AI Training Data Glossary
Plain-language definitions for the licensing, rights, and provenance terms used across AI training data.
C
Chain of rights
The documented trail showing who owns a piece of content and that each link, from creator to licensor to buyer, had the authority to grant the rights it granted. In AI training data, a complete chain of rights per asset is what separates rights-cleared data from data that merely has a license attached.
Read definition →Computer-use agent
An AI agent that operates software the way a person does: viewing the screen, clicking, typing, and navigating real interfaces to complete tasks. Training one takes two inputs, demonstrations of humans actually using software, and RL environments that replicate applications so the agent can practice with verifiable feedback. Computer-use is currently one of the fastest-growing categories of both environment building and training data demand.
Read definition →Content provenance (C2PA)
An industry standard for attaching tamper-evident origin records to media files, documenting where content came from and how it's been altered. In training data, provenance standards support the per-asset documentation that rights-cleared licensing requires.
Read definition →D
Dark data
Dark data is information an organization has collected or generated but is not actively using for its primary business purpose: old email archives, logs, closed tickets, and records kept for compliance. It used to be a storage cost. AI changed that, because some proprietary dark data contains real operational knowledge that exists nowhere on the public web. Not all of it is valuable; worth depends on uniqueness, connected context, visible outcomes, and whether the rights are clean enough to license.
Read definition →Data annotation
The work of labeling data so models can learn from it: tagging objects, transcribing speech, ranking outputs. Annotation companies process data you already have; they are a different category from licensing marketplaces, which supply the data itself.
Read definition →Data curation
The selection, cleaning, deduplication, and structuring of raw content into a training-ready dataset. Curation is where most of the practical value in a dataset gets added, and it's the difference between buying content and buying training data.
Read definition →Data exclusivity
A licensing term under which a rights holder agrees not to license the same content to other buyers. Exclusive data is the only data that differentiates a model, since content every lab can buy moves no benchmarks. On Troveo, 95 percent of licensors have signed exclusively.
Read definition →Data provenance
The documented origin and history of a piece of training data: who created it, who has held the rights to it, and how it was acquired. Courts in the AI copyright cases have treated acquisition as a separate legal question from training, which has made provenance a requirement rather than a preference for production models.
Read definition →E
Egocentric video
First-person footage shot from the point of view of the person performing a task, hands in frame. It is disproportionately valuable for robotics and physical AI because it approximates what a robot's own cameras will see while acting.
Read definition →Expert demonstrations
Training examples produced by domain specialists, such as a doctor writing an ideal diagnostic response or an engineer completing a real coding task, used to teach a model what excellent performance looks like. Expert demonstrations sit between raw training data and human feedback: they are data, but produced to order by vetted people. Vendors differentiate on how their experts are sourced, verified, and matched to tasks.
Read definition →F
Fair use
The legal doctrine allowing limited use of copyrighted material without permission, weighed on factors including the purpose of the use and its effect on the market for the original. In AI training, courts have split on when training qualifies, and the growth of the data licensing market itself weakens the defense over time, because a market that exists can be harmed.
Read definition →Fine-tuning data
Smaller, targeted datasets used to adapt a pretrained model to a specific task, domain, or behavior. Because fine-tuning data steers the model directly, quality and specificity matter more per example than in pretraining, which makes it a market for curated, licensed datasets rather than bulk collection.
Read definition →I
L
M
Model weights
The numerical parameters a model learns during training, the artifact that training data ultimately becomes. Weights matter in licensing because content absorbed into them cannot be simply deleted later, which is why training rights, unlike display rights, are effectively permanent grants.
Read definition →Multi-camera video
Footage of the same scene captured simultaneously from multiple angles. It teaches models spatial consistency, how a scene holds together across viewpoints, which makes it one of the most requested data types for world models and 3D-aware video generation.
Read definition →R
Reward hacking
When an agent trained by reinforcement learning finds a way to score well on the reward signal without actually doing the task, such as exploiting a loophole in how success is measured. Reward hacking is the central quality problem in the RL environment market: a realistic simulation with a gameable verifier teaches a model to cheat convincingly. Buyers evaluating environment vendors should ask how verifiers are tested against it.
Read definition →RL environment
A controlled, simulated task space where an AI agent practices completing goals and gets scored on the result. A production RL environment has three parts: a replica of real software or a real task domain, goal specifications usually written by domain experts, and a verifier that decides whether the agent succeeded. Environments train how a model acts; they do not replace the real-world data that teaches a model what the world is.
Read definition →RLHF data
Human feedback data used in reinforcement learning from human feedback: rankings, preferences, and evaluations that teach a model which outputs people consider better. It is produced by human workforces rather than licensed from rights holders, which makes it a different market from training data supply.
Read definition →S
Sim-to-real gap
The performance drop that occurs when a model trained in simulation is deployed in the physical world, caused by everything simulators fail to reproduce. It is the main reason physical AI teams layer real-world footage on top of synthetic environments.
Read definition →Synthetic training data
Training data generated by models rather than collected from the world. It fills gaps and covers edge cases cheaply, but inherits the limitations of the models that generate it, so labs treat it as a layer on top of real-world data rather than a replacement.
Read definition →System of record
A system of record is an authoritative business system where important operational information is stored, such as a CRM, ERP, support platform, HR system, or internal database. When sources disagree, it is the one the organization treats as true. For AI training data, systems of record hold the outcomes that turn activity into learnable examples, and they anchor cross-system context. They also demand the most care on rights and privacy, since they often contain customer and employee information that must be excluded.
Read definition →T
Teleoperation
Humans remotely controlling robots to perform tasks, recorded to produce robot-native training data. It yields the highest-fidelity demonstrations and costs the most per hour, which is why teams reserve it for the last mile and layer licensed real-world footage underneath.
Read definition →Training-ready data
Data delivered in the condition a training pipeline actually needs: cleaned, normalized, consistently formatted, with metadata and per-asset rights documentation. The difference between raw content and training-ready data is weeks of engineering and legal review, which is why buyers increasingly treat delivery condition as part of the price.
Read definition →V
Verifier
The component of an RL environment that decides whether an agent's attempt at a task succeeded, producing the reward signal the model learns from. Verifiers range from automated checks, like passing tests in a codebase, to expert-written rubrics scored by humans. Verifier quality is what separates environment vendors, because a weak verifier gets gamed and a mislabeled reward trains the wrong behavior at scale.
Read definition →Voice likeness rights
Legal protections covering a person's voice as an aspect of their identity, now explicit in several US states in response to voice cloning. In audio training data, voice rights sit alongside copyright, which is why properly cleared speech data requires speaker consent, not just a content license.
Read definition →W
Workflow trajectory
A workflow trajectory is a structured sequence of actions or state changes showing how a task progressed from an initial state toward an outcome: the request, the systems used, the actions and handoffs, and the result. Agentic AI models learn from complete sequences like these, not isolated records. A raw data export is not automatically a trajectory; turning records into clean, ordered, connected sequences takes structuring and annotation, which is where much of the value in enterprise training data is created.
Read definition →World model
An AI system that builds an internal simulation of an environment and can run it forward in time, predicting what happens next given a scene and an action. World models learn from watching worlds operate, which is why they train on video and gameplay rather than text.
Read definition →