Can AI Explain Your Moisturizer? Ours Can, Almost…

Written By Glow Getter

Overview:

This post continues the work on my ingredient interpreter project, this time focusing on prototyping behavior using ChatGPT ahead of model training. Read the previous post here.

Since then, we’ve focused on:

  • Behavioral Prototyping with ChatGPT: Simulated how our ingredient chatbot might respond using the same structured data we’ll use in our local Ollama model. This helps shape tone, accuracy, and fallback handling early.

  • Cross-Product Dataset: Pulled ingredient data from five globally sourced skincare products, each selected to reflect different formulation philosophies (e.g., biotech, sensitive skin, reef-safe). The ingredient data is sourced from INCIDecoder to align with our tone and educational goals.

  • UX-First Prompt Testing: Designed and tested prompts around sensitive skin, actives, and multi-functional ingredients to uncover how the chatbot handles ambiguity, risk, and scientific nuance.

 
 

This post builds on the original project overview. You can read the full context here. → Original case study

 
 

Step 1: Choosing 5 Products from 5 Countries

Before building a fully fine-tuned AI model, we’re starting with a lightweight MVP designed to test tone, behavior, and user experience. The idea is simple: train ChatGPT on the same dataset we’ll eventually use for our local prototype, so we can simulate how the interpreter might respond, and refine the system before scaling.

We selected five skincare products that represent different formulation philosophies around the world.

  1. 🇰🇷 Korea – No. 3 Skin Softening Serum
    → Ferments, humectants, and biotech actives

  2. 🇯🇵 Japan – Biore UV Aqua Rich Watery Essence
    → Lightweight SPF with skin-soothing agents

  3. 🇺🇸 United States – CeraVe Resurfacing Retinol Serum
    → Potent actives with marketing-led language

  4. 🇪🇺 Europe – Bioderma Sensibio H2O Micellar Water
    → EU-regulated, fragrance-free formulation for sensitive skin

  5. 🇦🇺 Australia – Blue Lizard Face Mineral Sunscreen SPF 30+
    → Reef-safe, mineral-based sunscreen with natural-leaning claims

This spread lets us test how the chatbot responds to different regulatory assumptions, ingredient styles, and product purposes.

 

Step 2: Ingredient Data from INCIDecoder

We’re pulling ingredient profiles directly from INCIDecoder because its tone aligns with our goals:

  • Neutral, science-first, and accessible

  • Avoids clean beauty fear tactics

  • Offers structured insights: functions, aliases, safety notes

We’re building a dataset of ~80 ingredients, each tagged with:

  • INCI name and aliases

  • Functional roles (e.g. humectant, preservative)

  • Controversies or common user concerns

  • Suitability for sensitive skin or acne-prone users

This dataset is used consistently across all stages of development.

 

Step 3: Training ChatGPT as a Behavioral Prototype

To simulate how our chatbot might eventually behave, we’re training ChatGPT on the same structured dataset that we’ll use for our final model. This allows us to:

  • Fine-tune tone and behavior early

  • Explore prompt logic and response boundaries

  • Identify inconsistencies before committing to local deployment

We’re running prompt tests such as:

  • “What do the key ingredients in CeraVe’s Retinol Serum do?”

  • “Is Ethylhexylglycerin suitable for sensitive skin?”

  • “What’s the role of Retinol vs. Niacinamide in this formula?”

This phase gives us a high-fidelity preview of how the model will perform — and lets us troubleshoot misunderstandings, edge cases, and UX challenges in advance.

 

UX + Error Handling

We're designing for a user experience that’s:

  • Informative but not overwhelming

  • Science-backed but easy to read

  • Transparent about limitations (e.g., “We don’t have data on this ingredient yet”)

We’re also testing fallback responses for:

  • Unsupported questions

  • Unknown ingredients

  • Debated safety claims

 

Phases:

Phase 1: Testing + Iteration (Current Phase)

  • Integrate with LLM or retrieval-augmented model

  • Conduct usability testing on language clarity and risk phrasing

  • Align outputs with accessibility and enterprise design systems

Phase 2: Testing + Iteration

  • Begin user testing with real ingredient lists

  • Test upload and paste workflows

  • Build a trust-feedback loop to guide future improvements

Future Plans

  • Layer in user profiles (e.g., acne-prone, sensitive skin, rosacea)

  • Enable “smart” questions (e.g., “Is this pregnancy safe?”)

  • Develop a Chrome extension for ingredient popovers while brows

Glow Getter
Next

The AI Brain Behind Ingredient Transparency