Go read the terms of service for your AI chat platform. I'll wait.

You probably didn't. Nobody does. So here's a summary of what you agreed to: the company stores everything you've ever typed. They can access it internally. In most cases, they've reserved the right to use it for training. And if someone with a subpoena asks for it, they'll hand it over.

For most AI use cases, this is fine. You're drafting emails. Summarizing documents. Asking what year the French Revolution started. Who cares if that's on a server somewhere.

But you're not drafting emails. You're writing fiction. And fiction — if it's any good — goes to places you wouldn't bring up at dinner.

What server-side storage actually means

When a platform says your data is "stored securely," they mean secured against outsiders. Not against themselves. Their engineers can query it. Their moderation team can read it. Their ML pipeline can ingest it.

• Employees can access your content. Internal moderation and debugging tools exist on every major platform. Your private fiction is readable by people you've never met.
• Your writing may become training data. Most platforms include broad data-use clauses in their ToS. Opt-out, where offered, is not the same as never-collected. Policies change. Companies get acquired.
• Law enforcement can request it. Server-side data is subject to subpoenas, warrants, and legal holds under the Stored Communications Act and equivalents in other jurisdictions.
• Breaches happen. AI companies are high-value targets. A breach wouldn't expose your email — it would expose every word of every story you've written on the platform.

None of these are hypothetical risks. They're the normal, documented, legal operation of cloud-based platforms.

The training data thing

This is the one people miss. When your fiction becomes training data, it doesn't get copied verbatim into other people's outputs. It's subtler than that. The model absorbs patterns, character archetypes, specific turns of phrase, plot structures. Your creative fingerprint gets dissolved into a statistical soup and served back to strangers.

Some platforms let you opt out of training. But "opt out" means the data was collected first. If the company changes its policy, or gets bought, or goes bankrupt, your opt-out preference is just a flag in a database that the new owner may or may not honor.

How local-first changes the equation

Underfiction stores story content locally by default — in IndexedDB on the web and local app storage on iOS and Android. If you enable story sync, server copies are stored encrypted at rest for your account so other devices can pick them up.

The tradeoff is clear: local-only storage gives you less cloud exposure, but clearing local data can delete work that has not synced. For those who want the safety net, story sync stores server copies encrypted at rest for your account.

The inference question

Local storage solves the data-at-rest problem. But what about when your prompt is being processed? The AI model has to see your text to generate a response.

This is where Venice comes in. Venice is a privacy-preserving inference provider that acts as a proxy. Your identity — account info, IP, session tokens — gets stripped before your prompt reaches the model. The AI processes anonymous text. It generates a response. The response comes back to you. Venice doesn't log prompt content.

This isn't a privacy policy. It's architecture. The system is designed so the model provider never receives identifying information. "We promise not to look" is not the same as "we can't."

What to do right now

1. Go read the ToS of whatever platform you're currently writing on. Look for "data usage," "training," and "retention."
2. Check if there's an export function. Download your stories. Some platforms make this easy. Some don't offer it at all.
3. Think about what you're writing. Casual chat? Server storage is probably fine. Deeply personal fiction? The storage model matters more than any other feature.