UC San Diego study finds GPT-4.5 can convincingly impersonate humans
Turns out, AI is getting seriously good at pretending to be human.
Researchers at UC San Diego tested advanced language models like GPT-4.5 and found they could fool people in the classic Turing test, where machines try to mimic human conversation.
GPT-4.5 nailed it by copying things like humor and tone when given specific prompts, making it tough for participants to tell if they were chatting with a person or a bot.
GPT-4.5 fooled 73% of participants
In the study, folks interacted with both humans and four different AIs (including GPT-4.5 and LLaMa-3.1-405B) and then guessed who was actually human.
GPT-4.5 tricked people 73% of the time, while LLaMa-3.1-405B managed 56%. Older models like ELIZA only fooled 23%.
The researchers say giving clear instructions helped AIs sound more real, but this raises big questions about trust online.
Co-author Ben Bergen pointed out that now machines can convincingly pass as human; risks like manipulation and fraud are more possible than ever.