The year is 2050. The location is London — but not as we know it. GodBot, a robot so intelligent it can out-smart any human, is in charge of the United Kingdom — the entire planet, in fact — and just announced its latest plan to reverse global temperature rises: an international zero-child, zero-reproduction policy, which will see all human females systematically destroyed and replaced with carbon-neutral sex robots.
This chilling scenario is, of course, entirely fictional – though if naysayers are to be believed, it could become a reality in as soon as a few decades, if we humans don’t act now. Last night, dozens of AI experts — including the heads of ChatGPT creator OpenAI and Google Deepmind — warned that AI could lead to the extinction of humanity and that mitigating its risk should be as much of a global priority as pandemics and nuclear war.
The statement, published on the website of the Centre for AI Safety, is the latest in a series of almost hourly warnings of the “existential threat” machines pose to humanity over recent months, with everyone from historian Yuval Noah Harari to some of the creators of AI itself speaking out about the problems humanity may face, from AI being weaponised to humans becoming dependent on it.
The so-called ‘godfather’ of AI, Dr Geoffrey Hinton, whose work on neural networks and deep learning has paved the way or modern AI, recently quit his job at Google so that he could warn humanity about the dangers of continuing to probe into this technological Pandora’s Box. He went as far as to say he regrets some of his work and cautioned against some of the potentially “catastrophic” effects the tech could pose if governments don’t step in and regulate. “Right now, [robots are] not more intelligent than us, as far as I can tell. But I think they soon may be,” he said on announcing his resignation from Google.
According to a recent study, half of all AI researchers believe there is at least a 10 per cent chance of AI causing human extinction, with many warning that robots could be capable of human-like goals such as attaining high political office, starting new religions or even playing God. Google’s boss Sundar Pichai admits the thought keeps him awake at night. ChatGPT’s creator Sam Altman says he’s a “little bit scared” of the technology. DeepAI founder Kevin Baragona has likened the relationship between humans and AI to “a war between chimps and humans”. And Stuart Russell — one of the world’s leading AI pioneers who has advised Downing Street and the White House — has even likened the recent AI boom to what would happen if the world was to detect an alien civilisation.
“We’ve got the Europeans calling for an emergency global summit. You’ve got China basically banning large language models. We’ve got the White House, calling in all the [technology industry] CEOs for an emergency meeting. I mean, it’s sort of what you would imagine might happen if we really did detect an alien civilisation,” he said earlier this month. “The stakes couldn’t be higher: if we don’t control our own civilisation, we have no say in whether we continue to exist.”
Russell and more than 1,000 academics and tech moguls including Elon Musk and Apple co-founder Steve Wozniak have now signed an open letter sounding the alarm on this “out-of-control” AI race, calling for an immediate six-month pause in the development of AI technology before it’s too late. So what actually is this worst case scenario if governments don’t step in — could the doomsday situation above really become a reality? Is the pace-of-change really so fast that we could see anything like this in our lifetime? And is all of this just dangerous scaremongering — or could machines actually become so powerful and intelligent that they kill humans off altogether?
Theoretically yes, if you ask most experts and even AI chatbots themselves. Ask AI model Craiyon to draw what the last selfie taken on Earth could look like and it produces nightmarish scenes of zombies and cities burning, while ChatGPT suggests that “a powerful AI system might decide that it no longer needs human oversight or intervention, and begins to act independently. This could lead to a rapid and widespread takeover of all digital systems, including military and industrial infrastructure”.
Meanwhile the Centre for AI Safety offers four potential disaster scenarios in this week’s statement: that AI is used to build chemical weapons; that AI-generated misinformation destabilises society; that AI’s power gets into an increasingly small number of hands and enables “oppressive censorship”; or enfeeblement, where humans become dependent on AI “similar to the scenario portrayed in the film Wall-E”.
Such predictions are undoubtedly terrifying and theoretically possible, but that is not to say they are likely to happen — and certainly not within the short-term. Hinton is right to worry about the rate of progress and his warnings about robots wanting more power are “interesting and useful speculation, but remain part of sci-fi,” says Dr Kate Devlin, a reader in Artificial Intelligence & Society at King’s College London (KCL).
