Reference
AI work glossary
The terms that decide whether you understand a task, a pay band, or a contract. Definitions are written for people applying to AI training and evaluation work, not for model researchers.
44 terms4 groupsLinked to role guides
Work types
The tasks that make up AI training and evaluation jobs.
- AI training
- Human work that improves a model: ranking responses, writing reference answers, grading output, annotating data, or probing for failures. The title covers several job families with different pay bands.Read the guide
- Data annotation
- Adding labels or structure to raw data so a model can learn from it. Common tasks include bounding boxes on images, text classification, audio transcription, and search relevance rating.Read the guide
- AI evaluation
- Judging model output against a rubric and writing the reasoning that supports the score. Evaluation differs from annotation because the written justification is the deliverable.Read the guide
- RLHF
- Reinforcement learning from human feedback. A training method where people rank model outputs, and those rankings teach the model which responses are preferred.
- Red teaming
- Structured probing of a model for unsafe, biased, or policy-breaking behavior. The output is a documented failure and a proposed mitigation, not a simple label.Read the guide
- Prompt engineering
- Designing and refining the instructions given to a model so it produces a better result. It is a skill used across evaluation and red teaming work, not a guaranteed job title.
- AI tutoring
- Writing explanations, worked solutions, and corrective feedback for a model or a learner. The task usually specifies the subject, the level, and the format.Read the guide
- Model evaluation
- Measuring a model against a standard: a benchmark, a rubric, a policy, or a set of expert judgments. Human evaluation is used where automated metrics miss meaning or safety.
- Benchmark
- A fixed test or dataset used to compare model performance. Benchmarks are useful for tracking progress and easy to overfit, which is why human review persists.
- Rubric
- The scoring guide for an evaluation task. It names the dimensions, the severity levels, and the examples that anchor each score.
- Golden response
- An expert-written answer used as the reference standard for judging model output. Golden responses are expensive to produce and valuable as training data.
- Inter-annotator agreement
- A measure of how often two reviewers reach the same judgment. Low agreement usually means the rubric is ambiguous, not that one reviewer is wrong.
Pay and contracts
How AI gig work is priced, paid, and taxed.
- Contractor
- A worker engaged per project or statement of work, not an employee. Contractors usually receive no benefits, paid time off, or guaranteed hours, and they handle their own tax.
- Statement of Work
- The project document that sets scope, rate, payment schedule, and deliverable. On expert marketplaces the final rate often appears in the project invite before the Statement of Work.
- Hourly rate
- Pay per tracked hour. The headline rate is gross, before tax, unpaid onboarding, and idle time between projects.
- Per-task pay
- A fixed amount for a completed task or batch instead of an hourly rate. Per-task pay is not comparable to an hourly band unless you know how long the task actually takes.
- Project cut
- A reduction in rate or scope when a contract is renamed, renewed, or rehired at a lower price. It is the most common pay complaint in AI training work.
- Idle queue
- A period when a platform has approved you but has no tasks available. Idle time is unpaid and is the main reason an hourly rate overstates real earnings.
- 1099
- The US tax form for independent contractor income. Contractors receive a 1099-NEC when payments reach the reporting threshold and owe self-employment tax.
- Effective rate
- What you actually earn per hour after unpaid onboarding, rejected tasks, idle time, and platform fees. It is the number to compare across platforms.
- Platform fee
- A cut taken by the marketplace. Many AI training platforms pay a posted rate directly; others take a margin before the contractor rate. Read the agreement, not the marketing page.
Hiring and screening
The gates between an application and paid work.
- Assessment
- The test a platform uses to qualify a worker. It may be a written task, a coding exercise, a rubric review, a language sample, or a domain exam.
- Track
- A pay and task category set by an assessment, such as generalist, coding, STEM, or expert. The track determines the rate ceiling more than years of experience do.Read the guide
- Credential verification
- Confirming a licence, degree, or professional record through a registry, transcript, or employer reference. Expert platforms verify before they sell your review.
- Identity verification
- Confirming that the person doing the work is the person who applied. It usually involves a government ID and, on some platforms, a live check.
- Trial project
- A short paid or low-paid assignment used to judge quality before a full contract. A trial is not the same as an unpaid assessment, and the rate should be stated.
- Invite-only
- A platform that opens projects by invitation rather than open application. Invite-only access is common on expert tracks and US-only programs.
Models and data
The technical terms that show up in task guidelines.
- Large language model
- A model trained on large text datasets to predict and generate language. GPT, Claude, and Gemini are examples. The term describes the architecture family, not a specific product.
- Token
- A unit of text a model processes. A token can be a word, part of a word, or a character depending on the tokenizer. Token counts set context limits and API costs.
- Context window
- The maximum amount of text, measured in tokens, that a model can consider at once. A longer context window lets a model handle longer documents and conversations.
- Hallucination
- A plausible but incorrect or fabricated statement from a model. Hallucination is a common evaluation target because it is hard to catch without domain knowledge.
- Grounding
- Connecting a model answer to a verifiable source. Grounded answers cite evidence; ungrounded answers rely on the model’s parameters.
- Fine-tuning
- Further training a pretrained model on a smaller, task-specific dataset. Fine-tuning is one route to specialization; human feedback is another.
- Supervised fine-tuning
- Fine-tuning on labeled examples of the desired input and output. It is often paired with preference data and reinforcement learning.
- Inference
- Running a trained model to produce an output. Inference is different from training, which is how the model learned its parameters.
- Embedding
- A numeric vector that represents the meaning of text, an image, or another input. Embeddings power similarity search and resume-to-role matching.Read the guide
- Temperature
- A sampling setting that controls output randomness. Lower values produce more predictable responses; higher values produce more varied ones.
- Chain-of-thought
- A prompting pattern that asks a model to show intermediate reasoning. It can improve accuracy on multi-step problems and makes errors easier to inspect.
- Alignment
- The work of making a model behave in line with human intent and safety expectations. Evaluation, red teaming, and preference training all contribute.
- Bias
- Systematic error that produces unfair or skewed output across groups. Bias can enter through training data, labels, or model design, and it is a common expert review target.
- Adversarial testing
- Deliberately trying to make a model fail through crafted inputs. It is the method behind red teaming and safety evaluation.
- Multimodal
- A model or task that handles more than one input type, such as text and images, or audio and video. Multimodal annotation is a growing task category.
- Synthetic data
- Data generated by a model rather than collected from the world. Synthetic data can scale training but needs review for errors and duplication.
- Overfitting
- When a model memorizes training examples and performs worse on new inputs. Evaluation sets are kept separate to detect it.
Put the vocabulary to work
Read the role guides for pay bands and screening, or upload a resume and see which live roles match your track.