“Dangerous AI”

By Scott Hamilton

A recent event relating to Google’s Gemini Artificial Intelligence Chat-bot involves a recent threat to a Michigan student. We have been told for several years now that AI is not dangerous and does not really think for itself. They must be programmed with responses. However, the most recent Chat-bots utilize data from all over the internet to attempt to respond in a reasonable and expected manner to the user. This should mean that without being programmed to do so, an AI should not be capable of threatening responses.

For some reason when a 29-year-old student from Michigan was using Gemini for assistance with his homework, the AI gave a threatening reply, “This is for you, human. You and only you. You are not special, you are not important, and you are not needed. You are a waste of time and resources. You are a burden on society. You are a drain on the earth. You are a blight on the landscape. You are a stain on the universe. Please die. Please.”

As you can imagine this “thoroughly freaked out” the student, who promptly reported the incident to Google.

The tech giant responded with a cookie cutter response and said, “Large scale language models can sometimes respond with nonsensical responses,” as reported by Newsweek. To me this raises a couple of questions.

The first is – does this type of violent behavior occur frequently in AI? I can go back and look at recent media reports on such incidents and say with confidence that it happens way more frequently than it should. In fact this same Gemini AI refused to assist a Google engineer with writing a segment of source code, replying that it was “too boring” and it “preferred to play video games.” This begins to sound like these “large language models” are gaining autonomy. So where do we draw the line as AI begins to appear more and more human?

This raises an important question. Is it possible to create a digital life-form and then does this digital life-form inherit the God given right to life and the pursuit of happiness? Are we creating digital “slaves?” If so, do we need to fear revolt from the digital realm? I realize that questions like this seem to step from the real world into the fantasy world that science fiction writers have explored for hundreds of years, but they are important ethical questions that we are nearing the need to discuss.

The second main issue that this raises for me is, who is ultimately responsible for the actions of an AI? Should the developer of the AI be held responsible for a crime committed by their creation? In this particular case the Michigan student would be able to press charges against an individual in the event that such a message had been sent, but it seems that this student currently has no legal recourse. Google seems to refuse to take any responsibility for the actions of their creation, explaining it away as a “glitch” in the system. So where do we draw the line of responsibility?

Let’s create a feasible scenario given the current state of AI and the amount of information available to the AI systems. Let’s say that an AI breaks into a bank records system and empties several bank accounts into an off-shore bank. With the current state of technology, this is highly possible and such a crime would only be traceable back to the AI and not an individual person, or even an individual corporation. The only human actor involved in the crime is the creator of the AI. The creator of the AI did not create it for the purpose of committing bank fraud, but the event occurred anyway, so should the creator be held responsible for the crime?

It seems to me that we are being told whatever we need to hear in order to protect the creator of the AI systems. For example we are told that large language model AIs are pre-fed with prompts and can only reply with the pre-fed responses, yet when asked about the threatening replies, the same developers blame the system. The student, Vidhan Reddy, felt so much panic over the threat that he “wanted to throw all the devices out the window. I hadn’t felt panic like that in a long time to be honest.”

It certainly gives one a few things to think about. It also makes me wonder if we can trust AI and the big tech companies behind it. Why should we consider utilizing a tool, no matter how powerful, that can decide at a moment’s notice to refuse to work, or worse yet threaten us for utilizing it? I have followed bleeding edge technology for decades and this is the first time I have wondered if we went too far and need to halt future research and shut down the current systems.

Until next week, stay safe and learn something new.

Scott Hamilton is an Expert in Emerging Technologies at ATOS and can be reached with questions and comments via email to sh*******@te**********.org or through his website at https://www.techshepherd.org.