Responsible AI
Showcase document. CookingVerse is a fictional product built to demonstrate design and engineering work. This document is a realistic template, not a binding agreement, and it is not legal advice. Replace it with counsel-reviewed text before using it for a real service.
The AI in CookingVerse decides what you eat, what you buy and how hot your pan gets. That earns some obligations. This page describes what the models do, how we test them, where they are not allowed to act alone, and what we will not do with your data.
Where AI is used
- Meal planning — ranking and sequencing recipes against your profile, calendar and pantry.
- Recipe adaptation — substitutions, scaling and technique rewrites.
- Cart compilation — normalising ingredients and matching them to store catalogue items.
- Voice guidance — speech recognition and conversational step guidance.
- Appliance orchestration — deriving a timeline of device commands from a recipe.
Hard limits
Some decisions are never left to a model:
- Allergens are rule-based, not inferred. A declared allergen filters candidates deterministically before ranking. A model cannot override it, and it cannot “reason” that a trace is acceptable.
- No autonomous purchasing. Every basket needs your explicit approval. Recurring delivery repeats a basket you already approved and pauses on any substitution.
- Appliance commands are bounded. Temperatures and durations are clamped to the manufacturer’s certified envelope. Requests outside it are refused, not clipped silently.
- No unattended ignition. We never start a heat source that was not already on without a confirmed presence signal.
- Food-safety thresholds are fixed. Minimum internal temperatures come from a curated table, not from generation.
Showing our work
Every adaptation renders a diff: what changed, why, and the expected effect on texture, timing and seasoning. Confidence is surfaced in plain words — “this is a common swap” versus “this is unusual, taste as you go”. Any AI-generated step is labelled as such, and one tap reverts to the original recipe.
How we test
- A held-out evaluation set of 12,000 recipes across 40 cuisines, refreshed quarterly.
- Adversarial suites for allergen leakage, unsafe temperatures and impossible timelines. These are blocking release gates, not dashboards.
- Cuisine-fairness review: we measure whether adaptation quality degrades for under-represented cuisines, and report it internally per release.
- Speech recognition benchmarked across accents, ages and atypical speech; known gaps are published on our accessibility page rather than buried.
- Professional chefs and a registered dietitian review sampled output every release.
Training data
Our models are trained on licensed recipe corpora, public-domain culinary texts and content our partners have granted rights to.
- We do not train on your private recipes, plans, pantry or voice data.
- We do not pass your content to third-party model providers for their training.
- Community recipes are used only where the author published them under our community licence, and we honour unpublishing in the next training cycle.
- We use aggregate, de-identified signals — such as how often a substitution is rejected — to improve ranking. You can opt out in Settings → Privacy.
Human oversight and redress
Rejected adaptations are reviewed weekly and feed a curated deny-list. Any output can be reported in one tap; safety reports are triaged within 24 hours. If a model materially misled you, write to [email protected] — a person, not a model, will answer.
We publish a model card and a changelog for each significant model release, including what changed, what improved and what regressed.
Contact
Questions about this document go to [email protected], or by post to CookingVerse Labs, Inc., 1180 Kearny Street, Suite 400, San Francisco, CA 94133.