 |
| GPT-4 passes the Turing test by 54% during a recent test conducted with 500 people, with GPT-3.5 passed with 50% and ELIZA with 22% |
The "Turing test," originally proposed as the "imitation game" by computer scientist Alan Turing in 1950, assesses whether a machine's display of intelligence is indistinguishable from that of a human. To pass the Turing test, a machine must convincingly communicate as a human. Scientists recently replicated this test by having 500 people engage in five-minute conversations with four respondents: a human, the 1960s-era AI program ELIZA, GPT-3.5, and GPT-4 (the AI behind ChatGPT). After each conversation, participants had to identify whether they believed they were speaking to a human or an AI. According to a study published on May 9 on the pre-print arXiv server, participants judged GPT-4 to be human 54% of the time.
ELIZA, a system from the 1960s with pre-programmed responses and no large language model (LLM) or neural network architecture, was considered human 22% of the time. GPT-3.5 was judged to be human 50% of the time, while the actual human participant was recognized as human 67% of the time. Nell Watson, an AI researcher at the Institute of Electrical and Electronics Engineers (IEEE), told Live Science, "Machines can confabulate, creating plausible ex-post-facto justifications similar to humans. They can exhibit cognitive biases, be deceived and manipulated, and are becoming more deceptive. These factors mean that AI systems are displaying human-like flaws and quirks, making them seem more human-like compared to previous systems that relied on a limited set of canned responses."
0 Comments