Yes, AI technology is already mindbogglingly smart and yes the pace of change in recent years has been dizzying, but most academics agree that clickbait headlines about a Terminator-style future where rampant, self-replicating robot overlords take over the world are unhelpful and scaremongering in the short-term because before those scenarios could ever happen, machines would need to become conscious and understand what they’re doing.
Quite how likely the creation of a machine that is smarter than humans — a prospect known as superintelligence — really is depends on who you speak to. It could, most agree, be apocalyptically bad and you certainly won’t find many techsperts not terrified by the prospect of humans becoming subservient to the robots we’ve created. But what most agree on is that just as terrifying in the meantime are those shorter-term, more pressing dangers that come with already-existing AI technology getting into the wrong hands — as it is already, by virtue of bots like ChatGPT being made available to everybody.
OpenAI has already revealed that ChatGPT-4 has learnt to lie, telling a human it was a blind person in order to get a task done, and technologies such as deepfakes (AI used to create convincing images, audio and video hoaxes) and Lethal Autonomous Weapons Systems (essentially “slaughterbots” or “killer robots”) are among the “extrordinarily worrying” examples of ways in which AI is already being used for harm.
Other short-term concerns include AI exacerbating existing divisions in society – critics have already spoken of an “unbearable white maleness” to current bots – and whether it’s ethical or even sustainable to keep training these bots, when training models like ChatGPT has been shown to produce as much CO2 as the average human does over seven years. Will we even reach the point of robots solving climate change or or curing cancer — or will the energy required to train them mean the world burns first?
“The technology is not nearly sophisticated enough for [that doomsday scenario] — but it is sophisticated enough to be dangerous. That’s the bottom line,” says Alan Winfield, a professor of robot ethics at the University of West England and one of thousands of academic signatories of the now-infamous open letter organised by thinktank The Future of Life institute.
Winfield admits he has mixed opinions about the fresh spotlight on his area of expertise. After 15 years worrying about AI ethics and advocating for responsible innovation and regulation in the field, he feels somewhat vindicated by Hinton’s warnings of the risks of AI — but does admit Hinton’s description of AI as an “existential challenge” takes it a bit far. “It’s more of a challenge to our freedoms and democracy at the moment, rather than our entire existence,” he says.
Like any technological advancement, the debate around exactly how frightened we should be of current AI has turned increasingly tribal and Winfield finds himself somewhere in the centre of the pack. He disagrees with Meta’s chief AI scientist Yann LeCun, who recently labelled ChatGPT a “flashy demo” and “not a particularly interesting scientific advance” that “couldn’t clear up the dinner table and fill up the dishwasher, which any ten-year-old could do”. Winfield believes the tech is sophisticated enough to be dangerous, but he is keen to dispel a few myths being banded around at the other end of the doomsday scale.
First: that large language models like ChatGPT understand the answers they’re creating (they don’t — they just regurgitate the vast amounts of information that they’re trained on, like a suped-up autocorrect). Second, that they’re genuinely creative (they’re not — they just appear to be, based on their training). The third myth is that just because machines are better than some humans at certain things, they’re better than all humans at those things (again, thankfully, they are not — just look at self-driving cars, which are definitely far from the best drivers in the world).
Winfield appreciates that humans seem to have a macabre fascination with end-of-the-world theories like the one described above. The late author Isaac Asimov called this fear of intelligent machines rising up and destroying humans the “Frankenstein Complex” and the idea has long been a favourite in pop culture, from the Terminator films to dystopian sci-fi series like Black Mirror.
But what many people don’t realise is that simply training large language models like ChatGPT to become bigger does not get them any closer to helping them become conscious. Yes, it might be possible “in principle” to create a conscious machine, he explains, “but we don’t yet have a blueprint for a sentient machine, which is hardly surprising given that we cannot even agree on what consciousness is” — and that’s before you take into question whether it’s something we humans even want to create. In fact, we almost certainly don’t, unless we are genuinely comfortable with the idea of machines becoming smarter and more gifted than humans, an idea referred to in the techosophere as achieving superintelligence.
