The_Ledger

Three weeks after closing a $30 million round at a $500 million valuation led by Craft Ventures, ComfyUI published a new Terms of Service on May 13. It carries a creator-friendly headline pledge, plus one sentence directly underneath it that quietly reserves the right to do something the pledge was supposed to prevent.

The pledge: “Comfy will not use Input or Output to train generative AI or diffusion models. Comfy may, however, collect and use limited metadata derived from Customer’s use of the Comfy Products, such as prompt classifications, workflow structures, and node configurations, to improve the performance, functionality, and user experience of the Comfy Products.”

The first sentence protects your content. The second carves out a third category, the behavior around your content, and gives it none of the same protection. And it names that category plainly: prompt classifications, workflow structures, node configurations. The structure of the workflow you spent months refining is exactly what Comfy reserves the right to collect and use.

In human terms: A VFX freelancer has spent two years on her workstation building texture-synthesis workflows for ad clients: the node sequence she refined by trial and error, the sampler choices, the ControlNet she chained to her inpainting pass. That craft lives in JSON files on her drive. This week a brief requires the job to run on hosted infrastructure, so she clicks “Launch Cloud.” The interface looks the same; the workflows transfer cleanly. What’s changed is that her workflow structure, the thing she actually got paid for understanding, is now metadata Comfy is permitted to use to improve its products. Her prompts get classified before they ever reach a training corpus, but the classification carries her intent forward. The pixels she generates are protected. The recipe that produced them is not.

Why this matters: ComfyUI sits underneath a large share of professional VFX, ad-studio, and game-studio AI pipelines. This is the cleanest example this quarter of a Tier-1 venture round reshaping the legal scaffolding around an open-source tool. The no-training pledge is real, but it has two holes the headline doesn’t reach: Partner Nodes downstream and the metadata carve-out upstream. If workflow intelligence becomes the next competitive layer in creative AI, that carve-out is the sentence the contract was built to enable, while still saying never.

The affiliates clause in ElevenLabs’ privacy policy is the list of legal entities authorized to process user voice data. Last quarter it had three names on it. This week it has five.

The prior list named “Eleven Labs Ltd., Eleven Labs Poland sp. z o.o., and Eleven Labs Japan Godo Kaisha.” The May 20 update inserts two more: “Eleven Labs Auto India Private Limited, Eleven Labs Technology Middle East Limited.”

“Eleven Labs Technology Middle East Limited” is a reasonably standard regional-operations entity. “Eleven Labs Auto India Private Limited” is more interesting.

The Seedance 2.0 video model is a ByteDance product. The official ByteDance Seedance project page lives at seed.bytedance.com. The model is available to creators through ByteDance’s Jimeng AI app (in China), Dreamina, and CapCut.

None of those is the site we tracked this week.

The site we spotted change in is seedance2.ai. It markets itself as offering Seedance 2.0 video generation. Its newly-updated privacy policy contains an unambiguous no-training pledge:

Your data is never used to train AI models. We do not use your prompts, uploaded content, or generated videos to train, fine-tune, or otherwise improve any AI model. Content you submit is processed solely to deliver the requested generation result back to you.

It is also the kind of pledge that rewards a closer read. The policy names no corporate operator — only “Seedance2.ai” and “we.” There is no company name,

LangChain’s May 20 ToS update is the kind of revision that reads as routine until you look at what each change does.

Three substantive moves. First: the training opt-out language in Section 4.2 picked up four new words. The prior clause read “LangChain agrees that it will not use Customer Data to train on, develop, or otherwise improve its products.” The updated clause adds: “…including any large language models.” On its face, the addition just closes a possible loophole — but those four words are a meaningful concession in an industry where “our products” and “models we train” have spent two years being negotiated as separate categories.