Stone — AI Disclosure

Last updated: 2026-08-27. Policy version 2026-08-27.

What changed in this revision. The previous version was stamped 2026-06-17 and described Stone as it was then. Since then Stone gained a coach that talks to you in a running conversation, a notebook it keeps about your training, standing orders it writes for itself, a spoken rest cue, and dictation. None of that was disclosed here. This revision covers all of it.

This page describes Stone's Cloud AI system end-to-end so you know what to expect before you opt in. Stone also keeps an on-device record of every cloud call (the data shape, a retention note, and the policy version that applied), which you can clear any time.

When does AI run?

Stone is cloud-AI-first. The pipeline tries each tier in this order:

1. Cloud AI (xAI Grok, reached through a privacy-preserving proxy). Every Quick Log save when Cloud AI is on in Settings. Also used to produce the Today recommendation, to answer Coach, and to propose entries on the Learned screen.
2. On-device deterministic parser — the safety floor. Parses canonical inputs like Bench 185 x 8 with no network and no model. Always available; runs when Cloud AI is unavailable, offline, or your consent settings opt out of cloud.
3. Apple Foundation Models — a live on-device AI tier on Apple-Intelligence-capable hardware, not a future plan. It does not run before the cloud: with Cloud AI on, tier 1 is tried first, and Foundation Models handles the parse when the cloud declines, fails or is switched off. On a supported device with Cloud AI off, this is what reads your note — locally, with no network call.

Manual confirmation surfaces only when even the deterministic floor produced an ambiguous result.

If you have Cloud AI turned off in Settings, every save skips straight to the on-device deterministic parser.

Which mode runs

Cloud AI uses xAI's Grok models through a privacy-preserving proxy, in two modes:

• A fast mode for chat (Coach) and Quick Log parsing, so replies are instant.
• A deeper-reasoning mode for your Today recommendation and learned-pattern proposals, where a little extra reasoning meaningfully improves the result.

The "AI quality" setting (Balanced / Smarter) changes only how deeply your recommendation is reasoned — chat and parsing always use the fast mode. Same provider, same privacy posture, same retention, same per-request cost either way.

What gets sent

The exact bytes are pinned per call. Stone records the data shape of every cloud call on your device:

• Workout Parse — sends the freeform note you typed in Quick Log (or the photo, if you used the photo path).
• Session Recommendation — sends a compact training summary: number of recent workouts, a humanized movement-balance string ("pressing-heavy", "leg-light"), your goal / equipment / preferred session length, any injuries or avoided movements you've entered (including exercises you've marked "not here" at your current gym, so the pick fits the place you're training), and (when you've accepted any) up to a few of your highest-confidence Learned memories. When Apple Health is on, a categorical readiness hint ("fresh" / "stale") derived on-device — never raw heart-rate numbers. Your logged sets DO go: for the exercises being considered, the last session's working sets ship as weight + reps, because the model cannot pick a sane next load without seeing the last one.
• Coach — sends your question plus the same compact training context, so the answer is grounded in your history.
• Memory Update — sends summaries of the last few workouts + the current focus and (optionally) the existing memories you've accepted, so the model can refine without repeating itself.

Your voice

This is the part the previous version of this page never mentioned, and it is the most sensitive thing Stone sends.

• Dictation — the audio itself leaves your device. When you tap the mic (to log a set, in Quick Log, or in the Coach composer) and Cloud AI is on, Stone records a mono 16 kHz AAC file and sends the whole recording, not a transcription of it, to Stone's worker, which forwards it to xAI for speech-to-text. There is no on-device transcription first and no redaction. Whatever is audible in the room is in that file.
• The recording is deleted from your device as soon as the call finishes, whether it succeeded or not.
• If the cloud call fails, Apple sees it instead. On any failure — offline, a recording too large, a quota exhausted, an empty result — Stone falls back to Apple's speech recognition on the same file. Depending on your device and OS, Apple's recognizer may process it on Apple's servers rather than locally. So a failed cloud transcription does not mean the audio stayed on your phone.
• With Cloud AI off, dictation never reaches Stone's worker or xAI. It goes to Apple's recognizer only, on the same terms.
• Dictated text can make a second trip. If the on-device parser cannot read a dictated set, the recognised text (first 500 characters) plus the exercise name goes to the cloud to be interpreted.
• The spoken rest cue sends text, never your voice. It sends one of six fixed English phrases and a voice id, and gets audio back. It is off by default and needs both the spoken-cue toggle and Cloud AI.