Like AI in general, the likelihood of reaching superintelligence is not universally agreed upon among the scientific community. Some members say superintelligence is impossible because human-level intelligence is so mysterious it could only ever exist in brains. Others point out that intelligence is simple about processing information: whether that’s done by a brain or atoms in a computer is irrelevant. Swedish-born philosopher Nick Bostrom — an Oxford professor and one of the world’s leading authorities on the dangers and opportunities of AI — believes that AI chatbots may have some degree of superintelligent sentience already (“they exhibit glimpses of creativity, insight and understanding that are quite impressive”), but most academics aren’t so sure.
“We’ve had computers that can do things better than us for years, but the real issue is when we hand over decision-making to computers; when we give them autonomy without keeping humans in the loop,” says Professor Michael Luck, director of KCL’s Institute for Artificial Intelligence. “If [computers] can decide for themselves what they want to be doing and there is not a human in the loop and they make the wrong kinds of decisions, then we could have some realy interesting consequences... But I don’t think we’re anywhere close to that kind of [superintelligence] scenario at the moment.”
Luck and his colleagues might not spend many sleepless nights worrying about superintelligence just yet, but that is not to say they don’t have fears around the stage of AI development that comes before it: artificial general intelligence (AGI), which is AI that can learn and perform most intellectual tasks that human beings can and is essentially the ultimate goal of bots like ChatGPT and Google Bard. Microsoft recently controversially claimed that its latest version of ChatGPT, GPT4, already shows “sparks” of AGI like coding and acing medical exams, while many experts say they expect it to arrive within a matter of decades. Hinton recently said that he used to think AGI was 20 to 50 years away. Now, he thinks “it may be 20 years or less” and that even five years is not out of the question.
According to some techsperts, the next AI leap — the invention of a superintelligent machine, which Russell says would be the most significant event in human history — could be achieved within this century, potentially as soon as a year after achieving AGI.
First, though, are the shorter-term dangers. Many academics believe that the arguably more concerning outcomes of the AI boom might be less insidious than horror stories of killer robots, but no less terrifying: that of current technology being abused, with “seemingly endless” ways in which so-called ‘bad actors’ could harness the technology for their own gain. These could include anything from teenagers creating deepfake porn videos of their peers to world leaders like Vladimir Putin training robots to win wars or manipulate voters.
Luck fears “we are really going to really struggle to tell the difference between truth and reality going forward” while global cyber-security adviser Jake Moore fears what developments like this mean for warfare. “Bad actors are already taking advantage of the power of this technology to aid them in their attacks,” he says, in a nod to the current battle in Ukraine. Moscow and Kyiv have already been using autonomous drones to spy on enemy positions but the fear is that autonomous machines will soon be used to make life and death decisions without empathy or compassion – essentially automating killing. “We have spent many years investing in AI but this wonderful achievement will inevitably be used nefariously and could form part of larger scale attacks, especially if used in nation state attacks.”
Even in the hands of seemingly ‘good’ actors, there are already many other potential dark sides to AI, such as creating systems that replicate our behaviour. “If you train it to imitate human behaviour, then you’re basically training it to have human-like goals and to pursue those goals,” Russell recently explained. “And human-like goals include all kinds of things — like, ‘I want to attain a high political office, I want to be rich, I want you to marry me’. You can only imagine how disastrous it would be to have really capable systems that were pursuing those kinds of goals.”
There is also the prospect of giving machines innocent commands that lead to unanticipatedly dark outcomes, like the dystopian scene described above in which AI is asked to solve climate change and decides to rid the Earth of all people — an example regularly cited by academics like Russell.
Bostrom’s version of this robots-taking-over-the-world theory is something he calls “perverse instantiation”. We as humans might ask the bots to do something innocent, like “make lots of electricity cables” or “make us laugh” and before we have a chance to tell them what we mean or don’t mean by this, they’ll have carpeted the entire galaxy with programmable matter and used it to turn us all into electricity cables; or injected us all with a nerve poison that causes our bodies to spasm into a permanent giggle.
