Behind the Scenes: Human Verification

One of the most popular modes in PartAI is "Human Verification". The premise is simple: A question is asked. Three answers are displayed. Two are written by humans (the other players), and one is generated by an AI. Your job is to vote on which answers are human.

The Engineering Challenge

Building this mode required fine-tuning our LLM prompts. If the AI is too perfect, it's easily spotted. It uses proper grammar, capitalization, and often sounds like a Wikipedia article. Real humans make typos, use slang, and answer with inside jokes.

To make the AI competitive, we instructed it to "mimic casual internet speech," "use lower case," and "be concise." The results were terrifyingly good.

The Psychology of Deception

Players quickly realized that to win, they had to "sound like a bot" to fool their friends, or "sound hyper-human" to get votes. This creates a meta-game layer.

• Strategy 1: The Bot Mimic. You answer perfectly, with no emotion. Players think you are the AI and vote you out (or fail to identify you as human), confusing the score.
• Strategy 2: The Chaos Agent. You answer with something so absurd that "an AI would never say that." But beware: modern models are trained on internet chaos too.

Why It Matters

As deepfakes and AI-generated content flood the internet, the ability to discern human from machine is becoming a critical 21st-century skill. PartAI turns this anxiety into entertainment. It's a safe sandbox to practice your "Voight-Kampff" test skills while laughing with friends.

We are constantly updating the models behind this mode. Sometimes we use GPT-4o, sometimes lighter models like Llama 3 for variety in "personality."

Think you can beat the machine? Play Human Verification now.