{"course":{"slug":"ai-and-you","title":"AI, Safety and What Goes Wrong","url":"https://addaly.com/learn/ai-and-you","level":"beginner","track":"start"},"lesson":{"slug":"licences-you-actually-have","title":"Reading the licence before you build on it","url":"https://addaly.com/learn/ai-and-you/licences-you-actually-have","position":61,"of":73,"minutes":8,"answer":"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.","keyIdea":"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."},"body":{"format":"markdown","text":"## Three separate licences, routinely confused\n\nWhen you use an AI system there are at least three distinct legal instruments in play, and people conflate them constantly.\n\n**The model licence**: what you may do with the weights. Applies when you download a model.\n\n**The service terms**: what you may do with the output of a hosted service, and what the provider may do with your input.\n\n**The data licence**: what the training data permitted, which is mostly the problem of [The training data question, unresolved](/learn/ai-and-you/the-training-data-dispute) and occasionally yours, if you fine-tune on a dataset.\n\nA 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.\n\n> Figure: Three instruments, routinely read as one. Layers, top to bottom: The service terms (What you may do with a hosted service's output, and what the provider may do with your input); The model licence (What you may do with the weights: Apache 2.0 and MIT are permissive; community licences carry user thresholds and use rules); The data licence (What the training data permitted; the previous lesson's problem, and yours if you fine-tune). 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. Each layer answers a different question.\n\n## Model licences, in the categories that matter\n\n**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.\n\n**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.\n\n**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.\n\n**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 ignore.\n\n**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**.\n\n## Output terms\n\nFor hosted services, look for three clauses.\n\n**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 does is stop the provider claiming it.\n\n**Restrictions on use.** Prohibited categories, competitive restrictions on using output to train a rival model, and requirements to disclose AI involvement in some contexts.\n\n**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.\n\n## Questions to ask before you build a business on it\n\n1. Which licence exactly, at which version? Licences change between releases of the same model family.\n2. Are there user or revenue thresholds that trigger different terms?\n3. Are there field-of-use restrictions that conflict with your customers?\n4. Can you fine-tune, and do the outputs of a fine-tune inherit the licence?\n5. Do you have to attribute, and where?\n6. 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.\n\n## Where the free path is\n\nIf 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.\n\nAnd 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."},"check":{"question":"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?","options":[{"text":"Identical in practice, since the startup will never come near 700 million users","correct":false,"tempting":"Fixes on the one threshold unlikely to bind, and ignores the use, naming and attribution conditions that apply from day one."},{"text":"Higher, since use rules, naming and attribution apply at once, and can change by version","correct":true},{"text":"Lower, since the licensor wants commercial adopters and will not enforce on a startup","correct":false,"tempting":"Swaps a prediction about enforcement for the terms actually agreed, which a customer or acquirer will read."},{"text":"Equivalent, because both allow commercial use, and that is the only term that matters","correct":false,"tempting":"Treats commercial permission as the whole licence, when the conditions attached to it are exactly what differs."}],"why":"A community licence is a contract with conditions Apache 2.0 does not carry, and its terms can change between versions."},"prev":{"slug":"the-training-data-dispute","title":"The training data question, unresolved","url":"https://addaly.com/learn/ai-and-you/the-training-data-dispute"},"next":{"slug":"disclosure-rules-that-bind-you","title":"When you have to say it was AI","url":"https://addaly.com/learn/ai-and-you/disclosure-rules-that-bind-you"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Source and assessment review continues. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Reading the licence before you build on it', in AI, Safety and What Goes Wrong. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-and-you/licences-you-actually-have (Accessed: 21 November 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Reading the licence before you build on it","courseTitle":"AI, Safety and What Goes Wrong","module":{"index":8,"slug":"law-and-ownership","title":"Ownership, law and the public record"},"answer":null,"keyIdea":"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.