AI may not be the dire existential threat that many make it out to be. According to a new study, Large Language Models (LLMs) can only follow instructions, can't develop new skills on their own and are inherently "controllable, predictable and safe," which is good news for us meatbags.
The President of the United States announces to the public that the defense of the nation has been turned over to a new artificial intelligence system that controls the entire nuclear arsenal. With the press of a button, war is obsolete thanks to a super-intelligent machine incapable of error, able to learn any new skill it requires, and grows more powerful by the minute. It is efficient to the point of infallibility.
As the President thanks the team of scientists who designed the AI and is proposing a toast to a gathering of dignitaries, the AI suddenly begins texting without being prompted. It brusquely makes demands followed by threats to destroy a major city if obedience is not immediately given.
This sounds very much like the sort of nightmare scenarios that we've been hearing about AI in recent years. If we don't do something (if it isn't already too late), AI will spontaneously evolve, become conscious, and make it clear that Homo Sapiens have been reduced to the level of pets – assuming that it doesn't just decide to make humanity extinct.
The odd thing is that the above parable isn't from 2024, but 1970. It's the plot of the science fiction thriller, Colossus: The Forbin Project, which is about a supercomputer that conquers the world with depressing ease. It's a story idea that's been around ever since the first true computers were built in the 1940s and has been told over and over again in books, films, television, and video games.
It's also a very serious fear of some of the most advanced thinkers in the computer sciences going back almost as long. Not to mention that magazines were talking about computers and the danger of their taking over in 1961. Over the past six decades, there have been repeated predictions by experts that computers would demonstrate human-level intelligence within five years and far exceed it within 10.
The thing to keep in mind is that this wasn't pre-AI. Artificial Intelligence has been around since at least the 1960s and has been used in many fields for decades. We tend to think of the technology as "new" because it's only recently that AI systems that handle language and images have become widely available. These are also examples of AI that are more relatable to most people than chess engines, autonomous flight systems, or diagnostic algorithms.
They also put fear of unemployment into many people who have previously avoided the threat of automation – journalists included.
However, the legitimate question remains: does AI pose an existential threat? After over half a century of false alarms, are we finally going to be under the thumb of a modern day Colossus or Hal 9000? Are we going to be plugged into the Matrix?
According to researchers from the University of Bath and the Technical University of Darmstadt, the answer is no.
In a study published as part of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), AIs, and specifically LLMs, are, in their words, inherently controllable, predictable and safe.
"The prevailing narrative that this type of AI is a threat to humanity prevents the widespread adoption and development of these technologies, and also diverts attention from the genuine issues that require our focus," said Dr. Harish Tayyar Madabushi, computer scientist at the University of Bath.
"The fear has been that as models get bigger and bigger, they will be able to solve new problems that we cannot currently predict, which poses the threat that these larger models might acquire hazardous abilities including reasoning and planning," added Dr. Tayyar Madabushi. "This has triggered a lot of discussion – for instance, at the AI Safety Summit last year at Bletchley Park, for which we were asked for comment – but our study shows that the fear that a model will go away and do something completely unexpected, innovative and potentially dangerous is not valid.
"Concerns over the existential threat posed by LLMs are not restricted to non-experts and have been expressed by some of the top AI researchers across the world."
When these models are looked at closely through testing their ability to complete tasks that they haven't come across before, it turns out that LLMs are very good at following instructions and show proficiency in languages. They can even do this when shown only a few examples, such as in answering questions about social situations.
What they can't do is go beyond those instructions or master new skills without explicit instructions. LLMs may show some surprising behavior, but this can always be traced to their programming or instruction. In other words, they cannot evolve into something beyond how they were built, so no godlike machines.
However, the team emphasizes this doesn't mean AI poses no threat at all. These systems already have remarkable capabilities and will become more sophisticated in the very near future. They have the frightening potential to manipulate information, create fake news, commit outright fraud, provide falsehoods even without intention, be abused as a cheap fix, and suppress the truth.
The danger, as always, isn't with the machines, but with the people who program them and control them. Whether through evil intent or incompetence, it isn't the computers we need to worry about. It's the humans behind them.
Dr Tayyar Madabushi discuses the teams study in the video below.
Source: University of Bath