Your conversation with the coach

• The last 38 turns of your chat are sent with each message, in your own words, each clamped to 4,000 characters. The full history stays on your device; only that window travels.
• If you opened the coach from a card, an exercise or a live session, the context of that thing is appended as an extra turn so the answer is about what you were looking at.
• Chat messages are stored on your device and, when iCloud sync is on, in your own private iCloud database. They are not stored on Stone's server: the chat route writes no request or response content to storage or logs.

The notebook and standing orders

• The notebook is what the coach remembers about your training between sessions. It lives on your device, capped at 40 entries, and the model rewrites it wholesale on a scheduled pass. Summaries of your recent workouts and your current focus are sent so it can be rewritten without repeating itself.
• Standing orders (Stone's word for a directive it writes for itself — a goal, an emphasis, a load trajectory and a horizon) are generated in the cloud from that same summary and then stored on your device and in your own iCloud. They shape what Stone suggests for the next several sessions.
• Both are cleared by Forget AI memories and by Delete all my data.

Who sees the data and how long they keep it

When this policy version changes, each call's on-device record stays pinned to the version that applied when the call was made — your historical record doesn't get rewritten by a later xAI ToS change.

What gets returned

For Quick Log, a structured workout the deterministic parser couldn't produce on its own. For Recommendations, a session title, training load, duration, and rationale bullets. For Coach, a grounded plain-language answer, optionally with links back to specific workouts in your history. For Memories, candidates with evidence pointing back at specific workouts.

Stone runs a server-side safety validator over the Today recommendation, Quick Log parses, and anything Stone learns about you, before those reach your device. It filters:

• Motivational hype and "crush it" language.
• Shaming or judgmental language about you.
• Diagnostic phrases and medical advice — that's a clinician's job, not Stone's.
• Prompt-injection echoes.

What the validator does not cover.

It only reads English. Its patterns are English phrases, and nothing in Stone tells the coach which language to answer in — it answers in yours. So if you write to Stone in another language the replies come back in that language, and this deterministic filter does not apply to them. What still applies are the coaching rules written into the model's instructions, which are language-independent, and Stone's refusal to diagnose is part of those. But the belt-and-braces filter described above is an English-only backstop today, and you should know that rather than assume otherwise.

Conversational replies in Coach, and the weekly and post-workout recaps, also do not pass through it. Those are governed by the coaching rules written into the prompt, plus a separate acute red-flag detector on the chat route that forces a stop-training-and-seek-care reply when your message describes something that needs a clinician rather than a coach. That detector is English-only too.

When your input mentions pain or injury, the safety validator forces a more conservative training-load suggestion and inserts a safety note pointing you at a qualified clinician. Stone never tries to diagnose.

Your controls

• Cloud AI is off by default. You opt in once at onboarding; toggle in Settings any time.
• Forget AI memories only clears the audit log + every AI memory. Your workouts stay.
• Delete all my data wipes local + iCloud + server-side state including the rate-limit and safety counters keyed to your device and account.
• Stone keeps an on-device record of every cloud call: timestamp, provider, data shape, retention note, and the policy version that applied at the time. It records the SHAPE, never the content — a dictation call is logged as "dictated-audio", not as your audio or its transcript. The newest 500 records are kept and older ones are deleted automatically. Clear it any time via Forget AI memories.
• Because that record syncs to your own private iCloud database, it is visible to you on your other devices and is included in the data export. It never goes to Stone's server or to xAI.

What Stone will never do

• Use your data to train a third-party model. (Stone doesn't train any model. xAI's stated policy is they don't either.)
• Send your data to any provider you didn't opt into. The one thing worth reading twice: turning Cloud AI ON opts dictation in, and dictation sends raw audio (see "Your voice"). Turning it off does not silence the mic — it routes dictation to Apple instead, which may still be a server call.
• Show motivational hype, shame language, or medical claims — enforced server-side regardless of what either model proposes.
• Send your raw heart-rate or HRV numbers off your device. When Apple Health is on, those readings are summarized to a categorical hint ("fresh" / "stale") on-device before any cloud call, and body weight is reduced to a trend phrase the same way. Sex, age and height are the exception and are sent as-is — they scale recovery windows and load, and no summary preserves that. See the Privacy Policy for the full HealthKit detail.

Open questions / honest limits

• Device integrity — every cloud call from Stone carries a cryptographic device-integrity check so the backend can refuse abuse. This is automatic; you don't need to do anything.
• Account credentials — Sign in with Apple is optional. If you sign in, Stone may occasionally ask you to re-confirm your Apple ID to keep elevated AI quotas active. You can refresh it anytime from Settings → Account.
• On-device AI tiers — additional fully-on-device AI surfaces are planned for capable hardware and are rolling out gradually. Until then the ordering is the one stated at the top of this document: with Cloud AI on, the cloud runs first.