1. What the document is: Meta's Terms of Service for the Llama API — the hosted inference/fine-tuning API for Llama AI models. This capture bundles the incorporated-by-reference Meta Global Processor Terms (MGPT) and a Digital Markets Act dispute-resolution exhibit.
What this wiki found — complete, every page
Presence facts, not verdicts — each flagged term links to its definition and the exact clause on this page. Absence of a badge means the term isn't currently flagged here, not that the page is risk-free.
Llama API – Terms of Service
Dataset: GenGA (Generative AI Governance Archive) — 11 AI services, 2025–present
GenGA (Generative AI Governance Archive) is this wiki’s Generative-AI dataset: raw policy snapshots captured by the Open Terms Archive
genai-euproject across 11 GenAI providers (ChatGPT, Claude.ai, DeepSeek, Google Generative AI Services, Le Chat, Llama API, Meta AI, Microsoft Copilot, Perplexity, Qwen Chat, xAI). Unlike PGAv2, GenGA has no pre-tagged risk scores — all risk scoring and keyword tagging on these pages is LLM-assigned by direct reading, spanning 2025–present.
Note: This page contains documented policy clauses. Risk assessment is qualitative and context-dependent. For analysis of patterns across platforms, see:
Source: sources/GenGA/Llama API/Terms of Service/ (raw Markdown captures, Open Terms Archive genai-eu dataset — no pre-tagged JSONL exists for this source)
Last updated: 2026-03-20 (content unchanged since; see methodology note)
Note on methodology: No pre-computed
risk_score/keywordfields exist for this source. Keyword tags and clause analysis on this page are LLM-assigned, using the same rubric/vocabulary as the PGAv2 pages plus new GenAI-specific tags (output ownership,gdpr,feedback). Treat as first-pass analysis, not externally verified ground truth. This wiki’s risk-scoring system was retired project-wide on 2026-06-21 (seemethodology.md§4) — no page, GenGA included, computes or displays a numeric risk score.Content-hash dedup: 90 raw captures (2025-11-11 through 2026-04-15) collapse to 10 distinct hashes, but only 4 genuinely substantive content states — the rest are cosmetic (a missing breadcrumb-link header) or transient scrape errors. The baseline and 41 other captures across the dataset window are a transient error page (“This page isn’t available”) with an incidentally-attached fragment of an unrelated Meta document (Data Security Terms / Meta Global Processor Terms) — excluded below as failed captures. The first successful real capture is 2025-11-13.
Overview
1. What the document is: Meta’s Terms of Service for the Llama API — the hosted inference/fine-tuning API for Llama AI models. This capture bundles the incorporated-by-reference Meta Global Processor Terms (MGPT) and a Digital Markets Act dispute-resolution exhibit.
2. Input rights: Meta makes an unconditional no-training commitment, the strongest in this dataset alongside Anthropic’s Commercial Terms: “Meta will not use your Content to train any artificial intelligence models.” Meta may still aggregate and anonymize usage metadata (not your Content itself) to “test, support, develop, train, and improve… Meta’s other products.”
3. Output restrictions: None beyond standard accuracy disclaimers. You “own all rights, title, and interests” in your Inputs and Outputs; “Meta does not claim any ownership.”
4. Non-explicit predatory clauses: A viral attribution requirement — any model “created, trained, fine-tuned, or otherwise improved” using Llama Outputs, if distributed, must have a name starting with “Llama” and display “Built with Llama” branding. There’s also a significant EU-specific access restriction (see below), and a standard uncompensated, no-obligation Feedback grant.
5. Regulatory references: GDPR/UK GDPR, cited via the incorporated MGPT (Standard Contractual Clauses, UK Addendum, Module Four Processor-to-Controller transfer terms). Unique finding: this is the only document in the 11-provider dataset that names the EU Digital Markets Act — “Regulation (EU) 2022/1925 (the ‘Digital Markets Act’ or ‘DMA’)” — offering “business users” an Article 6(12) alternative dispute settlement mechanism.
6. Regional variation: A direct, named EU restriction: “If you are an individual domiciled in, or a company with a principal place of business in, the European Union, you are not permitted to use the Services to… access any multimodal AI models that may be made available via the Services.” This only limits an EU developer’s own direct API access — it doesn’t extend to that developer’s end users.
7. Key risk to users: Low overall, thanks to the unconditional no-training commitment. The main risks are the attribution requirement’s viral reach — any sufficiently-derived model must carry “Llama” branding — and the EU multimodal-access restriction, which limits what EU-based developers can build directly with the API (though not what their end users can access through a deployed product).
Flagged Keywords & Risks (LLM-assigned)
output ownership— You own your Inputs and Outputs; Meta disclaims any ownership over them. Why it matters: a user-protective pattern, consistent with most other providers in this dataset.feedback— “Meta may freely use and disclose such Feedback, without obligation or compensation to you.” Why it matters: if you send feedback, Meta can use it freely and owes you nothing — the same uncompensated pattern every other provider in this dataset uses.gdpr— The GDPR/UK GDPR/SCC framework is incorporated via the Meta Global Processor Terms. Why it matters: this is the legal backbone for how Meta handles API customer data under EU/UK privacy law.
