Read the licence on the model

Two separate documents

The model licence governs the weights. It says whether you may use the model commercially, whether you may modify or redistribute it, and what uses are forbidden.

The terms of service of a hosted product govern the output. They say who owns what comes out, whether you may use it commercially, whether the provider may use your inputs for training, and what happens on a dispute.

These are independent. A permissively licensed model used through a restrictive service leaves you bound by the service's terms.

What model licences actually say

The range is wider than people expect, and the differences matter.

Fully permissive — Apache 2.0, MIT. Commercial use, modification and redistribution allowed. Several capable open models are here.

Non-commercial only. Some prominent, high-quality open models are released for research and non-commercial use, with commercial use requiring a separate paid licence. Using one for client work without that licence is a breach, and this catches people constantly because the model is freely downloadable and works perfectly.

Community or bespoke licences with conditions: revenue thresholds above which a commercial licence is required, attribution requirements, restrictions on training other models on the output, or clauses that terminate the licence if you bring certain legal claims.

Responsible-AI licences that permit broad use but forbid enumerated purposes — surveillance, discrimination, generating certain content. Their enforceability is untested and they express intent clearly.

Derived weights inherit conditions. A fine-tune of a non-commercial base model is non-commercial. A merge of several models carries every restriction in the set. Community model repositories are full of merges whose licence position is genuinely unclear, and "it was on the internet" is not a defence.

What service terms typically say about output

Most major services now assign whatever rights they may have in outputs to the user, and this is worded carefully — it does not create copyright where none exists, and it does not warrant that the output infringes nothing.

Then read the conditions:

• Commercial use may depend on your subscription tier, with free-tier output restricted.
• Training on your inputs — whether your uploaded images and prompts are used to improve the service. This is often on by default on consumer tiers and off on business ones. For client-confidential material, this is a contractual problem before it is a privacy one.
• Attribution may be required on some tiers.
• Retention and deletion. How long inputs are kept, and whether you can require deletion.
• Content restrictions beyond the law, which can be broader than you expect.

The checks worth running before a job

Five minutes, once per tool:

1. Find the model licence. If it says non-commercial, do not use it for paid work.
2. Find the output terms. Confirm commercial use at your tier.
3. Check whether your inputs are used for training, and turn it off if the material is confidential.
4. Check for attribution requirements and comply with them.
5. Save a copy of both documents with the date. They change, and what governed your use is the version in force at the time.

That last point is not pedantry. Terms are revised, sometimes materially, and a screenshot with a date has resolved more than one dispute.

Where this collides with client contracts

A client contract that warrants you have all necessary rights, combined with a non-commercial model licence, puts you in breach of one or the other. This is a common and entirely avoidable failure.

Equally, a client contract requiring confidentiality is inconsistent with a service that trains on your uploads. Read both sides before signing either.

One more provision worth looking for, because it catches people building on top of these tools: restrictions on using the output to train another model. Several services forbid it outright. If your plan involves generating a dataset and fine-tuning something on it — which is a common and otherwise sensible technique for building a house style — that clause decides whether the plan is available to you. Open models with permissive licences generally allow it; hosted services frequently do not.

The limitation worth stating: none of these documents tells you whether the output infringes somebody else's rights. They allocate rights between you and the provider. Everything about third parties — the melody, the memorised image, the recognisable face — sits outside them entirely, which is what the next lesson is about.