Software, data & AI training work: field summary

What does the work involve?

The tasks draw on your engineering judgement, not simple labelling. Four kinds come up again and again:

• Writing tasks that models fail. You build a realistic problem with requirements, failure modes and a reference answer. Terminal tasks bundle an instruction, a Docker image, tests, an example solution and a time limit. Outlier says these "typically take 6 to 7 hours to complete".
• Grading AI-written code. You check edge cases and whether the tests really test what they claim. You also flag code that passes review but would fail under concurrency, at scale or in error handling.
• Working with an AI agent. You run agentic coding sessions or review AI-generated pull requests. Terac states: "You never write production code for a client."
• Timed human baselines. Researchers measure AI agents against people doing the same task, in the same environment and with no AI assistants.

ML engineers may review PyTorch and JAX training scripts or judge advice on serving models with tools such as vLLM and Triton. Security specialists appear in Mercor's software pool and in Terac's top band for "security-critical review".

What pay do programmes post?

Pay depends on the programme, the project and often the country. This is exactly what each programme's page showed on 29 October 2026:

Entry requirements vary just as much. Outlier's general coding page asks for a degree from a "top 200 global" institution. DataAnnotation says no degree is required. Mercor says owning code in production matters more than where you trained. Turing asks for 3+ years as a developer. Handshake AI and Mercor do not support STEM OPT.

What did our roles board show for this field?

On 29 October 2026, we listed 35 open roles in this group, from three programmes' public pages. Software engineering accounted for 23 of them, cybersecurity 5, machine learning 4, data analysis 2 and IT management 1.

• 25 roles showed an hourly range. The lowest was $31–$60 an hour and the highest $120–$180.
• Seven were full-time staff posts with a yearly salary. Two were paid per task.
• Eleven stated a minimum, usually 10 to 15 hours a week. One asked for 40.
• Five limited where you must live.

Across the whole site there are 412 open roles today, checked on 6 Nov 2026. Each role card shows its own posted range and the date we checked it. Read how we vet listings and how we earn from programmes. We list programmes whether or not we earn from them.

How does vetting work?

The usual path is a profile, a skills assessment, often an AI-led interview, then identity checks. Some steps allow only one attempt:

• DataAnnotation's starter assessment: "There are no retakes or second chances."
• Mercor's AI interview takes about 20 minutes with camera and microphone. Mercor allows up to three retakes in total, and "hiring managers can see your retake count".
• From 17 July 2026, Mercor refuses sign-ins through a VPN, a proxy or iCloud Private Relay.

Passing does not mean work starts. Mercor says a match can come "within a week or several months". Mindrift says the path to a first paid task typically takes 1–2 weeks.

Can you use AI tools during tasks?

That depends on the task, and the rule cuts both ways. Mercor forbids using AI to assess model answers, to write your justification or to predict code behaviour. It says such submissions "will be identified and flagged". Agentic tasks, by contrast, require you to drive the agent they provide. Read each task's own rule before you start.

Which pay terms should you check?

• Per task vs per hour. Some programmes pay per task even when the rate is shown per hour. At Mercor, a deliverable contract has "no minimum hourly guarantee".
• Paused projects. At Mercor, time logged during a pause is not paid.
• Onboarding. Outlier says "lower rates apply" during initial project onboarding.
• Monitoring. Mercor's hourly work runs under time-tracking software that takes screenshots. Its tracker needs Windows 10 or later, or macOS 11 or later.
• Tax. Mercor "does not deduct taxes from your payments". For official rules, see your tax authority. The general guide points to them.

What does a per-task offer work out to?

This is an illustration with our own numbers, not a sourced figure. Take a task posted at up to $1,800. If it takes 7 hours and is accepted first time, that is about $257 an hour before tax. If review sends it back twice and it takes 14 hours, it is about $129 an hour. If it is rejected, the time may be unpaid.

What scams target developers?

• Malware "coding tests". A fake recruiter sends a private repository to build. Building it installs software that steals data. Open a stranger's code only in a throwaway virtual machine.
• Task scams. These pay small amounts at first, then ask you to deposit money. The FTC's rule: "Never pay anyone to get paid."
• Renting accounts. Mercor allows one account per person, in your own name.

Programmes do ask for your ID and payout details, so that request alone is not the warning sign. The warning signs are the channel and money: an unsolicited message, a domain that is not the programme's own, or any request to pay.

Before you apply

• Find the role on the programme's own site.
• Check the stated years, degree, country and hours.
• Know whether it pays per hour, per task or per deliverable.
• Read the rule on AI tools for that task.
• Read your employment contract's outside-work and IP clauses.

Our full field guide gives the source for each fact. Our sources showed no programme publishing guidance on employer conflicts beyond Terac's: decline a conflicting task "without penalty".