Tom Gruber speaks in a soft and deep American drawl. Passionate and methodical, he reflects on the moment he and two colleagues created Siri – Apple's virtual assistant – the high point of his 40-year career in Silicon Valley's pursuit of artificial intelligence.
"Around 2007-2008, we had everything in place to bring real artificial intelligence into everyone's hand, and that was thrilling.
"Siri was very playful. And that was by design," he declares with a wide grin and a laugh almost like a proud dad.
"Now it's used roughly a billion times a day. That's a lot of use. It's on 2 billion devices. It is absolutely woven into everyday life."
But what Mr Gruber and long-time colleagues working on artificial intelligence (AI) have seen in the past 18 months has scared them.
"There's something different this time," he says.
"And that something different is that the amount of capabilities that were just uncovered in the last year or two that has surprised the people who were building them, and surpassed all of our expectations at the pace to which these things were uncovered."
'A human trial without consent'
ChatGPT – produced by the Microsoft-funded OpenAI company – is the most well-known of all the new "generative" AI chatbots that have been released.
Trained on the knowledge of the internet and then released to be tested on the public, this new AI has spread at a record pace.
In Australia, it is already causing disruption. Schools and unis have both embraced and banned it. Workplaces are using it for shortcuts and efficiencies, raising questions about whether it will abolish some jobs. IBM's CEO has already said about 8,000 jobs could be replaced by AI and automation.
Microsoft told 7.30 this week that "real-world experience and feedback is critical and can't be fully replicated in a lab".
But Mr Gruber, and thousands of AI industry scientists, coders and researchers, want testing on the public to stop until a framework is put in place to — in his words — "keep this technology safe and on the side of humans".
"What they're doing by releasing it [to the world] is they're getting data from humans about the kind of stuff that normal humans would do when they get their hands on such a model. And they're like, learning by trial and error," Mr Gruber tells 7.30.
"There's a problem with that model — that's called a human trial without consent.
Toby Walsh is the chief scientist at UNSW's new AI Institute. He says another part of the concern is the rate at which ChatGPT is being adopted.
"Even small harms, multiplied by a billion people, could be cause for significant concern," he says.
"ChatGPT was used by a million people in five days, [and] 100 million people at the end of the first month.
"Now, six months later, it's in the hands of a billion people. We've never had technologies before where you could roll them out so quickly."
Blurring the lines between what's real and what isn't
Here's the rub — the new AI models are really good at being fake humans in text. They're also really good at creating fake images, and even the voices of real people.
If you don't want to pay a model for a photo shoot, or you want a contract written quickly without a lawyer, AI is a tool that's at your disposal.
But the new AI apps are also great for fraudsters and those who want to manipulate public perceptions.
CBC Canada's public broadcaster reported that police are investigating cases where AI-trained fake voices were used to scam money from parents who believed they were speaking to their children.
"Just to clarify – it only takes a few seconds now to clone someone's voice … I could ring up your answer phone, record your voice, and speak just like you," Mr Walsh says.
Mr Gruber is scared by how new AI can "pretend to be human really well".
"Humans are already pretty gullible," he says.
"I mean, a lot of people would talk to Siri for a while, mostly for entertainment purposes.
"But there are a lot of people who get sucked into such a thing. And that's one of the really big risks.
"We haven't even seen the beginning of all the ways people can use this amazing piece of technology to amplify acts of mischief.
"If we can't believe our senses, and use our inherited ability to detect whether that thing is fake or real, then a lot of things that we do as a society start to unravel."
'Complete mystery boxes' comparable to sociopaths
AI is not like computer-coded programs whose lines of script can be checked and corrected one by one.
"They're closer to [being] organic," Connor Leahy says.
The London-based coder is at the beginning of his career and already the CEO of his own company Conjecture, which aims to create "safe AI" and is funded to the tune of millions of dollars by venture capitalists and former tech success stories like the creator of Skype.
Mr Leahy's definition of safe AI is "AI that truly does what we want it to do, and that we can rely on them to not do things we don't want them to do".
