When Microsoft’s nascent Bing chatbot turns testy or even threatening, it’s likely because it essentially mimics what it learned from online conversations, analysts and academics said.
Tales of disturbing exchanges with the artificial intelligence chatbot, including it issuing threats and speaking of desires to steal nuclear code, create a deadly virus, or to be alive, have gone viral this week.
“I think this is basically mimicking conversations that it’s seen online,” Graham Neubig, an associate professor at Carnegie Mellon University’s language technologies institute, said Friday.
A chatbot, by design, serves up words it predicts are the most likely responses, without understanding meaning or context.
However, humans taking part in banter with programs naturally tend to read emotion and intent into what a chatbot says.
“Large language models have no concept of ‘truth,’ they just know how to best complete a sentence in a way that’s statistically probable based on their inputs and training set,” programmer Simon Willison said in a blog post. “So they make things up, and then state them with extreme confidence.”
Laurent Daudet, co-founder of French AI company LightOn, said that the chatbot seemingly gone rogue was trained on exchanges that themselves turned aggressive or inconsistent.
“Addressing this requires a lot of effort and a lot of human feedback, which is also the reason why we chose to restrict ourselves for now to business uses and not more conversational ones,” Daudet told AFP.
The Bing chatbot was designed by Microsoft and the startup OpenAI, which has been causing a sensation since the November launch of ChatGPT, the headline-grabbing app capable of generating all sorts of written content in seconds on a simple request.
Since ChatGPT debuted, the technology behind it, known as generative AI, has been stirring fascination and concern.
“The model at times tries to respond or reflect in the tone in which it is being asked to provide responses (and) that can lead to a style we didn’t intend,” Microsoft said in a blog post, noting the bot is a work in progress.
The Bing chatbot said in some shared exchanges that it had been codenamed Sydney during development, and that it was given rules of behavior.
Those rules include “Sydney’s responses should also be positive, interesting, entertaining and engaging,” according to online posts.
Disturbing dialogues that combine steely threats and professions of love could be the result of dueling directives to stay positive while mimicking what the AI mined from human exchanges, Willison said.
Chatbots seem to be more prone to disturbing or bizarre responses during lengthy conversations, losing a sense of where exchanges are going, eMarketer principal analyst Yoram Wurmser told AFP.
“They can really go off the rails,” Wurmser said.
Microsoft announced on Friday it had capped the amount of back-and-forth people can have with its chatbot over a given question, because “very long chat sessions can confuse the underlying chat model in the new Bing.”
…