Data Marketplaces & Intermediaries

Getty licenses its library into ChatGPT — but explicitly not for training, and that line is the whole story

Source: Industry reporting · Jun 21, 2026

A major stock-visual library agreed to let a leading AI assistant surface its licensed images inside search results — displayed and credited to the source, with the deal pointedly structured so the images are not used to train the model. The distinction is not a footnote; it is the entire design. One category of deal licenses data to be learned from; this one licenses data to be shown, and keeps the training rights off the table entirely.

The strategic read is that 'licensed data' is fracturing into clearly different products with clearly different terms. Display licensing, retrieval licensing, and training licensing are three distinct things a rights-holder can sell, and the smart ones are learning to sell them separately rather than signing away everything in a single blanket grant. The library here is monetizing visibility inside an assistant while protecting the thing that makes its catalog valuable — the right to train on it — for a separate, presumably pricier, conversation.

That precision is the point worth taking away. The value of a data deal is not captured by the word 'licensed'; it is captured by exactly what use was granted, to whom, for how long. A provider who can slice those rights cleanly — display here, retrieval there, training only under specific terms — extracts more value and carries less risk than one who hands over an undifferentiated dump. The granularity is the leverage.

The through-line with the rest of this market is consistent: provenance and clearly specified rights are what turn data into a durable, defensible asset. A deal that says 'shown, credited, not trained on' is only possible because both sides can define and enforce that boundary. The ability to draw the line — and prove it holds — is exactly what separates data you can license with confidence from data you can only hope nobody challenges.

Key Points

• A major stock-visual library struck a multi-year deal to surface its licensed images inside a leading AI assistant's search — shown and credited, not used to train
• The 'display, don't train' structure is becoming its own licensing category, distinct from training-data deals
• Explicitly carving out training rights is a provenance move: it keeps the licensor's core asset off the training table
• It underlines that 'licensed' is not one thing — the terms of use are the product