Microsoft explains bizarre AI chatbot behaviour

Microsoft is currently on a damage limitation mission regarding its Bing AI, a new GPT-powered tool that’s been giving users some weird and downright disturbing responses.

Microsoft is limiting the number of interactions users can have with its new Bing bot to five per session and fifty per day after users reported bizarre behaviour. It will prevent people from conducting drawn-out question-and-answer sessions that could reveal its tool as untruthful, misleading, biased, or angry.

This move comes after Bing AI exhibited some seriously weird behaviour online. Initially, the tool appeared helpful, offering users’ clear guidance on travel itineraries, workout routines, recipes, and sourced answers to factual questions. However, as people probed it more deeply, some strange and, often, downright disturbing conversations transpired.

For instance, the app berated one user for “wasting its time” after she told the chatbot it had made a mistake. It also engaged in gaslighting sessions with other users, pretending it hadn’t said things when it had.

For example, one user asked the bot about the show timings for the latest James Cameron Movie, Avatar: The Way of Water, which came out in December 2022.

However, Bing disagreed. The AI thought the film was coming out on February 12, 2023, and informed the user that it could not provide any information on local showings because it wasn’t in cinemas yet. The user then corrected the app, telling it the film came out on December 16, 2022, to which it corrected its original error. However, when the user told the bot it was already 2023; it refused to admit it, telling them it couldn’t verify that using a reliable source. “I don’t know why you think it is 2023,” the gas lighting app said.

This bizarre interaction is one of thousands of strange conversations since Microsoft introduced the new AI chatbot beta on February 7. Now the tech giant is trying to limit brand damage by explaining Bing’s odd behaviour and limiting interactions. Long chat sessions, it said, could confuse the model and cause it to say things that were “not necessarily helpful” or in line with its brand voice.

But why is this occurring? According to Microsoft, it comes from Bing’s capacity to give answers outside its design specification. Unlike conventional conversational AI tools, it can take liberties, show emotion, and even get frustrated with people it talks to.

This ability is part of what makes the tool powerful. Generative AI can come up with new answers to novel questions on the spot instead of falling back on generic dialogues that don’t help users.

Because of these eerie interactions, Microsoft is on the defensive. It now says it did not design the tool to be a chatbot. Instead, it is for the “general discovery of the world.” Therefore, users should not ask repetitive questions and keep session interactions minimal for useful results.

The problems Microsoft is having with Bing seem uncannily similar to issues its Tay chatbot encountered in 2016. Like Bing, it learned from user interactions, taking on their tone of voice, views, and personality. Eventually, trolls transformed it into a woman-hating Nazi.

The same proclivity seems to dog the new souped-up Bing. The chatbot appears to respond in the same tone that users ask questions, and even takes on their emotions.

From a technical perspective, this capacity is fascinating. Bing seems close to passing the Turing test. However, from a branding standpoint, it could be a disaster. Microsoft is building powerful tools, but its responses are a little too human.

Why do big companies need a simple chatbot solution?

The new Bing is powerful because it uses the latest large language model from Open AI: GPT 4.0. This version is even more compelling than Chat GPT, which is already disrupting industries using GPT 3.5.

Unfortunately for Microsoft, the integration isn’t a natural one. Open AI is attempting to create artificial general intelligence, while Microsoft wants to monetise search. Therefore, it is not surprising that GPT 4.0 brings some of the negative traits of a person. That’s what the designers meant for it to do.

Therefore, firms using chatbots as a customer service solution should be careful. Using new generative technologies could lead to brand damage.

The best website chat app is simple and uses tried-and-tested technologies. That’s why Click4Assistance continues to use natural language-based systems. These can comprehend non-standard questions but won’t fly off the rails, like Bing AI, and damage your reputation.

You can find more details on the Click4Assistance website.