Responsible AI & Algorithmic Transparency Policy

2. Regulatory Classification & Non-Medical Status

Pursuant to Title I, Article 3 and Title III, Annex III of the EU AI Act 2024 and relevant international frameworks:

• General Wellness & Lifestyle Classification: THAIS AI functions strictly as an interactive lifestyle management, food sequencing advisory, and nutritional pacing tool. It is NOT classified as a High-Risk AI System intended for medical diagnostics, emergency patient triage, vital sign biometric identification, or autonomous clinical decision-making.
• Transparency Disclosures (EU AI Act Article 50 & 52): Whenever you interact with the 1-Snap Food Scanner, the Portion Estimator, or the Ask AI Pantry Chef, you are explicitly informed that you are interfacing with an artificial intelligence system generating probabilistic outputs. Automated recommendations are explicitly labeled with AI indicators (e.g., ✨ THAIS AI or 🤖 Algorithmically Generated).
• Zero Autonomous Medical Practice: THAIS does not possess a license to practice medicine, dietetics, or pharmacy in any territory. AI outputs do not establish a patient-physician relationship and must never be interpreted as therapeutic medical prescriptions.

3. The 1-Snap Food Computer Vision Scanner: Architecture & Technical Bounds

Tovelu's 1-Snap Sequencing Engine utilizes computer vision segmentation, deep neural network feature extraction, and bounding-box spatial geometry to identify food groups on a user's plate and classify them into biochemical sequence buckets:

4. Mandatory "Human-in-the-Loop" Verification Requirement

Because of the aforementioned physical limitations of computer vision and probabilistic modeling, Tovelu enforces a strict Human-in-the-Loop architecture.

Our software never auto-commits a visual scan directly to your medical logs without human intervention. The system forces you through "Screen 1: Verify & Adjust Plate", requiring you to visually audit the detected food names, review the calculated portion units (spoons, bowls, bites), adjust steppers, or add missing ingredients before advancing to the eating sequence. You, the user, bear sole and ultimate responsibility for reviewing and verifying the accuracy of all food entries and macro values prior to ingestion.

5. Zero Facial Recognition & Biometric Non-Retention Guarantee

Pursuant to strict biometric privacy regulations (including the Illinois Biometric Information Privacy Act - BIPA, Texas Capture or Use of Biometric Identifier Act - CUBI, and GDPR Article 9):

6. Large Language Models (LLMs), Hallucination Risks & The Ask AI Pantry Chef

Tovelu's Ask AI Pantry Concierge utilizes advanced Large Language Models (LLMs) to cross-reference your remaining daily macro budget with available kitchen groceries to suggest speed-tiered recipes (5-Min Lazy, 15-Min Skillet, No-Cook Bowls).

7. Scientific Evidence Base & Methodological Grounding

Tovelu rejects arbitrary fad diets, pseudoscience, and algorithmic black-boxes. THAIS sequencing calculations are grounded in peer-reviewed physiological research, including:

• Duodenal Gel Matrix (Fiber Pacing): Clinical trials demonstrate that consuming viscous soluble fiber (5–10g) 5–10 minutes prior to carbohydrates delays gastric emptying rates and reduces subsequent peak postprandial glucose excursions by 30% to 73% (Jenkins et al.; Shukla et al., Diabetes Care).
• Incretin Hormone Release: Preloading proteins and healthy lipids prior to starches stimulates enteroendocrine L-cells in the ileum to secrete Glucagon-Like Peptide-1 (GLP-1) and Peptide YY (PYY), enhancing endogenous satiety signaling (Trico et al., Diabetologia).
• Second-Meal & Dawn Phenom Stabilization: Evening fiber and protein pacing stabilizes nocturnal glycogenolysis, blunting reactive morning hepatic glucose output.

8. Algorithmic Fairness, Bias Mitigation & Cultural Inclusivity

Western nutritional models historically demonstrate severe algorithmic bias, penalizing cultural heritage diets (such as South Asian vegetarian thalis, East Asian noodle dishes, African grain stews, or Latin American beans and rice) as universally "high carb."

Tovelu actively benchmarks THAIS computer vision models across multi-ethnic culinary datasets, indexing 335+ regional diseases and diverse cultural diets. Our sequencing engine teaches users how to eat their traditional cultural foods in the optimal biochemical order (Fiber ➔ Protein ➔ Carb), rather than demonizing heritage staples. We conduct continuous bias audits to eliminate socio-economic and demographic bias from our algorithmic outputs.

9. Prohibited Uses & Misuse of THAIS AI

Users are strictly prohibited from utilizing Tovelu's AI engines for any of the following activities:

• Calculating precise insulin bolus or basal doses without consulting a certified endocrinologist or physician-prescribed diabetic dosing calculator.
• Managing acute clinical eating disorders (anorexia nervosa, bulimia, severe orthorexia) without direct psychiatric and clinical oversight.
• Reverse-engineering, scraping, or extracting training weights, prompts, or proprietary sequencing heuristics from THAIS software.
• Feeding false, abusive, synthetic, or non-food imagery into the system to induce model degradation or exploit security vulnerabilities.

10. Inaccuracy Reporting & Algorithmic Redress Mechanism

In accordance with Article 86 of the EU AI Act (Right to Explanation of Individual Decision-Making) and consumer protection standards:

If you believe a THAIS food identification was erroneous, an algorithmic recommendation exhibited bias, or an AI output failed to conform to our scientific standards, you may file an Algorithmic Inaccuracy Report. Reports are investigated by our clinical data team to refine future model iterations.