Regulatory & Research Context
Llama API was not among the providers directly studied by Davidson et al. (2026) or Pandit et al. (2026) in this batch; applying Davidson et al.’s regulatory-gray-area framework by analogy, the EU-specific restriction barring EU-domiciled developers from accessing multimodal models directly (while leaving their end users unaffected) is a jurisdiction-contingent carve-out of the kind their framework treats as creating uncertainty for legitimate use. The unconditional no-training commitment flagged above — “Meta will not use your Content to train any artificial intelligence models” — is nonetheless a notable counterpoint to Pandit et al.’s (2026) general finding, in providers they did study, that “all terms mentioned that inputs and outputs would also be used for other purposes beyond” the immediate service.
Changes Summary
| Date | What changed |
|---|---|
| 2025-11-11 – various (41 of 90 captures) | Failed/transient captures — scraper hit a generic error page, with an incidental fragment of Meta’s unrelated Data Security Terms/MGPT document attached. Excluded from Version History. |
| 2025-11-13 | First successful capture — baseline real content: no-training commitment, EU multimodal restriction, attribution requirements, DMA dispute-resolution exhibit. |
| 2025-11-13 – 2026-02-23 | Byte-identical to the 2025-11-13 baseline (aside from a cosmetic missing-header variant). |
| 2026-02-23 | Cosmetic: “updating Meta Global Processor Terms” advance-notice banners added (announcing a March 20, 2026 effective date). |
| 2026-02-24 | Cosmetic: a UK ICO Addendum URL updated (link migration, no textual change). |
| 2026-03-20 | Genuine MGPT update takes effect: new Severability & Waiver clause added (§4.2/4.3 renumbering); the GDPR Module-Four SCC transfer trigger narrowed to apply only “if Company is located in a country… deemed not to ensure an adequate level of protection by the European Commission” (previously applied unconditionally to any such transfer). |
| 2026-03-20 – 2026-04-15 | No further substantive change (cosmetic header variants only). |
Version History
2025-11-13
- Explicit AI clause: YES
- Non-explicit predatory: NO
- Flagged keywords:
output ownership,feedback,gdpr
Clause: no-training commitment (not separately concept-tagged — the strongest in this dataset)
5.2 Aggregate Data. […] Meta may use Aggregate Data for its business purposes, including to test, support, develop, train, and improve the Services and Meta’s other products and services. Meta will not use your Content to train any artificial intelligence models.
Clause: output ownership
5.1 Content Rights. Between you and Meta, to the extent permitted by applicable law, you own all rights, title, and interests in and to your Inputs and Outputs, each as defined below (collectively, “Content”), and Meta does not claim any ownership of your Content.
Clause: feedback
5.4 Feedback. If you provide any feedback, comments, or suggestions (“Feedback”) to Meta, whether directly or indirectly, then Meta may freely use and disclose such Feedback, without obligation or compensation to you.
Clause: EU regional restriction (not separately concept-tagged)
1.3 Restrictions. If you are an individual domiciled in, or a company with a principal place of business in, the European Union, you are not permitted to use the Services to, and you agree not to use the Services to, access any multimodal AI models that may be made available via the Services. This restriction does not apply to End Users of your Integrated Products that incorporate any such multimodal models.
Clause: attribution requirement (not separately concept-tagged)
If you use the Llama AI models or any Outputs or results of the Llama AI models to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama” at the beginning of any such AI model name.
Clause: gdpr (via incorporated MGPT)
11.1 Processor Terms. […] In the event of international data transfers of Applicable Personal Data that are subject to the MGPT’s European Data Protection Requirements, Section 4 of the European Region Terms of the MGPT shall apply.
Clause: DMA citation (not separately concept-tagged — unique to this provider)
3. European Disputes. […] if you are a “business user” of our online social networking core platform service(s), please see the alternative dispute settlement mechanism offered subject to Article 6(12) of the DMA. “Business user” is defined under Article 2 paragraph (21) of the Regulation (EU) 2022/1925 (the “Digital Markets Act” or “DMA”).
2026-03-20
output ownership/feedback/no-training commitment/EU restriction/attribution requirement unchanged from 2025-11-13.
Clause: gdpr (narrowed — adequacy-decision trigger added)
4.3 […] To the extent that Personal Data which is subject to the GDPR is transferred by Meta EU Processor (acting as Company’s Processor) to Company (acting as Controller) in connection with the Applicable Terms and Company is located in a country or territory which is, at the time of such transfer, deemed not to ensure an adequate level of protection by the European Commission, such transfer shall be governed by Module Four of the SCCs (Processor-to-Controller).
Clause: Severability (new, not separately concept-tagged)
4.2. Severability & Waiver. If any portion of this MGPT is found to be unenforceable, then that portion will be limited to the minimum extent necessary to remain enforceable or, if necessary, severed and the remaining portions will remain in full force and effect.