UNTIL Microsoft curtailed the capabilities of its Bing chatbot – codenamed Sydney and powered by an advanced version of OpenAI’s ChatGPT model – there were a chaotic few days last month when it was , cajoling, and terrifying its beta testers.
Even journalists who regularly write about artificial intelligence expressed surprise: they know these programs are just statistical models of the language on the internet, but they still found unsettling and eerily human.
Advertisement
Bing’s chatbot has yet to be rolled out to the world at large and its curtailment has prevented it from going off the rails again, but it remains unnerving. If the original Sydney were introduced to a more unfamiliar audience, what actions might it convince people to take, and might they believe it is something more than a statistical language model – a mistake even experienced professionals sometimes make?
It has long been standard practice to imbue chatbots with a human-like personality, from ELIZA, a simple chatbot designed to emulate the questioning style of a Rogerian psychotherapist that was created by computer scientist Joseph Weizenbaum at MIT in 1964, to Microsoft’s more recent Tay chatbot in 2016, which was quickly taken offline after being tricked into racist rants.
It now seems a given for chatbots’ responses to resemble those of humans, from which problems constantly seem to spew: a Google engineer thought the firm’s LaMDA AI was sentient, for example, and ChatGPT has had its fair share of difficulties.
These human imitators can also disguise what is really happening under the hood. Murray Shanahan, a researcher at DeepMind in London, recently spelled out exactly how these AIs work in a preprint titled . He reiterates that, despite how advanced their plumbing has become and the emergent abilities they display, all they are doing is predicting the most likely subsequent word in a sequence based on all the relationships of words they have seen.
We can convince ourselves that these AI models are something more, even when we know that isn’t the case
When such a chatbot leads people to believe it is something more than that, , it is a form of deception on the part of the chatbot. Not only could this fool people into trusting its capabilities more than they should, but it might also play on people’s sympathies for what they feel to be their fellow human, coercing them into acts they might not otherwise do.
AI researchers have suggested explicitly disclosing that a chatbot is non-human to improve trust between human and AI, but a by Nika Meyer at the University of Göttingen in Germany and her colleagues revealed a problem with this approach. If a chatbot succeeds in its goals, people are less likely to trust the outcome if they know it is an AI model. If it fails to help, on the other hand, then people tend to be more sympathetic, knowing it is “just” an AI.
So it appears that, in some cases, we might have to make a trade-off between the psychological discomfort we feel from taking advice from an AI and the risks we take on by regarding a chatbot as being like a human and deceiving ourselves. We can try to protect ourselves by remembering that it is a deception, but, as experiences with Bing’s Sydney have shown, we can still end up convincing ourselves that these AI models are something more, even when we know that isn’t the case.
While the convention is still to make chatbots converse like humans, there is a growing movement to reconsider this. Linguist Emily Bender at the University of Washington in Seattle and her colleagues wrote a in 2021, “On the dangers of stochastic parrots”, in which they argue that the mimicking ability of large language models is one of the key factors in their potentially harmful downstream effects. “We call on the field to recognise that applications that aim to believably mimic humans bring risk of extreme harms,” they wrote.
For those who talk about a rapidly approaching doomsday AI future, humans being duped by convincing AI chatbots may feel like an inevitability. But there are many intentional choices when designing a large language model, like the personality with which it speaks, or whether it has one in the first place.
These choices should be made at a societal level, not in the backrooms of a handful of rich tech companies, such as OpenAI and Google. While companies like Stability AI have made open-source alternatives to text-to-image AIs, there is still no open-source version of something like ChatGPT. Until there is, our tendency to see human attributes where they don’t exist may lead us into murky waters.
Alex Wilkins is a Â鶹´«Ă˝ reporter covering artificial intelligence, physics and space. Artificially intelligent isĚýaĚýcolumn that cuts through the hype and looks at what AI is really capable of and what itĚýmeans for us. You can follow him @AlexWilkins22
Alex’s week
What I’m reading
The Earthsea Quartet by Ursula K. Le Guin – it is nice to live in another world for a little while where there are no chatbots to be seen.
What I’m watching
Welcome to Wrexham, a traipse through Welsh history and UK football culture with a sprinkling of Hollywood magic.
What I’m working on
After being struck down with a winter lurgy, I have been trying to get back on the bicycle in preparation for an upcoming off-road trip.
Ěý
