AI & Multimodal Vision Disclaimer
Unitrition employs advanced computer vision and natural language models to accelerate meal logging in the Meal Builder. However, algorithmic parsing is a PROBABILISTIC ESTIMATION. It is not a calibrated kitchen scale, chemical spectrometer, or medical diagnostic tool (SaMD Class II). You must always visually inspect detected ingredients and calibrate portion weights to match your actual meal plate.
NEVER rely on Unitrition’s photo scanner to verify the presence or safety of life-threatening allergens (such as trace peanuts, tree nuts, shellfish, soy, or hidden gluten for celiac disease). Computer vision cannot detect microscopic cross-contact, shared fryer oils, or blended thickening flours invisible to camera sensors.
Prior to egress, photos are re-encoded onto a browser canvas, stripping GPS coordinates, device identifiers, and camera metadata.
Photos are analyzed strictly in server RAM and immediately destroyed upon extracting food tags. Images are never written to disk.
Invocations operate via protected enterprise API endpoints. Your meal plates are never utilized to train third-party foundation models.
The model segments food items exclusively. All facial biometric identification, emotional analysis, or human profiling is explicitly excluded.
- Visible contours and plate geometry
- Surface color spectrum and coarse texture
- Approximate relative 2D-to-3D visual volume
- Open food taxonomy candidate matches
- Added cooking oils (e.g. 15g invisible pan oil)
- Dissolved sugars or sodium in internal marinades
- Exact scale weight measured on a digital scale
- Specific proprietary homemade recipe ratios
1. Scope of AI Usage in Unitrition
Unitrition incorporates assistive algorithmic capabilities strictly within three designated product workflows:
Optical ingredient segmentation from meal photos using GPT-4o-mini Vision.
Rapid ingredient parsing from natural text descriptions («150g chicken with rice»).
Research synthesis (Gemini Deep Research) reviewing open PubMed and USDA scientific literature.
2. Technical Limits & Volumetric Approximation
Computer vision algorithms are fundamentally constrained by optical perspective and 2D-to-3D projection. Volumetric portion estimation presumes standard reference dish geometry. Factors such as camera angle, harsh shadows, ingredient layering, and food density compression (such as purees or stews) introduce intrinsic mathematical variance.
Crucially, optical sensors cannot quantify invisible culinary components: pan-frying oils, salad dressings, dissolved sodium, or marinades. Macronutrient metrics computed from photo recognition must therefore be treated strictly as an initial editable draft, never as a laboratory assay.
3. Absolute Allergen & Cross-Contact Disclaimer
Unitrition explicitly prohibits relying on photo scanning to determine whether a food item is safe for individuals with life-threatening allergies (such as peanut or crustacean anaphylaxis) or severe autoimmune conditions (such as celiac disease). Artificial intelligence possesses zero biological capability to detect molecular-level allergens or cross-contact occurring in shared commercial fryers or packaging facilities.
Allergen verification must invariably be conducted by reviewing mandatory physical manufacturer labels (under EU Regulation 1169/2011 or FDA standards) or inquiring directly with food preparation staff.
4. Mandatory User Calibration (Human-in-the-Loop)
Unitrition’s software architecture mandates a «Human-in-the-Loop» verification step before any recognized food item is added to the Meal Builder plate or nutrition diary. The user must actively review identified candidates, adjust portion weights in grams, and confirm database matches.
The user retains full responsibility for the accuracy of logged meals. This requirement is paramount when available carbohydrate totals are converted into Bread Units (ХЕ / BU / Raciones), where uncalibrated visual estimates would distort nutritional accounting.
5. Research Methodology & Blog Literature Synthesis
Educational blog articles, methodological breakdowns, and nutrition guides on Unitrition are developed with the assistance of advanced bibliographic research tools (including Gemini Deep Research and structured Gemini Notebook analysis). These tools synthesize peer-reviewed literature and open public repositories (PubMed, USDA FoodData Central, CIQUAL ANSES, Open Food Facts, and dietary consensus statements from organizations like ADA or ESPEN).
All literature review content is published strictly for general educational self-monitoring. It does not constitute clinical nutrition therapy or replace individualized guidance from a registered dietitian.
6. EU AI Act (Regulation 2024/1689) & FTC Compliance
Pursuant to Article 50 of Regulation (EU) 2024/1689 of the European Parliament and of the Council (EU AI Act) regarding transparency obligations for deployers of AI systems, Unitrition explicitly discloses that users interact with artificial intelligence models when operating the photo scanner or natural language meal parser.
Unitrition’s software is classified under the EU AI Act as a Minimal Risk AI application. It does not conduct facial biometric recognition, surveillance, subliminal behavioral manipulation, or credit scoring. Furthermore, it strictly aligns with US Federal Trade Commission (FTC) guidelines against deceptive or exaggerated claims regarding AI capabilities.
7. Provider Architecture & Zero Data Retention
For image analysis and natural language processing, Unitrition connects to multimodal vision and language models (OpenAI GPT-4o-mini Vision / Google Gemini API) via encrypted TLS 1.3 enterprise API endpoints.
Under enterprise API data commitments from OpenAI and Google, data sent via API endpoints is NOT utilized to train or fine-tune foundation models. Once candidate food tags are returned, ephemeral payload data is purged from memory.
8. Frequently Asked Questions & Regulatory Answers
Q:Are meal photos saved or used to train OpenAI / Google AI models?
No. Unitrition operates under a strict ephemeral Zero Data Retention (ZDR) pipeline. Before leaving your browser, the photo is re-encoded on an HTML5 canvas to scrub EXIF and GPS geolocation metadata. The image is transmitted via encrypted memory solely to extract food candidates and is immediately discarded. Enterprise API terms ensure your images are NEVER used to train or refine foundation AI models.
Q:How accurate is AI portion and weight estimation from a 2D photograph?
Computer vision portion estimation is a theoretical volumetric approximation derived from 2D pixel geometry and dish perspective; it is not a calibrated kitchen scale. Internal food density, moisture content, and cooking methods naturally alter real scale weight. Users must always visually verify and calibrate portions before confirming meals.
Q:Can AI vision detect food allergens, hidden gluten, or cross-contamination?
Under no circumstances. Computer vision cannot detect molecular-level allergens (such as hidden peanut oils, soy lecithin, gluten thickeners, or fryer cross-contact). Never rely on Unitrition AI vision to assess food safety for severe allergies, anaphylaxis, or celiac disease.
Q:How does Unitrition comply with the EU AI Act (Regulation (EU) 2024/1689)?
In accordance with Article 50 of the European Union AI Act, Unitrition transparently discloses the use of generative and computer vision AI systems to assist with meal logging. The software is classified as Minimal Risk AI and explicitly does not conduct biometric facial recognition, emotional analysis, or high-risk categorizations.
Ready to test the Meal Plate Scanner?
Open the Meal Builder and experience vision-assisted logging with mandatory human calibration.