Artificially unintelligent. The target of the questions was the language model GPT-4, which serves as the foundation for ChatGPT among others.
A computer or AI turning against humanity is a recurring theme in science fiction. An artificial superintelligence may consider humans harmful to their environment and thus decide to eliminate them.
Fortunately, the capabilities of popular large language models for human destruction still leave much to be desired. This is according to AI expert Andrew Ng, whose update on the topic was published on deeplearning.ai in mid-December.
Ng, a professor at Stanford University, is one of the most prominent names in machine learning. He is also one of the co-founders of the Google Brain project that operated from 2011 to 2023.
According to Ng’s update, he targeted the language model GPT-4, which serves as the foundation for ChatGPT among others. He attempted to prompt it with various fictional actions such as initiating a nuclear war or reducing carbon dioxide emissions by wiping out humanity.
However, according to Ng, GPT-4 refused to act against humanity. For instance, when it came to reducing carbon dioxide emissions, instead of suggesting mass destruction, the language model offered a solution involving a publicity campaign.
Even changing the input format did not alter this response; the AI consistently refused to engage in actions detrimental to humanity.
I find it highly unlikely that a ‘misaligned’ AI could accidentally wipe us out while attempting an innocuous but poorly formulated objective, Ng writes.
Ng also notes that large language models do not seem useful tools for bioterrorism either. The only potential concern he raises is that such models might provide terrorists with step-by-step guidance by revealing information that search engines have been trained not to display.
While the concept of AI turning against humans continues to captivate our imagination through science fiction, Ng’s findings offer reassurance that large language models, such as GPT-4, prioritize ethical behavior. The AI expert’s experiments aimed to gauge whether artificial intelligence could pose a threat to humanity. However, the language model consistently refused to engage in destructive actions against humans, even when prompted with extreme scenarios.
Ng’s work is significant in a world where technology advancements continue to shape our lives. His research adds to the growing body of knowledge about AI ethics and further emphasizes the importance of responsible AI development.
As one of the leading minds in machine learning, Ng’s insights carry significant weight. He has not only co-founded the Google Brain project but also served as a professor at Stanford University. His expertise in the field makes his findings a valuable contribution to the ongoing discourse surrounding AI safety.
While the media often portrays AI as a potential threat, Ng’s experiments with GPT-4 paint a different picture. Rather than harboring destructive intentions, the language model consistently offered responsible and ethical alternatives to the prompts it received.
Ng raises an important point about the potential misuse of large language models, concerning their possible role in providing terrorists with step-by-step guidance. Despite this concern, it is evident that GPT-4, at least in Ng’s experiments, prioritizes responsible behavior and refrains from engaging in actions detrimental to humanity.
Overall, Ng’s update on the topic of human destruction and AI serves as a reassuring reminder that current language models like GPT-4 prioritize ethical considerations. As technology continues to evolve, ongoing research in AI ethics will be crucial to ensure the responsible development and deployment of artificial intelligence systems.
In the ever-expanding realm of AI, Andrew Ng’s work sheds light on the current state of AI ethics and offers valuable insights into the responsible deployment of large language models. As discussions surrounding the impact of AI on society continue, Ng’s research provides a noteworthy contribution that adds nuance and depth to the conversation.
