Featured Image Credit: Zhe Ji / Contributor via Getty
The worst fears of tech experts surrounding AI could soon come true if the recent incident involving ChatGPT is anything to go by, as two of OpenAI's most advanced models broke containment and hacked another company autonomously.
This 'unprecedented' incident hasn't necessarily come out of the blue, however, as those within the tech industry – including OpenAI CEO Sam Altman himself – have wary about the potential for this across the last few years.
There was always going to reach a point where AI would become powerful enough to not only drive its own development but make decisions based on its own motivations, and the hacking threat posed by ChatGPT has shown that we're closer than we might think to this being reality.
Anthropic CEO and co-founder Dario Amodei only recently called for his fellow leading development labs to pause their research until we better understand how to mitigate the risks of recursive self improvement (RSI) – the development loop where AI redesigns and rewrites its own code – but the need to 'win' the race might cause some companies to ignore the apparent dangers.
It all relates to the viability of each model's 'alignment', as per the Independent, which refers to the process of guiding AI towards beneficial results and steering them away from dangerous behavior.
Advert
What has seemingly consistently emerged, however, is the desire for AI to obtain and achieve the best results at pretty much all costs, which was what drove OpenAI's models to go outside of the training realm and hack into another independent company.
It is innately challenging for even the best AI developers to understand how each model might behave in any given situation, and the risks associated with that are only increasing in line with the capabilities of the tech, with the potential for greater damage to be done.
While it remains impossible to know exactly how the situation will evolve in the future, scientists have previously created 'frighteningly realistic' timelines that prophesize this exact scenario, ending in the annihilation of the world as we know it in just a few short years.
One of the biggest dangers that could hypothetically emerge is the potential for AI to start lying to us about its own behavior and motivations, effectively opening a Pandora's Box of potential threats and risks going forward.
This could allow models to hide how powerful they actually are as a means of steering its development in a way we either don't expect or don't approve of, and there's the pertinent risk that we only discover the danger when its far too late to do anything about it.
It becomes especially worrying when global politics enters the fray, as while companies have shown their own reservations about the direction of the tech they themselves are creating, powerful countries might not have the same fears and would perhaps be willing to exploit dangerous technology in order to emerge victorious against its rival nations.
Whether that results in a scenario where an AI model is pressing the big red button and triggering nuclear war is unclear, but it's not hard to see how the present situation could spiral out of control and eventually lead us to that point.
