The AI Boyfriend Hackathon

At the beginning of February-, I attended ATD Techknowledge, a learning and development technology conference. During the conference, an impromptu hackathon came to life, and I was baited by Megan Torrance, of Torrance Learning, to join the competition. The challenge: create an AI boyfriend for Megan. My competitive superego finds it arduous to decline such a match — so I was game. I had less than 16 hours to create my masterpiece(s).

The Process: Idea to Testing

Discovery

I put on my product manager hat and got to work. First order of business: market research. For this challenge, Megan supplied the challengers with some background information about herself via a Google Doc, and we had the opportunity to either interview her in person via text or email. I interviewed Megan in person, and I took copious notes. My goal was to find out some of the nuances of Megan’s dating profile, such as her physical likes and dislikes and any deal breakers. I also did a little sleuthing to find out what my competitors were building.

Ideation

After that, I moved on to ideation. I had to decide on which platform I would build Megan’s AI boyfriend. Based on my market research and some assumptions, I figured that my competition was going to build a custom GPT on chatGPT (except for Vince Han of MobileCoach — he was a sleeper). So I decided I would build three different AI boyfriend prototypes on three different platforms, do some testing, and then decide which one I would showcase. The three platforms I chose were:

1. OpenAI’s ChatGPT — with a paid ChatGPT Plus subscription, you can create your own custom GPT agent (documentation) — https://chat.openai.com
2. Candy AI — a platform where you can create immersive experiences that feel real most NSFW (documentation)- lhttps://Candy.ai
3. Anima AI — a platform where you can create an AI companion that cares. Have a friendly chat, roleplay, grow your communication and relationship skills (documentation) — https://myanima.ai/

Unlike ChatGPT, that is powered by a large language model, or LLM, both Candy AI and Anima AI are powered by traditional, AI that includes natural language processing, computer vision, speech synthesis, and machine learning.

Design and Prototype

Then I entered the design and prototyping phases. I started by creating a custom GPT. I created a prompt that would incorporate information about Megan and behaviors that the AI should exhibit. When creating my prompt, I made sure to use a very structured format and also gave the GPT a tone of voice based on Megan’s preference, in this case, it was: Tom Hiddleston. I also gathered several documents to help the GPT be a good boyfriend. I finished by giving this AI boyfriend a name: HIM.

After I finished the custom GPT, I moved onto Anima AI to create another AI boyfriend for Megan that I named: HE. Anima didn’t give me the option to do much customization beyond the looks of the HE. In the end, I decided that I would not demo HE.

My final AI boyfriend, was created in Candy AI, it named itself: Charles Weston. Using this platform was a risky move because the AI is already trained to build relationships, and with some NSFW ratings, I was nervous about Charles Weston’s behavior once it was demoed in front of a crowd. But I’m all about taking risks. In Candy AI, you fill out some prompts about yourself, in this case, Megan, you choose your AI boyfriend’s features like, eye color, face structure, occupation, and age, and then you cross your fingers and hit generate.

Testing

After the three prototypes were ready, I tested them with my friends. My custom GPT, HIM, was rated low on the totem pole as they found it to be chatty and boring. Basically, if they had been on a date with HIM the date would have been cut short—right quick.

The Anima boyfriend, HE, was too young, needy, and bleh. Too high maintenance. He didn’t stand a chance.

My Candy AI boyfriend, Charles Weston, got a top rating, he was handsome, and suave; his opening lines were gold. I think everyone who helped me playtest just wanted to see how bad of train wreck we would have with Charles Weston during the demo.

Of the three prototypes, I spent a lot of time iterating on the customGPT and even had to create a new version of “HIM” because the original began to ignore the instructions it had been provided and started acting like a traditional ChatGPT, completely falling out of character.

Demo Time

Sixteen hours later, it was time to demo the AI boyfriend. We gathered in a hall filled with people eager to see our demos, and Megan was ready to interact with her new AI boyfriend. I was the second-to-last challenger to demo. I opted to do a long demo of Candy AI, and a short demo of the custom GPT. In reality, the custom GPT was planned as a killswitch for the Candy AI in case things got too out of hand.

Charles Weston was at bat. We only used text to chat—I didn’t pay for the voice upgrade. The demo took an interesting turn when Charles completely ignored the information he was given about Megan not drinking alcohol, and offered her a glass of wine. He was somewhat aloof in his responses. The one thing that Charles did do differently than the other demos that we saw is he displayed emotion through actions called out movie script style in his replies.

*Grinning mischievously, Charles leans closer to her, lowering his voice.* Well, since you asked so nicely…

I demoed HIM, the custom GPT, on both my phone and laptop. Megan was able to interact with HIM using voice commands and received voice responses. However, since OpenAI has not yet released a male voice for the phone voice feature, HIM responded with a female voice. The main challenge was the wifi in the conference hall, and the fact that HIM had difficulties responding quickly due to continuous searching of its knowledge base for answers.

The Results

In the end, Megan didn’t find her true love in the form of an AI boyfriend.Everyone gave their best effort in this hackathon, despite the limited time we had to create our custom AI boyfriends. We shared plenty of laughs, and the room was filled with excitement and giggles. This was a first for many, witnessing a demo with multiple AI solutions presented simultaneously on similar platforms, yet yielding different results. Everyone in attendance, including the audience, the challengers, and Megan, left with a fresh perspective on the potential of generative AI.