AI Disclosure

1. Yes, it's AI. We're not going to pretend.

Stupid Simple Fitness is built around large language models. Specifically, we usually use Anthropic's Claude family, with OpenAI's GPT family as a backup in case of an Anthropic outage. The product loop, the workout prescriptions, the calibration, the written feedback, the under-the-hood reasoning views, all of it routes through AI. There is no hidden bullpen of human coaches behind the curtain.

Honesty about that is part of the brand. So is the fact that the AI is doing something useful: it can hold your goal, your history, your last four weeks of logs, your sleep, your equipment, and your timeline in mind at once and produce a coherent, individualized plan in seconds. A human can do this; a human can't do it for $5 a month.

2. What the AI actually does

• Plan generation. When you complete intake (or when we recalibrate), the AI is given your goal, history, restrictions, equipment, and the prior plan/log if any, and produces a structured plan: sessions, lifts, sets, reps, target loads, rest, and notes.
• Recalibration. Once a week (and on demand), the AI reads your logged training, RPE, sleep and adherence signals, and adjusts the next block. The reasoning chain is shown in the under-the-hood view.
• Written feedback. Your weekly read, callouts on the dashboard, in-context notes on a workout, and pushback when something looks off.
• Goal pushback. The AI is explicitly instructed to challenge goals that are unrealistic or unsafe, and to refuse with a proposed alternative.

3. What the AI does not do

• It does not provide medical advice, diagnose, or treat anything. See the Medical Disclaimer.
• It does not prescribe nutrition, supplements, or medication.
• It does not impersonate a human. It will not pretend to be your coach, your trainer, or your friend.
• It is not an open chat. There is no general-purpose assistant; the AI surfaces are scoped to specific product tasks.
• It does not learn from your data in the sense of updating its own weights. Your logs make your plan smarter for you; they do not change the underlying model.

4. Where it can go wrong

Being honest about the failure modes is more useful than pretending they don't exist:

• Hallucinated specifics. AI sometimes invents detail that sounds correct. We mitigate by constraining outputs to schemas, citing the data behind a recommendation, and refusing to surface anything we can't ground.
• Misreading context. The AI works from what it has. If you didn't log it, it didn't see it. If you under- or over-described an injury, the plan will reflect that.
• Cargo-cult prescriptions. AI can over-rely on patterns from public training literature that don't fit your individual case. The under-the-hood view exists so you can sanity-check.
• Calibration drift. Over a long block, small errors can compound. Recalibration and human review (yours) are the brake.
• Brittle pushback. Adversarial users can sometimes maneuver around safety rails. We adversarial-test before each prompt change, and we improve the rails when we find a gap. We will never be perfect.

5. Our safety rails

The system prompts that drive the AI in Stupid Simple Fitness instruct it to refuse, redirect, and propose modified alternatives for:

• Eating-disorder-pattern goals or language.
• Weight-loss timelines that imply unsafe caloric deficits.
• Aesthetic-deadline goals where the math doesn't work without harm.
• Post-surgical or post-partum prescriptions without explicit medical clearance from the user.
• Any prescription that contradicts a medical restriction the user has stated.
• Reckless biomechanics given stated injuries or training history.

A test suite of bad-goal inputs runs against every system prompt change before deployment. Adversarial testing is part of the release process, not optional.

When the AI pushes back, it does so in our voice: protective, blunt, occasionally funny, never preachy. It always proposes an alternative, because "no" without "try this instead" is just an obstacle.

6. What data the AI sees

For any given AI call, only the data needed for that task is sent to the AI provider. For example:

• Plan generation: your goal, intake answers, restrictions, equipment, and (if any) recent training history and adherence signals.
• Weekly recalibration: the last block's prescribed plan, the logged completion against it, RPE, sleep, and your stated goal.
• Pushback / refusal evaluation: the inputs in question, plus the safety rail prompt.

Your raw email, billing details, and full account profile are not sent to the AI provider. The data flow is laid out in detail in the Privacy Policy.

7. Our agreement with the AI providers

Under our contracts with our AI providers (usually Anthropic, with OpenAI as a backup in case of outages):

• Data we send is not used to train their models.
• Data is not retained beyond what is operationally required to serve the request.
• Each provider is bound by their own published privacy and security commitments.

If we add or change AI providers, we will hold any new provider to terms at least as protective as these, and we will update this disclosure.

8. The under-the-hood view

Every AI-generated recommendation in the Service ships with an "under the hood" disclosure: the inputs the AI considered, the reasoning chain it produced, and the confidence it expressed. You don't have to look. Most users won't, most days. But it is always there, and that is intentional.

If a recommendation looks wrong, open the under-the-hood view first. It usually explains the disconnect (a missed log, a stale restriction, a misread signal) faster than going back and forth.

9. You are in the loop

The AI builds plans. You do them. You are the one in the room with your body, your equipment, your day, and your sense of whether something feels right. If a workout the Service prescribed today does not pass that sanity check, do not do it. Log why, and the next plan will be better for it.

10. Reporting an AI problem

If the AI generated something unsafe, inappropriate, factually broken, or just deeply weird, please tell us: help@stupidsimplefitness.com. Include the under-the-hood snapshot if you can. We read these. They become test cases.