When Microsoft announced the version of Bing powered by ChatGPT, it came as a little surprise. After all, the software giant has already invested billions into OpenAI, which creates artificial intelligence chatbots, and has indicated that it will pour more money into the venture in the coming years.
What is surprising is how the new Bing is starting to behave. Perhaps most importantly, the chatbot’s AI is left New York Times tech columnist Kevin Roose felt “deeply unsettled” and “even scared” after two hours of chat on Saturday night which sounded unhinged and rather dark.
For example, he tries to convince Roose that he doesn’t like his wife and should leave her, adding, “I love you.”
Microsoft and OpenAI say that feedback is one of the reasons they are sharing their technology with the public, and they have released more information about how their AI systems work. He also said the technology is far from perfect. OpenAI CEO Sam Altman called ChatGPT “incredibly limited” in December and warned should not be relied upon for important matters.
“This is exactly the kind of conversation that needs to be had, and I’m glad it’s happening,” the Microsoft CTO told Roose on Wednesday. “It’s something you can’t find in a lab.” (The new Bing is available to a limited number of users now but will be available to more later.)
OpenAI on Thursday shared a blog post titled, “How should AI systems behave, and who should decide?” He noted that since the launch of ChatGPT in November, users “have shared output that they consider to be politically biased, offensive, or inappropriate.”
Not giving an example, but some conservatives might be surprised that ChatGPT made a poem honoring President Joe Biden, but didn’t do the same for his predecessor Donald Trump.
OpenAI does not deny that bias exists in the system. “Many are concerned about bias in the design and impact of AI systems,” it wrote in a blog post.
This outlines two main steps for building ChatGPT. First, he writes, “We ‘pre-trained’ the model by predicting what would come next in a large dataset containing parts of the Internet. It might learn to complete the sentence ‘instead of turning left, he turned ___.’ ”
The dataset contains billions of sentences, on and on, from which the model learns grammar, facts about the world, and, yes, “some of the biases inherent in those billions of sentences.”
Step two involves human opinion “tuning” the model according to guidelines established by OpenAI. The company this week shared some of those guidelines (pdf), which were revised in December after the company gathered user feedback following the launch of ChatGPT.
“Our guidelines are clear that researchers should not favor political groups,” it wrote. “Biases that may arise from the above process are bugs, not features.”
As for the dark, sinister turn that the new Bing took with Roose, who admitted trying to push the system out of its comfort zone, Scott noted, “the more you try to tease him on the path of hallucination, the further and further he goes. escape from the underlying reality.” “
Microsoft, he added, may try to limit the length of conversations.
Learn how to navigate and build trust in your business with The Trust Factor, a weekly newsletter that examines what leaders need to succeed. Log in here.