What we look for in Natton Crowd contributors

High-quality AI data depends on people who can make careful, consistent judgments. The work may involve labelling an image, comparing two model responses, transcribing speech, reviewing a safety case, or applying specialist knowledge to a complex example.

Natton Crowd is designed around that range of work. We look for contributors whose language, domain expertise, and working habits match the task—not for one generic profile that is expected to do everything.

Language ability must be demonstrated

For language work, native or near-native fluency is assessed through task-specific examples. Reviewers look beyond grammar to regional usage, natural phrasing, script familiarity, code-switching, and the ability to explain why a response is or is not appropriate.

A contributor may be excellent in one dialect or register without being the right match for another. We record those strengths so work can be routed accurately rather than treating a language as a single uniform category.

Domain expertise is verified separately

Some projects require professional judgment. Medical, legal, financial, scientific, and technical tasks may call for education, credentials, or demonstrated work experience. Where credentials are relevant, they are checked against the issuing body or supporting documentation.

Subject expertise does not replace annotation skill. Experts still need to learn the project rubric, apply it consistently, and document uncertainty.

Calibration matters more than speed

Qualified contributors complete a calibration set before working on production batches. The purpose is to align on the rubric, reveal ambiguous guidance, and establish a quality baseline.

We value reviewers who read the instructions closely, use evidence, flag unclear cases, and explain difficult decisions. Fast completion is useful only when the judgment remains accurate. Unusually fast work is reviewed as a possible quality signal rather than automatically rewarded.

Quality is measured continuously

Onboarding is not a permanent certification. Production workflows can include gold-standard tasks, blinded overlap, reviewer feedback, escalation, and periodic recalibration. Quality is examined by task type and difficulty so contributors receive feedback that is relevant to their work.

When agreement is low, the contributor is not always the problem. The rubric or source material may be ambiguous. Strong contributors help identify those cases instead of guessing.

Reliability and communication count

Data programs run on deadlines and carefully managed batches. Contributors need to communicate availability, protect confidential material, follow security requirements, and raise blockers early. Consistency across many small decisions is often more valuable than a single perfect task.

Fair work supports better data

Clear expectations, appropriate compensation, respectful feedback, and realistic workloads make it possible for skilled contributors to stay engaged. Retention preserves project knowledge and reduces the repeated calibration cost that comes with constant turnover.

How to prepare

Applicants should describe the languages, dialects, domains, and task types they genuinely know. During qualification, read the rubric before answering, explain difficult choices, and flag ambiguity rather than inventing certainty.

The best contributors combine expertise with disciplined judgment. They understand that their work is not simply producing labels—it is creating the evidence used to train and evaluate systems that people will rely on.