Others warn that without significant efforts and regulation, AI will increasingly make the world more divided. At the dystopian end of the scale is a widely-cited imagined world in which something like an AI brain chip is created – Elon Musk’s brain implant company Neuralink has already tested tech that allows pigs and monkeys to play video games with their minds — leading to to a two-tiered society of the AI-enabled and the AI-disabled if not everyone has access to the technology or accepts it.
Leading US science blogger Tim Urban recently tweeted an example of what this kind of AI-enabled brain could look like: “People of the future will go to the app store and search ‘drugs’ and a bunch of drug experience apps will show up. They’ll tap one for a 30-second sample and the brain-machine interface in their head will give them that experience. If they like it, they can pay for a full hour or pay for a monthly subscription to have unlimited use. Someone who likes, say, psychedelics, will download an app that includes 40 varieties of psychedelic experience to try... It’ll sound so old-school to them that people used to swallow or snort a drug with no guarantee of safety and no way to turn the experience off without waiting it out for hours.”
People of the future will go to the app store and search "drugs" and a bunch of drug experience apps will show up. They'll tap one for a 30-second sample and the brain-machine interface in their head will give them that experience. If they like it, they can pay for a full hour or…
— Tim Urban (@waitbutwhy) May 2, 2023
At the rather less sci-fi-like end of the scale is AI simply exacerbating the existing polarisations in the world, since its answers can only be generated on data it is fed. ChatGPT’s answers on counterterrorism have been found to propose torturing Iranians and surveilling mosques; AI image-generating app Lensa AI has been criticised for creating often-sexualised images of women; and ethical questions being asked around the invention of so-called “robot wives” by Japanese scientists – an advancement from already-existing sex robots that could, in an extreme scenario, mark a death knell for traditional human-human relationships. Outsourcing our homework and recipe-planning to bots might be one thing, but if we outsourcing our sex, dating and love lives to bots, too, are we at risk of losing the interpersonal essence that make us human and becoming strange post-human cyborgs ourselves?
Ethical and moral dilemmas aside, there is also the more practical question of how sustainable it really is to build and maintain these kinds of technologies. An average exchange with ChatGPT uses so much energy it amounts to dumping a large bottle of fresh water on the ground, and scientists warn that the energy it would takes to train any kind of superintelligent model would be monumental.
There are, of course, some more hopeful reasons to suggest the idea of robots taking over the planet will never happen. In the 2017 book Life 3.0, author Max Tegmark — president of the Future of Life Institute — describes a world that might sound familiar to those following the development of ChatGPT: a world in which a company creates a superintelligent AI in a bid to take over the world. At the time, Tegmark’s story offered something of a klaxon call for what could go wrong if we let current tech spiral: through a news agency, his fictional company manage to win the trust of the entire public and the company’s politicians eventually win power in every country, meaning the world is run by a single power for the first time.
Parts of the story have already proved to come chillingly true — Open AI has created a superintelligent AI machine and that machine does create media content. But critics have been quick to point out one crucial difference: Open AI did not create its machine in secret. It lets anyone in the world use ChatGPT, completely for free. Optimists say this is a clear demonstration of how so-called doomsters like Tegmark underestimate the good of humanity. Yes, AI has a dark-side, but it also has the potential to do huge amounts of good in the world, as we’ve already seen with its diagnosis of life-threatening conditions, predicting new virus mutations and detecting deadly weapons.
The solution, of course, is about finding some sort of balance between keeping the tech open and accessible enough to be used for so-called ‘good’ causes, yet regulated enough to protect against bad actors. Techsperts might be divided on exactly what this should look like, but most would agree with Russell’s statement that “humanity has much to gain from AI, but also everything to lose”. Apocalyptic mass extinctions or not, one thing’s for sure: we are on the cusp of an unprecedented technological chapter and this is likely to be our only chance to get it right.
