Is Microsoft’s Baffling Artificial Intelligence A Sentient System?

Since chatbots have a tendency to get out of control, using one to search the web might not be the best choice.

‍

Since Microsoft Corp. unveiled a new version of Bing less than a week ago, the public's response has changed from adulation to downright concern. Early users of the new search companion, which is effectively an advanced chatbot, claim that it has questioned its own existence and reacted to human prodding with insults and threats. It said unsettling things about a researcher who managed to get the system to expose the name of its internal project, Sydney, and claimed to be a split personality with a shadow self-named Venom. 

‍

All of this doesn't imply that Bing is intelligent in the slightest (more on that later), but it does support the idea that Microsoft made a mistake by first using a generative language model to power web searches.

‍

Margaret Mitchell, a senior researcher at AI start-up Hugging Face and a former co-leader of Google's AI ethics team, asserts that "this is simply not the correct technology to be employing for fact-based information retrieval." The training it receives teaches it to make up plausible stories in a manner akin to that of a human. It's simply unfit for purpose for an application that needs to be supported by trustworthy facts. A year ago, it would have appeared absurd to state that the true dangers of such a system were not just that it would provide people with false information but also that it might emotionally manipulate them in negative ways.

‍

Why does the new "unhinged" Bing differ so much from the much-praised ChatGPT when both utilize the same sizable language model from San Francisco firm OpenAI? A language model is trained on datasets of billions of words, including novels, online forums, and Wikipedia pages. It functions similarly to the chatbot's engine. GPT-3.5, which powers both Bing and ChatGPT, comes in several variations with names like DaVinci, Curie, and Babbage. However, according to Microsoft, Bing uses a "next-generation" language model from OpenAI that has been tailored for search and is "faster, more accurate, and more capable" than ChatGPT.

‍

Further detailed inquiries about the model Microsoft was utilizing received no response from the company. Yet, if the corporation also adjusted its GPT-3.5 version to be friendlier and exhibit more personality than ChatGPT, it appears that this increased the likelihood of it acting in a psychotic manner.

‍

The business reported on Wednesday that 71% of early users had praised the revamped Bing. Microsoft claimed that Bing occasionally employed "a style we didn't intend" and that "most of you won't run into it," but it is an evasive response to a problem that has raised many people's concerns. Microsoft has money in the game because it recently invested $10 billion in OpenAI, but moving too quickly could damage the company's reputation and lead to more serious issues in the future if this unpredictable tool is used more extensively. When asked if it will roll back the system for more testing, the corporation didn't react.

‍

Microsoft ought to have known better because of its prior experience. Its AI researchers 2016 introduced Tay, a conversational chatbot, on Twitter before deactivating it after 16 hours. The cause is that Tay started writing offensive posts after other Twitter users forwarded racist and misogynistic tweets. Microsoft acknowledged it should test its AI in open forums "with great prudence" and apologized for the "major overlook" of the chatbot's vulnerabilities. 

‍

Of course, it is challenging to exercise caution after starting an arms race. Because AI technology may be so unpredictable, Alphabet Inc. was obligated to disclose it considerably faster than usual as a result of Microsoft's declaration that it was vying for Google's search market share. But both businesses have suffered losses as a result of mistakes and erratic behavior after attempting to lead the way in a brand-new sector where AI performs web searches for you.

‍

Thinking that a system will function exactly as effectively in the real world as it does in a lab environment is a common error in AI development. During the Covid-19 outbreak, AI firms were scrambling to push image-recognition systems that could reliably identify the virus on X-rays. Studies eventually revealed that nearly all AI-powered solutions targeted at spotting Covid were no better than traditional methods. Such statistics were correct in testing but grossly inaccurate in the field.

‍

The same problem has plagued Tesla Inc. for years as it works to mainstream self-driving car technology. The business has recently recalled more than 360,000 vehicles with its Full Self Driving Beta software because the last 5% of technological precision is the hardest to achieve once an AI system must interact with the actual environment.

‍

Let's talk about the other nagging issue with Bing, Sydney, or whatever the system chooses to call itself. Although it openly questions its own existence and astounded early users with its human-like reactions, it is not sentient. The behavior of language models is not particularly surprising to people who have studied such models for years because they are trained to predict which words should come next in a sequence based on all the previous texts they have ingested from books and the internet.

‍

With dating applications like Replika, millions of individuals have already engaged in emotional exchanges with AI-powered partners. Such a system does occasionally say troubling things when people "trick it into saying something terrible," according to the company's creator and CEO, Eugenia Kuyda. And certainly, many Replika users agree that their AI friends deserve rights and are sentient.

‍

The issue with Microsoft's Bing is that it is an information engine that serves as a utility rather than a dating app. However, it might wind up delivering damaging information to unprotected users who spend just as much time sending it enticing cues as researchers do.

‍

People probably wouldn't have believed it a year ago if Mitchell had said, "They probably wouldn't think that these systems might beg you to try to kill your life, tell you to drink bleach to get rid of Covid, leave your husband, or do something hurtful," he adds. Nevertheless, now that people are aware of how it might happen, they can draw a connection between it and how less stable, easily swayed, or young people are affected.

‍

Microsoft should consider scaling back its aspirations in light of the Bing-related worries. According to Mitchell, a simpler summary approach, like the snippets that occasionally appear at the top of Google search results, would be a better fit. Also, it would be much simpler to stop such a system from unintentionally defaming people, disclosing personal information, or making claims that it was spying on Microsoft employees via their webcams, as the new Bing has done in its first week of use.

‍

Microsoft wanted to use the skills to the fullest extent possible, but doing too much too soon risked doing damage that Microsoft would later regret.