AI Transparency Statement | CalorieScience

1. Document Control

Field	Detail
Version	2.1 - Last updated: November 2026
Date	November 2026
Regulatory context	UK GDPR and EU GDPR Articles 13/14 (transparency), Article 9 (special category data) and Article 22 (automated decision-making); EU AI Act (classification assessment pending)
Owner	Sunil Rongala, Privacy Lead (sunil@caloriescience.ai)

2. Our Use of Artificial Intelligence

CalorieScience uses AI tools from two providers to assist nutritionists. Google Vertex AI reads uploaded lab reports and extracts biomarker values. OpenAI generates meal-plan and recipe suggestions, classifies recipe and ingredient tags, and supports a cohort intake chat that only the nutritionist types into. No AI feature communicates directly with clients.

AI features support professional nutritionists in their work; they do not replace professional judgement. The AI does not diagnose disease, prescribe treatment, triage emergencies, or make autonomous clinical decisions. The values in an uploaded lab report are extracted either by Google Vertex AI or by a trained CalorieScience data operator, who reads the report and enters the values. Either way, the nutritionist checks every value and can correct it. All AI outputs must be reviewed, modified where appropriate, and approved by the nutritionist before being used with a client.

3. How the AI Works

Stage	Detail
Inputs	Lab-report extraction (Google Vertex AI): the uploaded lab report file (PDF or image), including the patient's name and other details printed on it. Meal planning and recipes (OpenAI): pseudonymised, allow-listed planning data only: nutrition targets, allergies, intolerances, dietary preferences, medical-condition names (for example "type 2 diabetes"), age, gender, pregnancy trimester, activity level, BMI, city or county and country, cuisine preferences, and sanitised nutritionist instructions.
Processing	Google Vertex AI identifies biomarker values and the patient name in the report so the values can be matched to the right client record. OpenAI uses the planning data, together with nutritional databases (e.g. USDA, ICMR) and established guidelines, to suggest nutrient targets, meals and recipes and to classify recipe and ingredient tags. In the cohort intake chat, OpenAI responds to text typed by the nutritionist.
Outputs	Extracted biomarker values for the nutritionist to verify; suggested macro and micronutrient targets; identified nutrient gaps; out-of-range biomarker values highlighted for the nutritionist's review; proposed meal plans and recipes; recipe tags; intake-chat responses to the nutritionist

4. Data Protection Measures for AI Processing

Measure	Detail
Pseudonymisation for OpenAI	A sanitiser enforces an allow-list on every call to OpenAI. Names, email, phone, date of birth, record identifiers, lab or biomarker values, lab-report text or files and client free text are never sent. The data sent is pseudonymised, not anonymised: it is still personal data and special category data.
Provider retention and training	OpenAI: enterprise terms with a signed data processing agreement. Requests are configured not to be stored (store: false) and are not used for model training. Zero data retention applies, so OpenAI does not keep requests for abuse monitoring. Google Vertex AI: caching disabled and data-access audit logging enabled; data is not used for model training. The zero-retention opt-out for abuse monitoring is in place, so Google does not keep requests for abuse monitoring.
Encryption in transit	Calls to the AI providers are encrypted in transit.
Identifiable data and locations	Google Vertex AI receives identifiable lab reports, including the patient's name. Processing runs in Google Cloud's Frankfurt region, with fallback regions in the EU, London and Zurich. OpenAI processes pseudonymised planning data in the United States, under OpenAI's signed data processing agreement, which includes the EU SCCs and the UK International Data Transfer Addendum to the EU SCCs (UK Addendum).
Audit logging	AI calls are to be logged with metadata only (time, request identifier, hashed client reference, field categories sent and provider), never health values. Conformance of the live log to this specification is not yet verified.

5. Safety Rules

The platform applies the following safety rules to AI outputs. They reduce risk but do not replace the nutritionist's review:

Constraint	Implementation
Allergen exclusion	Ingredients tagged as allergens in a client's profile are designed to be excluded from suggestions. The nutritionist must still check every suggestion.
Medical dietary restrictions	Specific medical flags (e.g. renal-safe, low-FODMAP, diabetic-friendly) are designed to apply limits on nutrients such as potassium, sodium, or fermentable carbohydrates.
No diagnosis	The AI does not diagnose medical conditions. It identifies patterns in provided data and flags values outside reference ranges for professional review.
No prescription	The AI does not prescribe medication, supplements, or clinical interventions.

6. Human Oversight

CalorieScience does not engage in solely automated decision-making that produces legal or similarly significant effects on the client (Article 22 GDPR). The following safeguards are in place:

Safeguard	Detail
AI is advisory only	The AI acts as a decision-support tool for the nutritionist. It cannot independently alter a client’s dietary plan or provide advice directly to the client.
Mandatory professional review	Every meal plan, extracted lab value and dietary suggestion generated by the AI must be reviewed, modified if necessary, and explicitly approved by the nutritionist before being made available to the client.
Professional override	The nutritionist can override, modify, or reject any AI output. Final professional judgment and liability rest with the nutritionist.
No direct client intervention	No AI feature communicates directly with clients. The cohort intake chat is used only by the nutritionist. The AI cannot prescribe clinical interventions.

7. Your Choices: Consent and Withdrawal

Where the nutritionist enables AI-assisted features, the AI-assisted processing layer relies on the client's explicit consent under Article 9(2)(a) UK/EU GDPR, in addition to the nutritionist's primary Article 9(2)(h) condition. Clients can refuse or withdraw this consent at any time without affecting their nutritional care, which continues without AI assistance. Clients can record or withdraw consent through their client consent form, or contact their nutritionist or sunil@caloriescience.ai.

8. Continuous Improvement

Subject to the nutritionist's documented instructions in the DPA, CalorieScience may use genuinely anonymised, aggregated operational and preference analytics (for example recipe feedback or feature usage that cannot identify a Practitioner or client) only to operate and improve the service. It does not use client personal data or special-category health data to train, fine-tune, or improve any AI model, does not autonomously diagnose conditions, and does not build individual profiles for purposes unrelated to nutrition services.

9. Review

This statement will be reviewed upon any material change to AI processing, upon any change of AI provider (currently Google Vertex AI and OpenAI), when the EU AI Act classification assessment is completed, and at least annually.

CalorieScience's provisional view is that the platform is not a high-risk AI system under the EU AI Act. A formal classification assessment is pending.