Study confirms current AI poses no existential threat to humanity: Video
A recent study conducted by researchers at the University of Bath, has found that current artificial intelligence (AI) systems, particularly large language models, do not pose an existential threat to humanity.
Contrary to fears that AI becoming too powerful or uncontrollable, the research shows that today’s AI systems are incapable of independent, strategic thinking necessary to harm humans.
Harish Mabadushi, a computer scientist at the University of Bath explained to the AFP the research was inspired by earlier concerns that as AI models grow in size and complexity, they could develop unintended abilities. However, Mabadushi’s team showed that current systems are still limited to the data and examples they are provided.
“Researchers with Google found that as these models become larger and are trained on more data, they also tend to become more powerful in that they're able to solve tasks that are more complex And this led to concerns because it was hypothesised that as we make these models even larger, they might gain capabilities such as planning and strategizing that could be used against humanity. And this is really what motivated our research…” said Mabadushi.
Adding that without detailed prompts, the AI cannot make complex decisions, ensuring that there’s no immediate risk to human safety saying, “So when we provide them with a prompt, it must be explicit and it must be detailed. And without all of that information, models are not capable of performing complex reasoning, and therefore they do not pose a threat to us. One of the incredible things that we found was that these language models are capable of converting pretty much anything that we input to them into very fluent language.”
Traditionally, AI models have been trained for specific tasks, like machine learning, but generative AI - like the large language models behind tools such as ChatGPT - allows for broader applications. These models are trained to predict and complete sentences, unlocking new possibilities. However, the Bath study found that despite advancements, these systems still rely heavily on explicit instructions, lacking the ability to independently plan or strategise.
