What does Meta's Llama Acceptable Use Policy restrict, and what happens if you breach it?

## An acceptable-use policy with contractual force Many vendors publish usage guidelines that read as aspirational. Meta's is not one of them: the Llama Community License incorporates the Acceptable Use Policy by reference and conditions the grant of rights on compliance with it. Using Llama in a prohibited way is therefore a breach of the licence itself, with the remedies the licence provides — not merely a policy violation you argue about in support tickets. This is one of the two clauses that keeps Llama outside the Open Source Definition, because open-source licences may not restrict fields of endeavour. ## What the policy covers The policy is organised as categories of prohibited use. Without reciting it clause by clause, the substance spans: - **Illegality and rights violations** — using the model to break the law, or to violate the rights of others, including intellectual-property and privacy rights. - **Violence, terrorism, and weapons** — including the development of weapons, and materials facilitating attacks. - **Harm to children** — exploitation, abuse material, and related conduct. - **Cyber-offence** — creating malicious code, malware, or tooling for unauthorised access to systems, and disrupting the security of computer systems. - **Critical infrastructure** — interference with the operation of critical infrastructure and safety-critical systems. - **Deception and impersonation** — fraud, disinformation campaigns, impersonating a person or entity, and generating content that misrepresents its origin. - **Regulated-professional deception** — presenting model output as qualified professional advice without appropriate disclosure. - **Failure to disclose risks** — not informing end users of known dangers of your AI system. The policy also carries reporting channels for violations and for model issues. ## The two structural facts that matter in a design review **It can change under you.** The policy lives at a URL and is incorporated by reference. Meta can revise it without issuing a new licence agreement, meaning the operative restrictions on a deployment shipped last year may not be the ones you read last year. This is genuinely different from a static licence file vendored into your repository. The mitigation is procedural: snapshot the policy text on the date you accept it, and re-check it at each model upgrade and at a fixed cadence. **Breach terminates the licence.** The term-and-termination section provides that the agreement continues until terminated, that Meta may terminate it if you are in breach, and that upon termination you must stop using and delete the Llama Materials. The relevant provisions on survival, disclaimers and liability persist past termination. There is no cure period spelled out that you can rely on. For an on-prem deployment this is a concrete operational event: you would need to pull weights from every image, registry, cache, and backup — which is why teams that have thought about it keep an inventory of where the weights physically live. Separately, the intellectual-property section provides that if you institute litigation against Meta or any entity alleging that the Llama Materials or their outputs infringe intellectual-property rights, your licence terminates as of the date the claim is filed. That clause is worth flagging to counsel in any organisation with an active patent programme. ## Pass-through to your users Because you must provide a copy of the agreement to any third party who receives the materials, the acceptable-use restrictions bind your recipients too. If you operate a platform where customers deploy or download Llama-derived models, your terms of service should reflect the restrictions, and your abuse-handling should be able to act on them. Practically, teams building customer-facing products on Llama end up implementing: - Usage terms that mirror the prohibited categories. - Content filtering and abuse detection at the application layer, since the licence obligation is about *use*, and your product is the surface where use happens. - A disclosure surface telling end users they are interacting with an AI system and what its known limitations are, which addresses the disclosure category directly. - An incident path for reporting and responding to violations. ## The judgment an interviewer is testing The list of prohibited categories is not the interesting part — anyone can read it. What distinguishes a senior answer is recognising that (a) the restrictions are contractual and enforceable, (b) they are mutable at Meta's URL rather than pinned in your repository, (c) the remedy is termination with a delete-and-cease obligation that has real infrastructure consequences, and (d) the obligations flow through to your own users, so they must show up in your terms of service and abuse tooling rather than only in a legal file. That is the shape of an answer from someone who has actually shipped a Llama-based product rather than read the licence once.