AI may not be as much of an existential threat as many make it out to be. New research suggests that large-scale 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 humans.
The President of the United States announced to the public that the defense of the nation has been handed over to a new artificial intelligence system that will control the entire nuclear arsenal. War will become obsolete thanks to super-intelligent machines that, at the touch of a button, can never make mistakes, can learn any new skill they need, and get more powerful by the minute. It will be as efficient as it is foolproof.
After the president thanked the team of scientists who designed the AI and toasted the gathering of dignitaries, the AI suddenly began sending unprompted text messages, making blunt demands and threatening to destroy major cities if compliance was not forthcoming.
This sounds a lot like the nightmare scenarios we've been hearing about with AI lately: If we don't do anything (and it's not too late already), AI will spontaneously evolve, become conscious, and reveal that Homo sapiens has degraded to the level of a pet — unless it decides to wipe out the human race.
Oddly enough, the fable above is from 1970, not 2024. It's the plot of the sci-fi thriller Colossus: The Forbin Project, about a supercomputer that takes over the world with depressing ease. It's a story idea that's been around since the first real computers were built in the 1940s, and has been told countless times in books, movies, TV and video games.
This has been a very real concern among the most forward thinking thinkers in computer science for almost as long – not to mention magazines talking about computers and the dangers of their domination over humans as far back as 1961. Over the past 60 years, experts have repeatedly predicted that computers will achieve human-level intelligence within five years, and far exceed it within ten.
It's important to remember that this isn't pre-AI. Artificial intelligence has been around since at least the 1960s and has been used in many fields for decades. AI systems that process language and images have only recently become widely available, so we tend to think of the technology as “new.” These are also examples of AI that are more familiar to most people than chess engines, autonomous flight systems, or diagnostic algorithms.
It also threatens to send job losses into the minds of many who have so far avoided the threat of automation, including journalists.
But natural questions remain: Will AI pose an existential threat? After more than half a century of misinformation, will we finally find ourselves under the control of a modern-day Colossus or HAL 9000? Will we be inserted into the Matrix?
According to researchers from the University of Bath and Technical University Darmstadt, the answer is no.
AI, and LLMs in particular, are inherently controllable, predictable and safe, according to research presented as part of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024).
“The common view that this kind of AI is a threat to humanity is hindering the widespread adoption and development of this technology, and also distracting from the real problems that need our attention,” said Dr Harish Tayyar Madhavshi, a computer scientist at the University of Bath.
“There is a concern that as models get bigger and bigger, they will be able to solve new problems that we cannot currently predict, and as a result there is a fear that these large models will acquire dangerous capabilities such as reasoning and planning,” added Dr Tayyar Madabhsi. “This has generated a lot of debate – for example, at the AI Safety Summit held at Bletchley Park last year we were asked for comment, and our work has demonstrated that the fears that models will go off and do something totally unexpected, innovative and potentially dangerous are unfounded.”
“Concerns about the existential threat posed by the LLM are not limited to non-specialists but have been expressed by some of the leading AI researchers around the world.”
When we examine these models more closely by testing their ability to complete tasks they have never been exposed to before, we find that LLMs are extremely good at following instructions and have strong language skills, even to give just a few examples, such as answering questions about social situations.
They cannot take actions beyond what they are instructed to do or learn new skills without explicit instructions. LLMs may exhibit surprising behavior, but this is always due to their programming or instructions. In other words, there are no godlike machines, as LLMs cannot evolve beyond the way they were built.
But the team stresses that this doesn't mean AI doesn't pose a threat at all: these systems already have incredible capabilities and will soon become even more sophisticated, with frightening possibilities for manipulation, creating fake news, outright fraud, unintentionally providing false information, being misused as an easy solution, or concealing the truth.
The danger, as always, is not the machines but the humans who program and control them. Whether through malice or incompetence, it's not the computers we need to worry about – it's the humans behind them.
Dr. Tayyar Madabhsi explains his team's research in the video below.
AI Safe
Source: University of Bath
