Your name never reaches the model: how Ethyx pseudonymizes identity

If you use ChatGPT, Claude, or Gemini directly, the provider knows your account. Your conversations live in their product, tied to your billing profile and login. That is convenient when you want one vendor's app—but it also means every prompt is easy to associate with you as an individual customer.

Ethyx is built differently. You sign into Ethyx; we send your turns to upstream models on Ethyx's operator API credentials, not your personal OpenAI or Google account. On top of that separation, we pseudonymize account identity—your name, email, and attachment filenames—before each provider request, then restore the real values in the reply you see.

This post explains that two-layer model: what providers actually receive, how pseudonymization works in a streaming chat app, and what we still cannot hide.

Two layers of separation

Layer 1: Ethyx's credentials, not yours. User bring-your-own-key is retired. Subscribed Ethyx users run on operator-managed upstream keys configured by us. When OpenAI or Google processes a turn, the billing and organizational identity on their side is Ethyx, not your personal ChatGPT or Gemini login. Your chat history lives in Ethyx (encrypted at rest), not in the provider's consumer chat product under your name.

That does not mean your prompts disappear into an anonymous shared pool. Each turn is still a distinct API request. Providers still process your message text to generate an answer. What changes is cross-linkage: a provider is less likely to connect that request to a separate personal account you may also have with them.

Layer 2: Pseudonymized identity inside the request. Even on Ethyx-originated traffic, we replace injected account metadata before the HTTP call:

Field	What the provider sees (default)
First, last, and preferred name	Deterministic fake names (e.g. "Jordan Hollis")
Account email	A fake address on a dedicated domain (e.g. @private-user.ethyx.example)
Attachment filenames	Pseudonymized basename
The reply in your browser	Real values restored before you see them

If pseudonymization is ever disabled on our side, Ethyx withholds your name and email from the model entirely rather than sending them raw—it fails closed.

You can see this in the product: the chat composer shows a Pseudonymized indicator when identity obfuscation is active, and Security settings summarize encryption, pseudonymization, and retention posture in plain language.

Why streaming makes this harder than it sounds

Pseudonymization in a one-shot API call is straightforward. In a streaming chat client with conversation history, it is not.

Every turn, Ethyx replays prior messages from our database (real text on our side), assembles context, and sends a fresh provider request. The same real name must always map to the same fake name across turns—otherwise yesterday's thread would not match today's. Our pseudonyms are deterministic per user, keyed with an HMAC derived from the application secret, so we can compute identical fakes at every injection site without shared runtime state.

When the model streams tokens back, it may echo a pseudonym ("Email Jordan Hollis at…"). Ethyx keeps a per-turn reverse vault mapping fake tokens to real ones and swaps them back as the stream arrives, before the text reaches your screen or is persisted as the assistant message.

Providers cannot reverse the mapping without Ethyx's secret. Ethyx can—which is why we call this pseudonymization, not anonymization. We still know who you are; we are limiting what identity metadata crosses the provider boundary by default.

What this does not do

We will not bury the limits in footnotes.

Your prompt text still goes to the provider. Inference requires plaintext at the model. Encryption at rest protects stored history on our side; it does nothing to the live request. This is not end-to-end encryption.

You can still identify yourself in the message. Pseudonymization targets account metadata we inject (system context, optional account-context tool, filenames)—not free-typed content. If you write "I'm Jane Doe at Acme Corp," the model reads Jane Doe.

Image and document contents are visible. We can pseudonymize a filename like Jane_Doe_Resume.pdf; we cannot pseudonymize what is inside the file or a photo. The composer warns you when attachments will be visible to the provider.

No-retention flags are requests, not guarantees. For OpenAI, xAI, and Gemini chat APIs, Ethyx sends store: false by default where supported. Anthropic and Perplexity describe no training on commercial API traffic in their terms. Some catalog models carry a Bronze privacy rating because upstream retention or training posture is weaker—we show that before you pick a model on the catalog. We do not have zero-data-retention agreements with providers, and org-level ZDR in a provider console is an operator step, not something automatic from using Ethyx.

Ethyx still holds your account. We encrypt chat history at rest, operate the pseudonym vault, and could—in principle—link activity back to you. Honest privacy copy admits that.

For the full trade-off list (MFA, retention opt-outs, what we encrypt), see How honest is your AI chat app about your data?.

ChatGPT or Gemini directly vs through Ethyx

	ChatGPT / Gemini directly	Ethyx
Provider billing identity	Your personal account	Ethyx operator API account
Chat history location	Provider's product (for that account)	Ethyx (encrypted at rest)
Name / email in API metadata	Typically your account profile	Pseudonymized or withheld
Message body	Provider processes it	Provider still processes it
Retention / training	Their consumer or API terms	Their API terms plus Ethyx retention flags where supported

Ethyx adds separation layers—organizational credentials plus pseudonymized identity metadata. It does not make you invisible to upstream models, and it is not a substitute for avoiding cloud AI when your use case demands stricter confidentiality.

For a broader comparison of product shape (history, projects, multi-model workspace), see Ethyx vs ChatGPT.

Where to read more

• How honest is your AI chat app about your data? — encryption at rest, retention flags, and the full caveat list
• Model catalog — privacy rating per provider/model
• Privacy policy — what Ethyx collects and how we use it
• Security settings (/account/security after sign-in) — in-app transparency panel

We would rather explain the boundary precisely than sell a privacy guarantee we cannot keep. If that matches how you want to use AI, we built Ethyx for you.

---

Ethyx is in closed testing with an access code today. Subscriptions will launch after closed beta.