AI training data
By Gareth HoyleLast updated How we research
AI training data is the text, code, images, audio, video or interaction records used to teach a machine learning model, adjust its behavior or test it.
AI training data is any collection of examples that a machine learning model learns from or is measured against. For large language models it is used for pre-training on large text and code corpora, for fine-tuning on curated sets and for evaluation.
What training data do AI companies pay for?
The training data AI companies pay for is mostly data they cannot collect from the open web. DataVendor, a marketplace run by HUD, says labs pay more for human-written code, complex products and systems languages such as Rust, Go, C and C++. Internal communications with linked context, evaluation datasets and agent trajectories are also in demand.
Public data, scraped data and contact lists rarely sell, because buyers can already get them or cannot clear the rights. The guide to what data AI companies buy has the full list.
Is training on copyrighted data fair use in the US?
Training on copyrighted data is fair use in the US only case by case, under the four factors in 17 USC 107. The US Copyright Office's May 2025 report on generative AI training concluded that some training uses will qualify as fair use and some will not. A license removes that question for the licensed data.
Who owns training data once it is licensed?
Training data stays owned by the seller under most data licenses, and the license decides whether the buyer's rights are exclusive, time-limited or perpetual. A free data valuation gives a realistic range for your records.