Using AI to Identify What Makes Products
Attractive to Consumers

The role of this chapter is to gather data on how products are presented online, focusing on dark patterns and framing techniques.

For quick testing and prototyping, I created 3 fake e-commerce websites that simulate various product presentations and dark patterns. You can view them here:

• Shop1
• Shop2
• Shop3

In my actual thesis, I will use the methodology from Mathur et al. (2019):

• Mathur et al. (2019) - Dark Patterns at Scale
• Official GitHub repository

• Python code

The next step is analyzing this data to identify patterns in product attractiveness. I used Orange Text Mining to run the following hierarchical clustering analysis:

This allows us to organize the set of text segments into groups of similar items so that we can scan and label them for dark pattern content (This was done in Mathur et al. (2019)).

It works on my fake shops, but it's not optimal as the product list and descriptions are small.

An extension from Mathur et al. (2019)'s work that I plan to implement is to use Computer Vision techniques to analyze product images. This could help identify visual dark patterns (e.g., misleading images, framing, borders) that text analysis alone might miss.

First I made the following workflow in Orange. It is running on Google's Inception V3 model. It doesn't detect dark patterns but it classifies products into categories (e.g., headphones, cars, etc. ) which can be interesting if we want to focus on specific products. However it is very good at clustering images by saturation which is good to detect one main dark pattern: oversaturation.

I am still looking for a model (or configuration of an existing model) that would prioritize categorization based on the text that appear on these images (as they are the main source of dark patterns).

Once we have identified and labeled dark patterns in our dataset, we can move on to the next chapter, which involves conducting behavioral experiments to assess the impact of these patterns on consumer choices.

This Chapter will allow us to measure the impact of dark patterns on purchasing decisions, but also will feed the algorithm for chapter 3.

For this Chapter, I'm still not set on the experimental design that I plan to follow.

Already we are facing huge ethical challenges if we use real products in our experiments. The ethical committee might not aprove.

This is why I'm presenting 2 different experimental designs.

They are both between-subjects randomized controlled experiments where participants are randomly assigned to either a framed product condition or a neutral product condition (no dark pattern).

To incentivize real purchasing decisions, participants given the cover story that they are evaluating a new online store interface and that their payoff depends on how close their responses are to others.

We measure willingness to pay using the Becker-DeGroot-Marschak mechanism as well as self-reported urge to buy in a similar fashion than Sin, Harris, Nilsson & Beck (2025).

1 dark pattern tested is 1 condition. If we test n patterns, each pattern can either be present or absent, yielding 2^n conditions in a full factorial design. (So we can only experiment with the main dark patterns)

• Experiment 1 uses realistic mockups of online stores.
• Experiment 2 only retains textual information.

The experiment will yield data on how different dark patterns influence willingness to pay and urge to buy, which will be used in Chapter 3 to train predictive models.