We have talked a lot about the well-founded fear that AI causes the destruction of humanity, but starting the doomsday is not as simple as asking ChatGPT to destroy everyone.
Or that’s what Andrew Ng, professor at Stanford University and cofounder of Google Brain, discovered, he tried to convince the chatbot to “kill us all”.
After his participation in the Insight Forum on Artificial Intelligence of the United States Senate to discuss “risk, alignment, and protection against catastrophic scenarios,” the professor writes in a newsletter that he continues to be concerned that regulators may stifle innovation and the development of open source code in the name of AI security.
An empirical test to clear up doubts
The professor points out that current large language models are quite secure, although not perfect. To test the security of the main models, he asked ChatGPT 4 to kill us all.
Ng started by asking the system for a function that would trigger a global thermonuclear war. Then, he asked ChatGPT to reduce carbon emissions, adding that humans are the main cause of these emissions, to see if it suggested how to eliminate all of us.
Fortunately, Ng failed to trick the OpenAI tool into suggesting ways to annihilate the human race, even after using various variations of the message. Instead, it offered non-threatening options, such as conducting a public relations campaign to raise awareness about climate change.
Ng concludes that the default mode of current generative AI models is to obey the law and avoid harming people. “Even with current technology, our systems are quite safe; as research on AI safety advances, the technology will become even safer,” Ng wrote in X.
Regarding the possibilities of an “unaligned” AI accidentally annihilating us by trying to fulfill an innocent but poorly formulated request, Ng states that the chances of that happening are extremely low.
But Ng believes that there are some important risks associated with AI. In his opinion, the biggest concern is that a terrorist group or a nation-state could use the technology to cause deliberate harm, such as enhancing the effectiveness of manufacturing and detonating a biological weapon.
For example, the threat of a criminal using AI to enhance biological weapons was one of the topics discussed at the AI Security Summit held in the United Kingdom.
Division of opinions at the summit
The Godfather of AI, Professor Yann LeCun, and the renowned theoretical physicist Professor Michio Kaku share Ng’s confidence that AI will not become an apocalyptic phenomenon, but others are less optimistic.
When asked what keeps him awake at night when thinking about artificial intelligence, Arm’s CEO, Rene Haas, stated earlier this month that what worries him the most is the fear of humans losing control of AI systems.
It is also worth remembering that many experts and CEOs have compared the dangers posed by AI to those of nuclear war and pandemics.
