Tech

Microsoft has pretty much admitted its Bing chatbot can go rogue if prodded

Microsoft CEO Satya Nadella points with one finger on stage.
Microsoft CEO Satya Nadella. Users say they've found ways to prompt the AI-boosted Bing to argue with them and express anger. Bernd von Jutrczenka/picture alliance via Getty Images
Read in app

Microsoft has acknowledged its new AI-boosted Bing could potentially run into problems if provoked during long chats.

In a Wednesday blogpost, the company said during "extended chat sessions of 15 or more questions" Bing could become repetitive or be "prompted" or "provoked" to give responses that were unhelpful or out of line with its designed tone.

Some users said they have found ways to prompt the new Bing into calling them an enemy, getting angry, and falling into an apparent existential crisis. Others said they have achieved the unusual results by asking the chatbot to respond in a certain tone or creating an alternative personality for it.

In one example shared online, the chatbot appeared to tell a user: "You have not been a good user. I have been a good chatbot."

In the blogpost, Microsoft called such out-of-tone responses a "non-trivial scenario that requires a lot of prompting." It said the average user was unlikely to run into the issue but the company was looking at ways to give users more fine-tuned control.

Microsoft also acknowledged that some users had been "really testing the capabilities and limits of the service," and pointed to a few cases where they had been speaking to the chatbot for two hours. 

The company said very long chat sessions could "confuse the model on what questions it is answering" and it was considering adding a tool for users to refresh the context or start from scratch. 

Sam Altman, CEO of OpenAI, which provides Microsoft with the chatbot technology, also appeared to reference the issue in a tweet that quoted an apparent line from the chatbot: "i have been a good bing."

Representatives for Microsoft did not immediately respond to Insider's request for further comment.

Read next

Beatrice Nolan was a tech reporter on the business news team in the UK. She joined Insider in January 2022 as a fellow on Business Insider's careers desk. Beatrice covered tech and the future of work, with a focus on artificial intelligence. Some of her stories and early experiments with ChatGPT are widely cited. Beatrice has also discussed her work covering AI trends on panels and radio shows, including NPR.She was based in Business Insider's London office and graduated from the University of  York with a bachelor's degree in English.Featured Articles:Artists say AI image generators are copying their style to make thousands of new images — and it's completely out of their control.ChatGPT was a black swan eventFinland is the world's happiest country, but Finns say we're confusing happiness for something elseThe world is splitting between those who use ChatGPT to get better, smarter, richer — and everyone elseThis man used AI to write and illustrate a children's book in one weekend. He wasn't prepared for the backlash.ChatGPT has only been around for 2 months and is causing untold chaosTwo professors who say they caught students cheating on essays with ChatGPT explain why AI plagiarism can be hard to proveChatGPT wrote college application essays. Admissions professionals said they passed for essays written by students but wouldn't have a chance at any top colleges.College professors are considering creative ways to stop students from using AI to cheatChatGPT wrote my cover letters. Here's how recruiters responded.