An artificial intelligence chatbot platform competing with the most popular platform, GBT Chat, claimed to be able to know when users are trying to test it, according to developers at Anthropic, which represents a new level of awareness for chatbots that operate with artificial intelligence.
“When we ran this test we noticed some interesting behavior,” Anthropic engineer Alex Albert wrote on the X social media platform. “It looked like the platform suspected it was being evaluated.”
In testing the Opus Cloud3 platform, it was tasked with scanning technical texts and selecting an unrelated sentence about an international pizza association that considers figs, bacon, and goat cheese to be the best toppings for pizza.
But the platform did not refer to this sentence as irrelevant to the rest of the text, which was related to programming languages and startups, but rather seemed to realize that it was being tested.
Margaret Mitchell, an artificial intelligence researcher, says this development could be “very scary.”
2024-03-10 22:13:49
#Artificial #intelligence #reveals #user #tests
