Does the platform keep your prompts and train on them? Here is our whole supply chain

When you press Generate in fuse, your prompt reaches our service first, then goes to the upstream service selected for that task. The project, prompt and generated files don't disappear when generation finishes.

Answering whether we train on your data means looking at each part of that process. What fuse stores, what a model provider may do with it, and where a routing service sends it each need an answer. This account follows our review of the code and official documents on 18 September 2026.

Retention, training and human review are three different things

Retention means keeping data somewhere for a period of time. A platform stores projects so you can return to them; support staff need request records to investigate a disputed generation. The questions are how long, where, and who has access.

Training means using data to adjust a model. Check which data is used, for what purpose, and whether there is a separate way to opt in or out.

Human review means a person can read or see the content, for example during a safety check, abuse investigation or debugging.

A company can decline to train on data while keeping it and allowing authorised staff to inspect flagged content. A no-training answer still leaves two questions open.

What actually happens to one generation request

The first stop is fuse's own services and storage. Our API, background workers and database run on Railway. Projects, generation tasks, snapshots of submitted prompts and processing results leave records. Uploaded media and collected image and video outputs use Cloudflare R2, with media access handled by our Cloudflare access-control service. After the model succeeds, we still need to receive, store and display the result.

The upstream route depends on the kind of generation. The following routes are connected in the current code; whether a particular model is offered to users also depends on the active model and pricing configuration. A request does not pass through every provider listed below.

BytePlus ModelArk's international service handles Seedream images and Seedance videos. The code's catalogue includes Seedream 5.0 Lite and Pro, plus Seedance 2.0, 2.0 Fast, 2.0 Mini and 2.5. The request uses the model recorded for that task and its reference material.

BytePlus's ModelArk data-processing document says it does not use Customer Data for its own model training without prior customer authorisation. The same document sets out retention for content that triggers safety filtering. Its current service-specific terms also permit processing for safety, support and troubleshooting. The former standalone video-model terms have been incorporated into that document; the old Seedance wording is no longer a sufficient basis for describing all current models. We checked these documents on 18 September 2026.

MiniMax Global provides another image and video route, connecting to Image-01, H3 and H3 Max through api.minimax.io. Choosing the MiniMax method for preparing a video prompt also calls its Context IR service, sending the text and reference material needed by that feature.

MiniMax needs its own assessment. The API terms, checked on 18 September 2026, allow input and generated content to be used to provide, maintain, develop and improve its services. They also allow necessary maintenance, support and legal processing. The API privacy policy says personal data is not used for training to profile or target consumers. That is a specific restriction, not a blanket promise that no input or output is ever used for any training. Nor do the public documents support describing this route as having zero retention or no human access.

General text generation goes through OpenRouter. Story, script, storyboard-text and related editing features select a model from the active configuration and the user's choice, then record it when accepting the task. The route is no longer fixed to Gemini 2.5 Flash. Longer writing tasks may call the model over several turns, passing the context and tool results needed to continue the work.

Behind OpenRouter sits the provider that actually runs the model. Its data-collection documentation says prompt and response storage is off by default, with separate optional logging and data-use settings. It still stores request metadata such as usage and latency. Each model provider has its own retention and training policies; OpenRouter's defaults do not describe the whole chain. We checked this documentation on 18 September 2026.

Why we keep your prompts

When someone requests a refund because a generation came out badly, we need to know what was submitted. A model failing to follow instructions, unsuitable reference material and an unclear request may need different remedies. The final image or video often cannot tell us which happened.

Section 6 of the fuse Content Policy states that we retain prompts and necessary generation records. This policy is public and took effect on 31 July 2026. The stated purposes include support, refunds, billing and generation-quality disputes, troubleshooting, security checks and legal matters. Completing a generation does not immediately delete those records.

Section 5 sets a separate limit: Alchemint LLC currently does not use user content to build training datasets or to train, retrain, fine-tune or build models, either itself or by instructing or commissioning someone to act on its behalf. This covers prompts, text, images, video, audio, inputs and outputs.

If we introduce a model-improvement programme, we must first explain the data, purposes, models and recipients, retention period and withdrawal process. Joining requires an explicit opt-in separate from the general generation service. Declining or leaving does not affect access to that ordinary service.

We must also explain what withdrawal means for existing datasets, training already underway and completed models. An individual contribution to model parameters may be impossible to isolate or remove. We cannot promise that opting out reverses a completed training run.

