Reading the licence before you build on it
Three separate licences, routinely confused
When you use an AI system there are at least three distinct legal instruments in play, and people conflate them constantly.
The model licence — what you may do with the weights. Applies when you download a model.
The service terms — what you may do with the output of a hosted service, and what the provider may do with your input.
The data licence — what the training data permitted, which is mostly the previous lesson's problem and occasionally yours, if you fine-tune on a dataset.
A model can be free to download and restricted in use. A service can grant you all rights in the output and disclaim all responsibility for it. These are separate questions with separate answers.
Model licences, in the categories that matter
Genuinely permissive. Apache 2.0 and MIT. Use commercially, modify, redistribute, no user limits, no field restrictions. Several strong models ship under these, including much of the Qwen family and some Mistral releases. If you need certainty, this is where it lives.
Community licences with conditions. Meta's Llama licences are the prominent example: free for almost everyone, with an acceptable use policy, attribution and naming requirements, and a clause requiring a separate licence if your product exceeds 700 million monthly active users. Google's Gemma terms have their own use restrictions. These are usable and they are not open source in the traditional sense, because the freedoms are conditional.
Responsible AI licences (OpenRAIL and similar). Permissive except for enumerated prohibited uses — surveillance, discrimination, generating disinformation. Whether use restrictions are enforceable and how they interact with downstream distribution is not fully tested.
Non-commercial and research-only. Common for research releases. "Research use" is often undefined, and internal evaluation at a company is a grey area many organisations quietly ignore.
What "open" means is contested. The Open Source Initiative published an Open Source AI Definition in 2024 requiring, among other things, sufficient information about the training data for someone to recreate the system. By that standard most models described as open are not, since the weights are published and the data is not. The term you want is usually open weights.
Output terms
For hosted services, look for three clauses.
Assignment of output. Most major providers now assign to the user whatever rights they may have in the output. Note the hedge — they cannot give you copyright that does not exist, and as the earlier lesson explained, purely generated material may have no copyright owner at all. What the clause really does is stop the provider claiming it.
Restrictions on use. Prohibited categories, competitive restrictions on using output to train a rival model, and requirements to disclose AI involvement in some contexts.
Indemnification. Several providers now indemnify business customers against third-party copyright claims arising from output, usually conditional on using their safety features and not deliberately prompting for infringement. Read the conditions, because that is where the coverage actually lives.
Questions to ask before you build a business on it
Six, and they take an hour.
- Which licence exactly, at which version? Licences change between releases of the same model family.
- Are there user or revenue thresholds that trigger different terms?
- Are there field-of-use restrictions that conflict with your customers?
- Can you fine-tune, and do the outputs of a fine-tune inherit the licence?
- Do you have to attribute, and where?
- What happens if the provider changes terms or withdraws the model? Model deprecation with short notice is common, and your product depends on something you do not control.
Where the free path is
If licence certainty matters — for a government contract, an academic release, a product you cannot revisit — build on an Apache 2.0 or MIT model rather than a conditional community licence. The capability gap between the most permissive open-weights models and the conditional ones has narrowed considerably, and the legal simplicity is worth more than a few points on a benchmark for most organisations.
And keep a record of which model, which version, which licence, on which date, for anything you ship. Reconstructing that two years later, after three model deprecations, is the sort of task that consumes a week and produces an approximation.
The one thing to keep
The model licence, the service terms and the data licence are three separate instruments — and most models called open are open weights rather than open source, with conditional licences that carry user thresholds and use restrictions.
Before you move on
A startup plans to embed a model licensed under a community licence with a 700 million monthly-active-user threshold, and treats it as equivalent to Apache 2.0. What is the real risk profile?
Pick the one you would defend. Nobody sees your answer.