Skip to content
SignUnder home

Inference platforms

Does your inference platform attach Content Credentials or a watermark, and what do its terms say about who carries the marking obligation? One page per platform, read from its own documentation.

Of eleven inference platforms that host open-weight image, video and audio models, one attaches Content Credentials to generated output, and it signs them in its own name. The other ten attach nothing, and all eleven place responsibility for output on the customer in their terms. Read from each platform's own pages on 12 October 2026.

Platform Marks output Watermark What the terms say EU region Terms dated
fal.ai Yes, signed as fal (hosted apps) Yes "Customer must use the AI Features and the Output Content only (i) in a lawful manner" None stated; US-based 8 September 2026
Replicate No No "CUSTOMER IS ULTIMATELY RESPONSIBLE FOR ITS USE OF ANY MODELS OR OUTPUT" None stated 1 April 2026
Together AI No No "You shall be responsible for all of Your Content and Output, including ensuring that it…" Dedicated endpoints only 19 May 2026
Runware No No "Users are solely responsible for ensuring that their inputs and the use of AI-generated…" Yes: EU pods, announced 4 Aug 2026 29 September 2026
Fireworks AI No No "You acknowledge and agree that you are entirely responsible for your Content, and for a…" "Contact sales" for EU residency 10 July 2026
Baseten No No "Customer assumes and shall perform all obligations imposed on the developer, provider,…" Yes: eu region 15 September 2026
Hugging Face Inference Providers No No "You are solely responsible for the Content you post, publish, display or otherwise make…" UNVERIFIED 15 September 2022
Modal No No "Customer is solely responsible for reviewing, validating, editing, and amending any Output" Yes: EU regions; some metadata in the US 1 May 2026
Novita AI No No "YOU ARE SOLELY RESPONSIBLE FOR EVALUATING AND VERIFYING ALL OUTPUT BEFORE RELYING UPON IT." Germany GPU zones 5 August 2026
WaveSpeedAI No Sells a remover "Customer is responsible for their use of the Services, including any content generation…" None stated 28 September 2026
DeepInfra No No "it is solely responsible for its own products, services, and end-user relationships, fo…" None stated 17 August 2026

What "marks output" means

A platform marks output when it attaches a C2PA manifest, signed with its own certificate, to the file it returns. fal does, for its hosted applications, and a verifier shows fal as the signer. No platform among the eleven signs in the customer's name, and under the C2PA Certificate Policy none can: a signing certificate is issued only to a product on the conforming-products list, with a subject that matches the listing.

Why the terms column matters

The Commission's guidelines on Article 50 of the AI Act, adopted on 20 July 2026, treat a business that integrates a model by API into its own user-facing system as the provider of that system (our reading of 8 October 2026). The terms say the same from the other side: each platform makes the customer responsible for output, and Baseten does so using the AI Act's own words. Both readings land in the same place. If your users see what these platforms generate for you, the machine-readable marking that Article 50(2) requires is yours to add.

Method

Each page quotes the clause from the platform's current terms, with the section and the date read, and says whether the platform's documentation mentions Content Credentials, C2PA or watermarks. Region information was read on 8 October 2026. The reads are repeated when a platform changes its terms, and every change is logged on the page. To cite this table: SignUnder, "Which AI inference platforms mark their output?", signunder.com/platforms, as read on the date shown.

Changes

  • 2026-10-12 — First published, replacing the table on the homepage, which had been read on 8 October 2026.

SignUnder is not a law firm and this is not legal advice.