Third-party inference, subprocessors and a second privacy policy

Many scoring tools send your photo to another company's hosted model rather than running their own, so the file is seen by two organisations and both sets of terms apply. Building and hosting a vision model is expensive, so small teams call an API instead and build their product as a layer on top.

What "subprocessor" means here

Privacy language borrowed from enterprise software calls this second company a subprocessor: an outside organisation the primary service relies on to deliver part of the product, given access to some of the same data as a result. An inference API is a textbook subprocessor. Under the GDPR the chain is meant to be controlled: Article 28(2) says a processor "shall not engage another processor without prior specific or general written authorisation of the controller." The company you uploaded to has a relationship, usually contractual, with the company whose model actually looked at the pixels, and that second company's own retention and usage practices now apply to your file whether or not the first company's marketing mentions it.

A policy that is honest about this says so directly - names the model provider, links to that provider's own terms, and states what data the provider is contractually allowed to do with a request. A policy that only describes the primary service's own practices and stays silent about where inference actually runs is not lying, but it is answering a narrower question than the one that matters.

What a subprocessor's terms typically allow

This varies by vendor and by the specific agreement in place, which is exactly why it is worth checking rather than assuming. Some inference providers contractually commit to zero retention - the image is processed and discarded, never logged or used for anything else. Others retain requests for a fixed window for abuse monitoring, which is defensible but is still a second copy of your file living somewhere the primary service does not control directly. The two largest providers both default to this: OpenAI's API data controls set a standard 30-day abuse-monitoring retention and scan submitted images for CSAM even under zero retention, and Anthropic's privacy center says API inputs and outputs are deleted within 30 days, with exceptions such as usage policy enforcement. A minority reserve the right to use submitted content to improve their own models unless the customer opts out at the account level, which means the primary service's own promise not to train on your data can be undermined by a subprocessor whose default points the other way. Whether uploading actually trains anything at either layer is a related question worth separating from this one, since a "we don't train on your data" claim about the primary service says nothing about the API it calls underneath.

Why this rarely gets a clear answer

Naming the specific vendor is a competitive and contractual choice most companies would rather not make publicly, and terms with API vendors sometimes include confidentiality clauses that make disclosure genuinely difficult even for a service that wants to be transparent. That is a real constraint, not just an excuse, and it means the absence of a name is not automatically a red flag. What is worth expecting regardless is a statement of the category - "we use third-party inference providers" - and the retention posture those providers commit to, even without a company name attached. A model card, where one exists, sometimes fills this gap from the vendor's side rather than the product's.

What to look for

The practical question is simple: does the privacy policy acknowledge that inference may run outside the company you uploaded to, and does it say anything about what that means for retention. Rate Cock states plainly which parts of its pipeline call external inference and what those providers commit to, rather than describing only its own storage as though that were the whole picture. The same layered-vendor question shows up wherever a service outsources any specialised step - how Measure My Cock's measurement method handles data that passes through a third party is worth reading with the same skepticism. A tool that instead runs everything itself, with nothing sent to an external API at all, removes this specific question entirely; self-hosted, open-weight scoring is one way that trade gets made. None of this applies to a human reviewer in the same way - a commissioned review on Rate Penis involves a person rather than a chain of API calls, which is a different set of privacy tradeoffs, not fewer of them. Whichever tool you use, comparing what different tools actually disclose about their pipeline is worth doing before comparing what they score.