Sounds simple enough — until he describes the current AI apps.
"They are complete mystery boxes, black boxes, as we would say in technical terms," he says.
"There's all these kinds of weirdness that we don't understand, even with, for example, relatively simple image recognition systems, which have existed for quite a while.
"They have these problems, which are called adversarial examples.
"And what this means is that you can completely confuse the system by just changing a single pixel in an image; you just change one pixel and suddenly [the system] thinks that a dog is an ostrich.
"This is very strange. And we don't really know why this happens. And we don't really know how to fix it."
This "black box" has led OpenAI to develop a tool to help identify which parts of its AI system are responsible for its behaviours.
William Saunders, the interpretability team manager at OpenAI, told industry site TechCrunch: "We want to really be able to know that we can trust what the model is doing, and the answer that it produces."
Each large language model he's referring to is a neural network. And each individual neuron makes decisions based on the information it receives, a bit like the human brain. That neuron then sends its answer to the rest of the network.
OpenAI says their tool could only "confidently" explain the behaviour of just 1,000 neurons out of a total of 307,200 neurons in its GPT-2 system. That's two generations back.
Meanwhile, GPT-4 has an estimated trillion neurons.
Ironically, OpenAI is using GPT-4 to run its tests on GPT-2, which underscores the point that it has released something into the world it barely understands.
Do not harm does not apply
Science fiction writer Isaac Asimov famously wrote the three laws of robotics — the first of which Mr Gruber expands upon: "Robots should do no harm to humans or not cause harm to happen through inaction."
That doesn't apply to AI at the moment because it's not a law it can understand at a conceptual level, "because the AI bot or the language model doesn't have human values engineered into it".
It's a big word calculator.
"It's only been trained to solve this astonishingly simple game of cards [in which each card is a word]," Mr Gruber says.
"It plays this game where it puts a card down and then guesses what the next word is. And then, once it figures that out, you know, OK, it guesses the next word. And so on."
Those words are what it calls a response to a question asked by a human.
"It plays the game a trillion, trillion times — an astonishing amount of scale and computation, and a very simple operation. And that's the only thing it's told to do."
The new generation of AI can mimic human language very effectively but it cannot feel empathy for humans.
"They are a very blunt instrument — we don't know how to make them care about us," Mr Leahy says.
"This is similar [to] how potentially a human sociopath understands that doing mean things is mean, but they don't care. This is the problem that we currently face with AI."
This is all happening now — not in some doomsday future scenario on a Hollywood scale where sentient AI takes over the world.
It's no wonder, then, that so many in the industry are calling for help.
Regulation and fear
Tech insiders are now calling for their wings to be clipped – even in America, where it is almost unheard of that US corporations ask to be regulated.
But that is precisely what happened this week, when the head of OpenAI Sam Altman appeared before the US Congress.
In stunning testimony, the 38-year-old declared: "If this technology goes wrong, it can go quite wrong; we want to be vocal about that, we want to work with the government to prevent that from happening."
Mr Altman has been quite open about his fears, telling ABC America in an interview earlier this year he is "a little bit" scared of AI's capabilities, before adding that if he said he wasn't, he shouldn't be trusted.
Mr Leahy is also outspoken.
"There is currently more regulation on selling a sandwich to the public than there is to building completely novel powerful AI systems with unknown capabilities and intelligence and releasing them to the general public widely onto the internet, you know, accessible by API to interface with any tools they want," he said.
"The government at the moment has no regulation in place whatsoever about this."
The challenge now is how fast safeguards can be installed and whether they are effective.
"It's kind of like a sense of … futurists' whack-a-mole," Mr Leahy told 7.30.
"It's not that there's one specific way things go wrong, and only one way, how unleashing intelligent, autonomous, powerful systems onto the internet that we cannot control and we do not understand ... there's billions of ways this could go wrong."
Watch 7.30, Mondays to Thursdays 7:30pm on ABC iview and ABC TV