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"caption":"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. Each layer answers a different question.","layers":[{"label":"The service terms","detail":"What you may do with a hosted service’s output, and what the provider may do with your input"},{"label":"The model licence","detail":"What you may do with the weights: Apache 2.0 and MIT are permissive; community licences carry user thresholds and use rules"},{"label":"The data licence","detail":"What the training data permitted; the previous lesson’s problem, and yours if you fine-tune"}],"title":"Three instruments, routinely read as one","type":"stack"},"alt":"Figure: Three instruments, routinely read as one. Layers, top to bottom: The service terms (What you may do with a hosted service’s output, and what the provider may do with your input); The model licence (What you may do with the weights: Apache 2.0 and MIT are permissive; community licences carry user thresholds and use rules); The data licence (What the training data permitted; the previous lesson’s problem, and yours if you fine-tune). 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. Each layer answers a different question."}},"blocks":[{"t":"h2","id":"three-separate-licences-routinely-confused","runs":[{"text":"Three separate licences, routinely confused"}]},{"t":"p","runs":[{"text":"When you use an AI system there are at least three distinct legal instruments in play, and people conflate them constantly."}]},{"t":"p","runs":[{"text":"The model licence","b":true},{"text":": what you may do with the weights. Applies when you download a model."}]},{"t":"p","runs":[{"text":"The service terms","b":true},{"text":": what you may do with the output of a hosted service, and what the provider may do with your input."}]},{"t":"p","runs":[{"text":"The data licence","b":true},{"text":": what the training data permitted, which is mostly the problem of "},{"text":"The training data question, unresolved","href":"https://addaly.com/learn/ai-and-you/the-training-data-dispute"},{"text":" and occasionally yours, if you fine-tune on a dataset."}]},{"t":"p","runs":[{"text":"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."}]},{"t":"h2","id":"model-licences-in-the-categories-that-matter","runs":[{"text":"Model licences, in the categories that matter"}]},{"t":"p","runs":[{"text":"Genuinely permissive.","b":true},{"text":" 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."}]},{"t":"p","runs":[{"text":"Community licences with conditions.","b":true},{"text":" 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."}]},{"t":"p","runs":[{"text":"Responsible AI licences (OpenRAIL and similar).","b":true},{"text":" 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."}]},{"t":"p","runs":[{"text":"Non-commercial and research-only.","b":true},{"text":" Common for research releases. “Research use” is often undefined, and internal evaluation at a company is a grey area many organisations ignore."}]},{"t":"p","runs":[{"text":"What “open” means is contested.","b":true},{"text":" 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 "},{"text":"open weights","b":true},{"text":"."}]},{"t":"h2","id":"output-terms","runs":[{"text":"Output terms"}]},{"t":"p","runs":[{"text":"For hosted services, look for three clauses."}]},{"t":"p","runs":[{"text":"Assignment of output.","b":true},{"text":" 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 does is stop the provider claiming it."}]},{"t":"p","runs":[{"text":"Restrictions on use.","b":true},{"text":" Prohibited categories, competitive restrictions on using output to train a rival model, and requirements to disclose AI involvement in some contexts."}]},{"t":"p","runs":[{"text":"Indemnification.","b":true},{"text":" 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."}]},{"t":"h2","id":"questions-to-ask-before-you-build-a-business-on-it","runs":[{"text":"Questions to ask before you build a business on it"}]},{"t":"ol","start":1,"items":[[{"text":"Which licence exactly, at which version? Licences change between releases of the same model family."}],[{"text":"Are there user or revenue thresholds that trigger different terms?"}],[{"text":"Are there field-of-use restrictions that conflict with your customers?"}],[{"text":"Can you fine-tune, and do the outputs of a fine-tune inherit the licence?"}],[{"text":"Do you have to attribute, and where?"}],[{"text":"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."}]]},{"t":"h2","id":"where-the-free-path-is","runs":[{"text":"Where the free path is"}]},{"t":"p","runs":[{"text":"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."}]},{"t":"p","runs":[{"text":"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."}]}],"check":{"question":[{"text":"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?"}],"options":[{"text":[{"text":"Identical in practice, since the startup will never come near 700 million users"}],"correct":false,"tempting":[{"text":"Fixes on the one threshold unlikely to bind, and ignores the use, naming and attribution conditions that apply from day one."}]},{"text":[{"text":"Higher, since use rules, naming and attribution apply at once, and can change by version"}],"correct":true,"tempting":null},{"text":[{"text":"Lower, since the licensor wants commercial adopters and will not enforce on a startup"}],"correct":false,"tempting":[{"text":"Swaps a prediction about enforcement for the terms actually agreed, which a customer or acquirer will read."}]},{"text":[{"text":"Equivalent, because both allow commercial use, and that is the only term that matters"}],"correct":false,"tempting":[{"text":"Treats commercial permission as the whole licence, when the conditions attached to it are exactly what differs."}]}],"order":[0,2,1,3],"why":[{"text":"A community licence is a contract with conditions Apache 2.0 does not carry, and its terms can change between versions."}]},"prev":{"slug":"the-training-data-dispute","title":"The training data question, unresolved","minutes":10},"next":{"slug":"disclosure-rules-that-bind-you","title":"When you have to say it was AI","minutes":8}}}