineer (Coding Tasks and Verifiers) lists $30 to $100 an hour for this.
Auditing other people's tasks is a third kind. Mercor's SWE-Bench Task Auditor, at $70 to $90, audits benchmark tasks; its listed skills are Python, code review and benchmark auditing.
Reviewing agent runs is the fourth. Mercor's AI Developer Trace Task Auditor lists code review, debugging and rubric-based evaluation, and its browser test validation reviewer pays $30 to $60.
Run it yourself
Mercor's LLM usage policy bans using language models to predict "code behavior or runtime results", noting that LLMs "are often inaccurate in complex programming scenarios". Its pay policy names "assets that are non-runnable" among common quality problems. So execute the code, paste the actual error into your rationale, and include the input that broke it.
Languages and pay
In October 2026 this site listed 86 jobs in software fields: 41 from micro1, 33 from Mercor, 9 from Invisible Technologies and 3 from Alignerr. Python was the most requested skill (28 listings), then Java (21), code review (19), Rust and TypeScript (18 each). The 80 hourly listings paid $20 to $350, with a median midpoint of $75 an hour. A few pay per task instead, such as Mercor's Go codebase Q&A role at $130 a task.
DataAnnotation's own FAQ gives two figures for coding projects: "starting at $40-$150+ per hour" in one answer and "$50 to $100 per hour" in another. Its role page says no degree is required.
Current openings are under Software & IT.
Questions
- Which programming languages do AI coding training tasks use?
- Python most of all. Among the 86 software listings on this site in October 2026, Python was named in 28, Java in 21, Rust and TypeScript in 18 each. DataAnnotation says most of its coding tasks use Python, C++, Rust, Go or JavaScript.
- Do I need a computer science degree for AI coding tasks?
- Not on DataAnnotation, whose software engineer page says no degree is required and asks for professional or serious open-source experience instead. Requirements on other platforms vary by listing.
- Can I use an AI assistant to check the code?
- Not on Mercor. Its LLM usage policy forbids using language models to predict code behaviour or runtime results, and says submissions evidently produced by AI will be flagged. Check each project's own rules.
Platforms covered here
Put this into practice
Every listing shows its pay and who it is open to.