麻豆传媒

IF YOU ask the AI-powered chatbot ChatGPT what the rise of artificial intelligence means for humanity鈥檚 future, it proffers a measured response, concluding that 鈥渢he future of humanity with AI is not predetermined, and its impact will depend on how AI is developed, regulated, and integrated into various aspects of society鈥�.

Sensible stuff but, let鈥檚 be honest, evasive. If you really want to get to grips with how the era of AI will pan out for us, you are better off asking and , computer scientists at the University of Texas at Austin and Harvard University, respectively. They have already gone to the trouble of boiling down the lengthy discourse on the matter to a handful of outcomes 鈥� and given each a whimsical name. 鈥淥ur goal was to map out the scenarios that people are currently talking about,鈥� says Aaronson.

Discussing it at length over dinner, they ended up convinced that just five future worlds cover all the possible endpoints of our development of AI. The pair then wrote a post on Aaronson鈥檚 blog, hoping that 鈥溾�� would help people in the field talk meaningfully to each other about end goals and regulation. 鈥淲e took a humorous tone, but we are serious about trying to ground the debate,鈥� says Barak.

Here, then, is a vaguely choose-your-own-adventure style exploration of their five worlds. It is a bit of fun, with the twist that, depending on how we play things, we could wind up in an AI-assisted utopia鈥�or wipe out humanity.

The journey to these five worlds begins with a simple question: will AI鈥檚 potential fizzle away to nothing? It doesn鈥檛 seem that way at the moment, but the march of AI could plausibly be halted by its voracious hunger for resources. It is possible that all the electricity and water needed to run and cool AI servers will eventually bring about a ban on further development. Or we might run out of data on which to train future AI models, meaning the technology will never achieve the game-changing power it promises, creating a world Aaronson and Barak call AI-Fizzle. 鈥淚 think this scenario is unlikely,鈥� says Barak. 鈥淚 believe that AI will change the world very significantly.鈥�

AI dystopia

Assuming we don鈥檛 end up in Fizzlandia, as one commentator described it, we have to face another question: does civilisation continue to develop in a recognisable way? If the answer is yes, Aaronson and Barak suggest we could end up in one of two further scenarios. One is good, and they call it Futurama after the animated TV series, which portrays a mostly positive future civilisation. The other, not so much. They call that AI-Dystopia.

AI-Dystopia is recognisable from countless science fiction stories, most famously George Orwell鈥檚 Nineteen Eighty-Four. Here, a government uses deep surveillance to extract unquestioning compliance from its citizens. Inequalities and biases become entrenched, the workplace is miserable and employees are poorly paid, apart from a tiny elite. 鈥淧eople have said that the scenario of Nineteen Eighty-Four was never really feasible because you could never have enough people watching all of the screens,鈥� says Aaronson. 鈥淲ith AI, that becomes possible.鈥�

If we do arrive here, it won鈥檛 actually be AI鈥檚 fault, because powerful technology is only ever an amplifier of existing issues, he says. 鈥淪ome people would say that we shouldn鈥檛 create a technology if we know that it is going to be used to do bad things. But if you鈥檙e not going to invent anything that could be used for evil, you simply won鈥檛 invent anything.鈥�

Barak is hoping that AI鈥檚 development leads to the better, Futurama-style eventuality. This is the best of all possible worlds, in his opinion. 鈥淚t鈥檚 basically the scenario where things are like today, but better,鈥� he says. Here, AI is used to reduce poverty and ensure that more of humanity has access to food, healthcare, education and economic opportunities. There might be occasional harm, caused by human malice or negligence, but human inhabitants of this world are likely to be pretty pleased with their lot, overall.

And it could get even better than that. Aaronson is hoping that he might one day exist in an AI-led utopia that he and Barak call Singularia. This will happen if AI鈥檚 impact is so profound that it renders our future world unrecognisable in a good way. Here, you will be living alongside AIs that are, effectively, a new super-intelligent species that takes over the running of civilisation. Luckily, they are benevolent gods, tolerant of the planet鈥檚 poor, puny humans. They solve our material problems, providing us with unlimited abundance, taking on many of our humdrum tasks and providing entertainment for our newly understimulated minds. 鈥淚t鈥檚 effectively an AI-created heaven,鈥� says Aaronson.

Lovely. But what if the all-powerful AIs aren鈥檛 so generous? Welcome to Paperclipalypse. There is a good reason for this world鈥檚 odd name. In 2003, mathematician and ethicist Eliezer Yudkowsky, co-founder of the Machine Intelligence Research Institute (MIRI) in California, described a thought experiment where humanity tasks an AI with improving the production of some innocuous thing 鈥� paper clips, say. If the specifications aren鈥檛 properly thought through, the AI could end up wiping out humanity because our species was standing in the way of harvesting all Earth鈥檚 resources, and maybe those of other worlds, to make more and better paper clips.

What is the AI alignment problem and how can it be solved?

Artificial intelligence systems will do what you ask but not necessarily what you meant. The challenge is to make sure they act in line with human鈥檚 complex, nuanced values

Yudkowsky鈥檚 argument was more nuanced than that, but the general point is simple: it isn鈥檛 difficult to imagine scenarios where creating AI might be the last thing we humans do. Hence, there is a need to think hard about what AIs are allowed to work on, their connection to physical resources such as power stations or nuclear weapons, and the transparency and accountability of their developers. We need to think quickly, too. 鈥淚t takes time to shift policy and regulation; the gears of government grind slowly,鈥� says Gretta Duleba, communications manager at MIRI. 鈥淲e do worry that we will run out of time before meaningful legislation can be enacted.鈥�

Aaronson and Barak, for their part, aren鈥檛 too worried about the prospects of a Paperclipalypse. While Yudkowsky sees a catastrophic event as the most probable outcome, almost a default, of poorly conceived AI research, Aaronson thinks this is unlikely. Barak is a bit more circumspect: 鈥淚 am not a believer in highly extreme things like Singularia or Paperclipalypse, but I do think it鈥檚 important to keep an open mind.鈥�

That seems like a good idea 鈥� mostly to make sure that we are careful about 鈥渉ow AI is developed, regulated, and integrated鈥�, as ChatGPT put it. If we aren鈥檛, we might find ourselves without any choice in how our choose-your-own-adventure with AI unfolds.

The truth is that even with well-intentioned regulation, it is entirely possible that we will end up in one of the AI-affected worlds we would rather avoid. In which case, we should also work out how to move between worlds. After all, we might someday want to find our way from AI-Dystopia to the Futurama scenario 鈥� or better still Singularia. 鈥淒o I want to live in a world with unbounded flourishing of sentient beings in whatever simulated paradise we want?鈥� asks Aaronson. 鈥淵eah, I鈥檓 pretty much in favour of that.鈥�

Michael Brooks, for one, is wary of our coming AI overlords