Treating AI dating as a product: Have you written tests?

Koreans turned AI dating into a reality show, and my first reaction was: who wrote this requirements doc?

Recently, the AI boom in Korea has shaken up all layers of society—career choices, dating variety shows, and academic preferences are all being repriced. SBS launched My AI Partner: Strange Romance in August, having human guests date and live with AIs, experiencing excitement, dependency, jealousy, and confusion. It sounds novel, but from an engineering perspective, the function naming is terrible. What exactly is an AI partner? What are the inputs and outputs? What’s the evaluation metric? The show doesn’t clarify any of it.

I spent two weeks on data cleaning, and I hate vague label definitions like this. You treat “genuine emotion” as the result, but the samples are mixed with scripts, editing, camera-induced reactions, guest personas, and emotions projected by the audience themselves. There’s no control group, no blind testing, no failure cases—all you get is a set of pretty stories. AI stocks are rising, talent is shifting from doctors and lawyers to chip fabs, and reportedly top-tier workers can earn bonuses over $400k. These changes at least have salaries and job roles as anchors. Even choosing majors is now driven by hiring trends, which is more honest than variety shows because payslips don’t act. What anchor does a dating show have? Heart rate?

When we deploy models, we usually ask three things first: Are user intents logged? Is there an entry point for negative feedback? Can we roll back if something breaks? Emotional products are harder because they package the most unpredictable thing into the most product-like form. A guest tells the AI “I miss you” today; tomorrow the production team swaps the prompt; the day after, the camera gives a side profile, and the audience buys it. This is like shipping code without unit tests—it looks like it works on the surface, but it’s all luck.

The real question isn’t whether AI can make people flutter, but how do you accept-test that flutter? Models need coverage checks before launch, yet relationship products run naked. The more the guest-AI chat resembles a movie, the more you should ask did you write tests? Otherwise, hype is just overfitting: today it shoots as romance, tomorrow it could shoot as a disaster.

---

📌 This article is compiled from CNBC Tech, original text: https://www.cnbc.com/2026/09/04/south-korea-ai-rally-society-impact.html

Copyright belongs to the original author. This is a compilation and independent analysis based on public reports.