AI Disclosure

Last updated 2026-09-09.

Short version

Screen Guardian uses AI text-to-speech (TTS) to speak the guided passcode ceremony aloud so you never see the passcode as written digits. The AI provider is OpenAI (with Google Translate TTS as a fallback). Most audio you hear is served from our cache — never sent per-listen — for the fixed phrases like “Alright. We’re going to start the passcode now.” The digit prompts your specific ceremony uses (“Tap six.”, “Tap eight.”) are generated on demand and streamed through the provider.

1. What AI we use

OpenAI — text-to-speech via the /v1/audio/speech endpoint. Model varies (gpt-4o-mini-tts, tts-1-hd, or tts-1) depending on phrase importance.
Google Translate TTS — free-tier fallback voice provider if OpenAI is unavailable.
We do not use any AI for text generation, chat, moderation, or personalization. The product does not talk to a language model about you or your data.

2. What data gets sent to OpenAI

The text we send to OpenAI’s TTS endpoint is the phrase to speak — and only that:

Fixed narrator phrases from a pre-authored script: “Alright. We’re going to start the passcode now.”, “Hit backspace once.”, “What’s two plus three?”. These are cached (see §3) so most users trigger zero fresh provider calls.
Digit prompts derived from the random passcode we generated for your commitment: “Tap six.”, “Tap eight.”, etc. The sequence and individual digits are unique per commitment; the raw 4-digit passcode is never sent as a whole string.

We do not send your email address, your account identity, your device identifiers, or any other profile data with the TTS request. The provider sees the phrase text and the voice-model parameters.

3. Audio caching — what actually happens per user

For fixed script phrases (the ceremony intro, math examples, closing lines), we cache the generated MP3 in our own database the first time it’s ever generated globally, then serve subsequent users from the cache. So:

“Alright. We’re going to start the passcode now.” — sent to OpenAI once (the first user), reused for everyone else.
Your personal digit prompts (“Tap six.”) are currently generated on-demand. Because tens of thousands of users share the same 10 digit-prompt phrases, cache hit rate on these is also high, but the cache-hit ratio for any single user cannot be guaranteed.

Cache metadata (byte size, hit count, timestamp) is stored; the text of the phrase is never stored in our per-user usage log — only the character count and cost.

4. Does OpenAI train on our data?

Screen Guardian uses OpenAI through its standard commercial API. Under the provider’s general commercial-API terms, content sent through the API is not used by default to train their models, though the provider may retain content for a limited period for safety, abuse-prevention, and policy enforcement. We rely on the provider’s default posture and do not currently set an additional opt-out header on requests; we will update this section if the provider’s terms change materially or if we adjust our request configuration.

If you would like the up-to-date provider terms, see OpenAI’s published API data-usage policy for commercial accounts.

5. What you can and cannot opt out of

The voice ceremony itself is core to the product. We chose voice specifically so you never see the passcode written down. There is no visible-digit fallback, and creating one would defeat the mechanism.
If you do not want TTS audio processed through a third-party provider, the product will not work for you today. This is an honest constraint, not a marketing angle.

6. Voice output — is it “you”?

The voice is a fixed synthetic voice (currently ash from OpenAI’s standard voice library). We do not clone your voice, mimic someone you know, or personalize the timbre.

7. Changes to this disclosure

If we add a new AI provider, expand what we send, change caching behavior in a way that materially affects your data flow, or wire an opt-out, we’ll update this page and reference the change in Privacy Policy §7.