Episode Details

Send us Fan Mail

In this episode, we speak with Arno De Caigny (IESEG School of Management) to discuss his article, "Is Similarity in Advertising a Double-Edged Sword? The Role of AI-Generated Faces," coauthored with Gudrun Roose (Ghent University), published in the June 2026 issue of JAR. Arno explains the question driving the paper: decades of research show people respond better to ads featuring faces that resemble their own, and generative AI now makes it cheap to push that resemblance as far as an advertiser wants, so is more similarity always better? 

The team first built an automated facial similarity measure by adapting the same face recognition technology that unlocks your phone, converting faces into embeddings and scoring resemblance with a cosine distance, then validating the algorithm against human annotators. Across two studies pairing participants' social media profile photos with AI-generated spokespersons in car insurance ads, the relationship between facial similarity and purchase intention traced an inverted U. Low to medium similarity helps, consistent with the familiar similar-to-me effect, but past an optimal point around 57% the effect reverses and purchase intentions fall, which the authors call the too-similar-to-me effect.

The practical takeaway for advertisers is to resist the pull toward near-perfect matches and aim instead for the 50% to 70% sweet spot, and Arno adds a warning worth hearing: because generative AI makes it cheap to spin up millions of ad variants, a black-box optimization algorithm can drift into high similarity without anyone deciding it should, so similarity is worth tracking, and the paper's measure runs entirely on open source tools. Context matters too, since hyper-personalization carries more risk in sensitive categories like financial services and healthcare, while transparency and consent can turn the same technology into something consumers actually value, as with virtual try-on experiences. Our conversation also digs into the project's origins in the early wave of deepfakes, the data collection headaches that come with profile pictures as personally identifiable information (Prolific and MTurk were off the table, hence student samples and careful ethics review), and future directions like explaining the mechanism behind the effect and pinning down which dimensions of similarity, from eyes to skin tone to build, actually drive it.

Read the full paper here: https://doi.org/10.1080/00218499.2025.2586185

To keep up to date on the latest JAR news sign up for our newsletter:
https://lp.constantcontactpages.com/su/mtD04QN

And watch this and more content on our YouTube page:
   / @journalofadvertisingresearch 

And follow us on LinkedIn:

linkedin.com/company/journal-of-advertising-research/