Grounding license vs training license: what each one permits

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

The two rights cover different acts on the same content

The core difference is what happens to the content: training changes what the model knows, while grounding changes what the model shows and credits in a given answer [2]. Industry deal trackers describe these as the two main shapes of publisher AI deals, and many recent agreements bundle both under one contract [3]. That bundling is where most scope errors start, because a single fee line can hide two very different permissions.

In a training license, the licensed acts are reproduction into a training corpus, preprocessing (tokenization, deduplication, filtering) and the use of that corpus in pre-training, continued pre-training or fine-tuning. The output is a set of weights that do not store the corpus as retrievable documents (though memorization of some passages remains possible), so the deal usually turns on a one-time or per-volume fee and on what happens to derivative models. For the rights grant itself, see rights for foundation-model pre-training.

In a grounding license, the licensed acts are crawling or API access, chunking, embedding into a vector index, retrieval at query time, and display of excerpts or summaries to end users. Because content is fetched every time a user asks, grounding deals tend to be ongoing and metered rather than paid once [1]. The broader context sits in our retrieval, RAG and grounding data hub and the retrieval-augmented generation glossary entry.

Side-by-side comparison of grounding and training grants

Most of the negotiable terms differ in kind, not just in degree, between the two license types. The table below lists the terms counsel typically checks and how each license usually treats them.

Vendors already sell these as separate tiers: some price training, derivative-model and RAG or vector-index use as distinct license levels [6]. If a content owner offers one price for "AI use," ask which column of this table the price actually covers.

A training license rarely covers RAG, and the reverse is worse

A training data license does not cover RAG unless it expressly grants retrieval, indexing and display; nothing in a corpus-and-weights grant implies the right to show licensed text to end users [1]. The reverse is riskier: if your product logs retrieved passages and those logs flow into a fine-tuning set, you may be training on grounding content without a training right. Commentators note that public deal reporting often leaves exactly this point unclear [4], so the fix is a written training exclusion or an explicit training grant.

Pay attention to the paths through which grounded content can reach training jobs without anyone deciding it should:

• Conversation logs that store retrieved chunks and are later sampled for supervised fine-tuning or preference data.
• Retrieval-augmented fine-tuning recipes (for example, training on query, retrieved context and answer triples), which by design train on retrieved text.
• Synthetic question-answer generation that uses the licensed corpus as seed context.
• Embedding or reranker training on the licensed documents, which may count as model training under the contract even if the generator is never touched. See training data for domain embedding models.
• Evaluation sets built from licensed passages and later reused as training data.

Define "grounding" in the contract, because the market does not

Deal language uses grounding, RAG, real-time access and content inference as loose near-synonyms, so the contract must define each term it relies on [5]. Industry glossaries list RAG alongside fine-tuning and model development as separate AI activities that a license may or may not permit [10]. A definition that only says "AI use" or "machine learning purposes" leaves both parties guessing about embeddings, caches and summaries.

Write definitions around observable acts rather than labels. "Grounding Use" can be the retrieval of Licensed Content at inference time to inform a response shown to an end user, including storage of embeddings and a cache limited to the retention period. "Training Use" can be any use of Licensed Content, or of outputs that reproduce it, to create or modify model parameters, including adapters, embedding models and rerankers. Tie both to a field of use that names the products and audiences, and cross-check against our general guide to AI data license terms.

Match the license to the product you are shipping

The product surface, not the model architecture, decides which right you need. Use this decision table as a first pass with product and legal in the same room.

Illustrative example: invented to show structure; it does not describe an available dataset.

For the internal-versus-public split in more depth, see licensing content for internal RAG vs customer-facing answers. For index-level rights, see embedding and vector index rights.

Legal background differs for training and for retrieval

Statutory exceptions written with training in mind may not reach retrieval, which is one reason a contract is still needed for grounding. Commentary on Japan's draft guidance on Article 30-4 places the collection of material for RAG outside the exception where the retrieved works are meant to be reproduced in outputs for people to read [7]. In the US, the Copyright Office's Part 3 report, still a pre-publication version as of October 2026, concludes that many acts involved in AI training implicate copyright owners' rights [8].

For general-purpose model providers in the EU, Article 53(1)(c) of the AI Act requires a copyright compliance policy that honors machine-readable rights reservations under Article 4(3) of Directive (EU) 2019/790 [9]. A training license with clear provenance records makes that policy easier to document. Grounding content still needs its own scope check, because a reservation against training does not answer whether display is allowed.

Clause checklist for a bundled training and grounding deal

A bundled deal works if each right has its own grant, fee basis, reporting duty and end-of-term rule. Use this checklist when one agreement covers both.

Illustrative example: invented to show structure; it does not describe an available dataset.

1. Separate grants. One clause for Training Use, one for Grounding Use, each with its own defined acts and field of use.
2. Explicit training exclusion for grounded content. "Content retrieved under the Grounding Use grant, including logs and caches containing it, shall not be used for Training Use except as expressly licensed in Section X."
3. Fee allocation. State which fee pays for which right, so a dispute over one does not reopen the other [3].
4. Display limits. Maximum excerpt length, summary rules, and required attribution or link-back format.
5. Derivative assets. Who may keep weights, adapters, embeddings and indexes, and for how long.
6. Termination split. Index and cache purge on grounding termination; a separate rule for trained models.
7. Usage reporting. Retrieval and display counts for grounding; corpus manifest and model list for training.
8. Third-party processing. Whether licensed content may be sent to external model APIs at inference time.

Related clause-level guidance is in RAG content license terms and usage-based pricing for RAG content.

Where operational company data fits

Not every grounding or training need is published editorial content; many teams need operational records such as support histories, engineering records or documents. SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, uses, term and delivery. You can describe the content and uses you need, stating separately whether you intend training, grounding or both.

License the right use for operational data

If you need operational data from US businesses for training, grounding or both, SourceX finds companies that hold the described data, assesses data and licensing permissions, and agrees pricing and allowed uses in a license; nothing is contracted until a supplier agrees. Start by describing your data and intended uses to SourceX.