Following up on my post about inflated claims in skincare, I wanted to explore the possibility of building a “plain claim(s)” model. This model would take a skincare product’s ingredient list and provide a breakdown of key ingredients, along with realistic expectations for the product’s effects. The goal is to enable consumers to cut through the hype and get a more objective perspective on their skincare products, rather than relying solely on marketing claims.

You can try out the model here: link

Building the Model

Initial Testing with AI Models

I began by conducting prompt tests with ChatGPT and Claude. The process involved:

1. Selecting an ingredient list and product description from popular retailers like Sephora or ULTA.
2. Asking ChatGPT to generate a description based only on the ingredient list.
3. Having Claude judge which statement (human-written or AI-generated) better reflected the ingredients’ functions.
4. Reversing the process, with Claude generating the statement and ChatGPT judging.

Interestingly, the AI-generated statements were often deemed more accurate. This could be due to various factors, including:

• My prompting technique
• Potential bias in interpreting the judgments
• The possibility that AI models have similar ways of expressing information due to correlated latent spaces (embeddings)

Enhancing Output with Structured Input

To improve the model’s ability to capture the full range of a product’s functions, I incorporated structured input based on our work at Revea. We’ve mapped the skin benefits of cosmetic ingredients and built a model that predicts skincare benefits across six parameters:

1. Skin hydration and barrier function
2. Skin texture
3. Skin tone
4. Redness / inflammation
5. Vitality
6. Firmness

This machine learning model maps individual ingredients to their skin benefits, applies an exponential decay function based on the ingredient’s position in the list, and produces a score across the six parameters. We normalize these scores using a large dataset of real product ingredient lists.

By feeding the AI model with both the calculated product scores and an optimized prompt, we achieved better results.

Choosing the Right AI Model

After testing Claude vs. ChatGPT outputs across various products, I found Claude’s writing style more suitable for our purposes. But this is solely based on my personal opinion and no “proper” consumer testing was done here.

Implementation and Production

With designs and infrastructure already in place from our ingredient scanner and exploration tools, implementation was relatively straightforward:

1. Frontend: We updated copy, modified a text box, and added rendering for the output.
2. Backend: We created a small cloud function on Google Cloud Platform (GCP) that:
   * Formats the ingredient list
   * Calculates product scores using the ML model
   * Calls Claude with the scores, optimized prompt, and ingredient list
   * Parses Claude’s output into a structured format for easy rendering

The entire process, from concept to deployment, took approximately three days: Two days for prompting LLMs and testing the basic idea. One day for coding, implementation, and basic testing before deployment

Consumer Value

In a world of increasingly bold marketing claims, I believe in providing a more balanced and nuanced perspective. The Plain Claim Model aims to offer an easily accessible and digestible alternative for understanding skincare products. By cutting through exaggerated marketing language, we hope to empower consumers to make more informed decisions about their skincare routines.