In brief
Training data is usually grouped by modality (text, audio, image, video, sensor) and by how it was produced (collected, created by experts, generated synthetically). Each combination has different uses and different rights risks.
On this page
The taxonomy
| Type | Examples | Typical AI uses | Main licensing considerations |
|---|---|---|---|
| Text | Publications, manuals, transcripts, forums you own | Pre-training, fine-tuning, retrieval | Copyright ownership; contributor terms |
| Audio & speech | Recorded speech, podcasts, call audio, music | ASR, TTS, audio understanding | Speaker consent; music rights |
| Images | Photo archives, medical or industrial imaging | Vision models, multimodal training | Releases; sensitive categories |
| Video | Production archives, demonstrations, egocentric footage | Video generation, scene understanding, robotics | Releases; embedded third-party content |
| Expert & reasoning data | Graded answers, rubrics, domain annotations | SFT, preference data, evaluation | Authorship and assignment; confidentiality |
| Industry & structured | Logs, records, sensor tables | Domain models, forecasting | Contractual restrictions; personal data |
| Robotics & sensor | Teleoperation logs, lidar, multi-camera | Embodied AI, manipulation | Capture-site permissions; bystanders |
| Synthetic | Simulation output, generated data | Augmentation, edge cases | Rights in seed data and generators |
Permissible use is the filter
Whatever the type, the gating question is whether the data may lawfully be licensed for AI training. Openly available material is not automatically licensable — check terms, and for Creative Commons material, the specific license conditions.
Own a dataset with clear rights? Start a private inquiry or email hello@callwavemedia.com