Looking for a fun way to pass the time? How about a Turing Test that shows how close language AIs and humans have come?
Human-machine tests have long been an integral part of the Internet. Captchas, for example: by performing small tasks, usually recognizing shapes or images, humans have to prove their humanity from time to time in order to use a service.
But how long can this concept withstand technological advances? Already, for example, AI-formatted homework is fooling teachers without students having to seriously use their own brains.
Human or not? Only two minutes to decide
A new website called humanornot.ai is a new twist on whether humans can still distinguish AI from other humans. Users are connected either to a large speech model or to other users via text chat. After two minutes, they have to decide: Was it a human or a neural network on the other end? This not only gives you a chance to test your skills as a Turing tester, but also puts you in the role of being tested - can you convince the other side that you are not a machine?
The idea for "Human or Not?" came from Amos Meron, a product manager at AI21 Labs. The Israeli AI startup offers the GPT alternative Jurassic-2.
The model joins OpenAI's GPT-4, Anthropic's Claude, and other undefined language models in the chat. To prevent the bots from giving themselves away too quickly, typos or "nonsensical humor" have been integrated into the response generation.
Data collection for instruction tuning?
"The game is free to play, and its goal is just to help all of us better understand the nature of these AI models in the context of not just a productivity tool, but members of our online world," Meron said. HumanOrNot is part of a research project being conducted by AI21 Labs.
But the data generated by the game could also be valuable to AI21 Labs for instruction-tuning its own language models. According to the terms of use, users agree to make their data available to AI21 Labs for "developing new technologies and services."
