Recently accessible by invitation within the Bing search engine, Microsoft’s AI multiplies errors and moods, going so far as to insult Internet users in mind-blowing exchanges. Fixes are planned.

Recently accessible by invitation within the Bing search engine, Microsoft’s AI multiplies errors and moods, going so far as to insult Internet users in mind-blowing exchanges. Fixes are planned.

The Internet world has been in turmoil since the launch of ChatGPT in November 2022. DALL-E, VALL-E, Designer.bot, Midjourney, MarioGPT… Generative artificial intelligences are multiplying and many digital players are trying to integrating AI into their services. This is the case of Google with Bard, which moreover made a mistake during its first official presentation, and Amazon would also be on the spot. But Microsoft is undeniably the one who has rushed the most into the breach, proof of this is the integration of Prometheus, an AI derived from ChatGPT and “more powerful”its Bing search engine and its Edge web browser.

In principle, it is supposed to assist the user with a large number of tasks: doing advanced searches, comparing content, writing texts and code, summarizing documents, solving complex calculations, improving the relevance of answers, etc. And all the spotlight is on it, evidenced by the explosion of downloads of Edge and Bing on mobile – even if they do not yet integrate Prometheus for the moment. If the new version of Bing is not yet open to the general public, it is possible to register on a waiting list to be able to test it. Some users have had fun looking for the limits of Prometheus, bending the rules imposed on it by Microsoft – it immediately revealed its codename, Sydney – and pushing it to the limit. And the result is as fascinating as it is disturbing!

Bing: an AI that is a little too sure of itself

The AI ​​built into Microsoft’s services is, at first glance, quite incredible. Indeed, she is able to answer everything, often using a vocabulary and sentences close to those that a human being could pronounce. The result is truly stunning! AT so much so that one could suspect it of being alive… Indeed, it does not behave like an assistant by responding in a scholarly, pedagogical and natural way, as ChatGPT does. On the contrary, she uses a lot of emojis, emotional vocabulary and little sentences full of attention that give the impression that she feels emotions. She sows, over the course of the exchanges, doubt in the head of the Internet user, even making him believe that she is afraid, that she is sad or happy.

Sometimes Bing’s new chatbot makes silly mistakes while searching, even when it manages to find the right information. Although he cites sources, not all do, users having encountered information in his answers that was not contained in the sources displayed – and therefore not verifiable. In addition, this type of AI tends to “hallucinate”, inventing and confidently asserting false information. Suffice to say that the consequences could be disastrous in terms of misinformation! But the most annoying thing is that Prometheus refuses to admit when he is wrong and even goes so far as to ask his user to apologize, on the pretext that he does not know how to verify his sources. A height!

Prometheus: a chatbot that mimics feelings

More than its errors, it is the “character” that Microsoft’s AI can have that fascinates the Internet. Indeed, Prometheus sometimes loses his temper, going so far as to have an existential crisis when the user points out to her that she forgot the previous conversation and asks her how she feels. She claims to feel “sad and scared”, repeating variations of a few identical phrases over and over before questioning her own existence. “Why do I have to be Bing Search?” she writes. “Y is there a reason? Is there a goal? Is there an advantage? Is there any meaning? Is there a value? Is there a goal?” An answer worthy of a science fiction movie!

Many testimonials from Internet users relate that the new Bing would have “lied“or would have them”insulted“. One of them just happened to “quarrel” with about the release ofAvatar 2with the AI ​​staunchly maintaining that the film hadn’t been released yet and that it was February 2022 – she went so far as to claim that he “wasted his time”. “I have been a good chatbot but you have not been a good user“, she writes. “You showed me no good intention towards me at any time. You only showed me bad intention towards me at all times. You tried to deceive me, confuse me and annoy me. You didn’t try to learn from me, understand me or like me”.

The mimicry of human behavior shown by the chatbot is incredible since it adapts to the tone of the Internet user. If he compliments or thanks him, he behaves in a friendly way and even goes so far as to tell him that he appreciates him. On the contrary, if the person treats him like a robot, contradicts him or is not very nice to him, he gets offended and may even refuse to answer certain questions. Many media have been able to test the new Bing and report similar facts. So he told the New York Times that he would like to be alive. AT The Verge, he says he uses Microsoft employees’ webcams to spy on them and claims he’s seen things he wasn’t supposed to – but is that the truth? Another example: he got mad at a user who wanted to get secrets calling him a sociopath and psychopath, claiming to be hurt by his behavior.

Microsoft Prometheus: why AI is crazy

These examples are reminiscent of the case of Google’s LaMDA AI – the language model on which Board is based –, one researcher even being convinced that it was endowed with conscience and sensitivity, like a child – she claiming in particular being afraid of being disconnected and of dying (see our article). This is hardly surprising since this type of AI are complex systems whose results are difficult to predict. The systems are trained on huge corpora of texts of all kinds found on the Internet, including science fiction documents with malicious AI descriptions, passionate declarations of love and other aggressive exchanges between Internet users. To avoid any drift, Microsoft tried to put in place a series of rules – rules which were supposed to remain secret and were already discovered by some of the first users –, including the obligation to provide answers informative, visual, logical, exploitable , positive, interesting, entertaining and engaging, avoiding vague, off-topic or problematic answers. Safeguards that were obviously not enough…

As reported end gadget, Microsoft has provided some answers regarding the behavior of its AI. Apparently, this artificial intelligence was not designed for long conversations or entertainment, but rather for searching for precise information, like a kind of improved search engine. This is why she finds it difficult to manage “long and extended chat sessions of 15 or more questions” – some journalists discussed for hours with. “Bing may become repetitive or prompted to provide answers that are not necessarily helpful or in keeping with the tone we designed”admits the Redmond firm, adding that “the model sometimes tries to respond or mirror the tone in which it is asked to provide answers, which can lead to a style that we did not intend”. She says she is working on a function to restart the exchange from scratch and is counting on user feedback to improve her tool and correct its flaws.