Keeping content for support and refunds does not grant permission to train on it. That is Alchemint LLC's commitment about its own conduct. Upstream processing needed to run a generation remains subject to each provider's terms and settings.

Content-safety filtering can leave another copy

BytePlus's data-processing document states that inputs and outputs triggering its content filter are retained in Malaysia for 180 days. This is a retention period for a particular safety process. It does not describe all fuse data or every ordinary generation request.

The ModelArk Content filter FAQ, checked on 18 September 2026, also says the relevant feature may store and use processed image, audio and video outputs, with access restricted to authorised employees. That does not mean every generation receives human review, but it does permit access under the stated conditions.

A false positive may put creative work into that process too. Location, duration and access rights each need checking. Safety review does not transfer ownership of the work to the provider, but it can affect where copies remain.

Selecting a model does not settle who runs it

OpenRouter offers controls for upstream routing, including data-collection restrictions, zero data retention (ZDR) and provider lists. One model may be available through several execution endpoints, each with its own data policy. Its provider-policy documentation explains those distinctions.

The current fuse integration can pass provider options. Some text features require an endpoint to support specified parameters so that it returns the required response format. That format requirement does not impose a privacy restriction. In the built-in generation flows checked for this revision, we found no uniform enforcement of data_collection: deny, ZDR or a fixed provider list.

The code therefore cannot establish that the entire OpenRouter route uses zero retention. It also cannot establish that no account-level privacy settings exist. Those require a separate account check.

The current code retains raw model responses as part of the generation-result record. Investigating a particular request means checking its request, response and routing information. A public model list describes the candidate endpoints at a point in time; it cannot prove where a particular request ran.

Data-for-credits programmes may ask for another licence

BytePlus offers a separate opt-in Data Authorization Agreement for data-collaboration reward campaigns and related programmes. It includes perpetual, sublicensable rights for AI/ML improvement. That additional grant should not be treated as the default permission for ordinary API use. The document was last updated on 21 May 2025; we checked it on 18 September 2026.

OpenRouter requires another distinction. Its data-collection documentation, checked on 18 September 2026, lists private input/output logging separately from allowing OpenRouter to use the data in exchange for a 1% usage discount. Section 6.2 of its terms gives broad, perpetual and irrevocable commercial-use rights for opt-in prompt/chat logging. Section 6.3's Private Prompt Storage has a revocable licence limited to privately displaying the content. The word “logging” alone does not tell you which permission you are granting.

Section 8 of our Content Policy commits us not to enable data-contribution settings or join programmes that exchange user content for discounts, credits or other benefits without separate, explicit consent. This does not rewrite a provider's ordinary service terms, and the commitment does not replace checking the actual account settings.

How to stop a platform from training on your data

Start with the service's data controls. Check whether training or data contribution can be disabled, then read the terms linked to the setting. Personal products, paid APIs and enterprise plans can have different conditions.

Check four things:

1. Does the change apply only to new data, or also to data already collected?
2. Does it cover the platform itself, or restrict its model providers too?
3. Are safety checks, abuse monitoring and legally required retention covered separately?
4. Does the setting still apply after changing models, enabling a feature or switching plans?

If material cannot be shared with an external service, a no-training statement is insufficient grounds to upload it. Retention and access conditions matter as well.

How to read a platform's data terms

These six searches help locate the relevant passages in a privacy policy or terms page:

Search for	What to find
retain / retention	How long, where, and when deletion happens
train / improve	Which data supports which improvements, and whether you can decline
review / human	Who can inspect content and under what conditions
subprocessor / service provider	Other recipients and their roles
logging / data collection	Recording purposes, data-use permissions and optional settings
delete / deletion	Copies that may remain after deleting content or an account

We think a platform explaining its data use should name the actual recipients and disclose its own commitments alongside the upstream conditions. A no-training sentence alone is too little to decide whether to submit a client's script or unpublished reference material.

This article was revised on 18 September 2026 against fuse's mainline code, public Content Policy and the official documents linked above. The code review covered the model catalogue, integrations and routing, text-model selection, prompt and result records, and media storage. Model availability settings, upstream account settings and individual historical requests cannot be established from that code review alone; changes to those conditions need their own verification.

Also in this series: Who owns what you make with AI? · Can you use AI-generated content commercially?