Safety standard

CHECKED BEFORE
ANYONE LIFTS

AI programme safety, published as a standard

AI can draft a training programme in minutes. The question that matters is what stands between that draft and your client. Our answer is five layers, and this page documents every one of them.

Why we publish our safety system

Ask a general chatbot for a training programme and you get a plausible answer with a disclaimer attached. Plausibility is not safety. When a coach puts their name on a plan, the plan needs to have been checked the way software is checked: by rules, by code, and by an independent review that was not involved in writing it.

This page is the full standard we hold every Forge AI programme to. It is written to be cited: if you are a coach weighing AI tools, a journalist covering them, or another vendor deciding what to build, this is what we believe the minimum bar should be. We are an AI programme generator, so we hold ourselves to it in public.

A STRUCTURED INTAKE, NOT A CHAT

Injuries, medical conditions, allergies and intolerances are first-class fields the coach fills in, not sentences the model might remember. The intake captures age, measurements, up to three ranked goals, training experience, activity level, equipment, session length and the client's exact weekly availability. Every layer after this one reads from the same structured record, so nothing depends on the model having 'noticed' something in a paragraph.

CONSTRAINT-AWARE GENERATION

The brief the model works from is built after the constraints are. Allergies expand to their derivatives before generation begins: a peanut allergy also rules out satay and groundnut oil, a gluten intolerance rules out couscous and seitan. Contraindicated movement patterns are excluded from the brief entirely. A vegan plan does not slip in whey, because the constraint is enforced at the ingredient level rather than trusted to the model's intentions.

PROGRAMMATIC VALIDATION

Every stage of the plan is checked by code, not vibes. Allergen and diet compliance on every ingredient. Exercise screening against stated injuries and conditions. Schedule conformance, so a Tuesday rest day stays a rest day. Set, rep and rest ranges that make physiological sense. Superset structure that is actually a superset. Macro arithmetic that adds up. Plans that fail validation are corrected or regenerated before a coach ever sees them.

AN INDEPENDENT AI SAFETY REVIEW

A separate review pass, uninvolved in generation, reads the finished plan against the client profile and hunts for what rules cannot catch. This is the judgement layer, and it has earned its place: in testing it flagged a seven-minute continuous cardio block written for an asthmatic client, a deload week whose calorie drop risked hypoglycaemia for a diabetic client, overhead pressing and breath-hold lifting for a client with high blood pressure, and deep knee-flexion loading for a client with a knee injury.

THE COACH

If the safety review finds anything that stands, the plan arrives marked 'needs review' with specific, plain-English notes naming what was removed and why. A contraindicated exercise is removed, never silently substituted: the professional fills the gap, because guessing at a replacement is exactly the failure this system exists to prevent. Every plan, flagged or not, is reviewed by the coach before a client sees it. That is not a formality; it is the final layer.

Testing evidence

What the review layer actually caught

Those four catches above are real, from validation runs across simple, complex and adversarial client profiles. Complex intakes (multiple conditions, allergies, injuries) consistently produce reviewable plans with specific notes rather than confident mistakes, and the meals meet their calorie targets by construction because portions are arithmetically scaled. We publish the categories, not the internals, because the internals are a moving target and the bar should not be.

What this system does not do

Honest limits

It cannot see the client. No camera, no movement assessment, no form check. The coach on the floor still owns everything that happens in the room.

It does not diagnose anything. It screens against conditions the coach has stated, and escalates judgement back to the professional. It is an engineering safeguard, not a clinician.

It will not catch what it was never told. An undisclosed injury is invisible to every layer, which is exactly why the structured intake (layer one) exists and why the coach's review is the last line, not the first.

It does not remove responsibility. The coach reviews and owns every plan. A flagged plan is the system working, not a product failure.

AI safety questions

Is AI-generated programme design safe for clients?
It can be, when safety is engineered rather than disclaimed. NForge's Forge AI runs every plan through five layers: a structured intake, constraint-aware generation, programmatic validation, an independent AI safety review, and the coach's own review before anything reaches a client. Plans that need professional judgement arrive flagged with specific notes, not smoothed over.
What happens when the AI gets something wrong?
Validation catches most errors before anyone sees the plan, and Forge AI's independent safety review reads the finished result against the client profile for what rules cannot catch. When it finds a conflict, the exercise or ingredient is removed and the plan is flagged with a specific note. Nothing contraindicated is ever silently substituted, and the coach decides what fills the gap.
How does this compare to asking ChatGPT for a programme?
A chatbot's safety system is a disclaimer. Forge AI's is five layers, including code that checks every ingredient and exercise against the client's stated constraints and a separate review pass that reads the finished plan again. ChatGPT may produce a fine programme; it has no mechanism that guarantees the seventh ingredient was checked against the stated allergy.
Does the AI know about my client's medical condition?
Only what the coach states in the intake. Medical conditions are structured fields, screened during generation and re-checked in the independent review, and plans that interact with a condition arrive flagged for the coach's judgement. Forge AI does not diagnose, infer, or ask clients for health data behind the coach's back.
Who is responsible if a client is injured following an AI-generated plan?
The coach, and we design for that honestly rather than around it. Every Forge AI plan is explicitly a draft the coach reviews, edits and owns; flagged plans say so in plain English, and the app shows the coach a review notice until the plan is approved. The system escalates to the professional by design, because that is where responsibility belongs.

SEE THE STANDARD
IN PRACTICE

Generate a free plan and watch the review discipline yourself, or read the full Forge AI feature.

Keep reading: The Forge AI feature · The free generator · Pricing — free, 7.9% only when clients pay

AI Safety for Training Programmes — Our Standard | NForge