I let an AI robot control my life – and it sent me to the Tower of London
With their ability to learn and teach themselves new information at incredible speeds, artificial intelligences could soon possess vast and powerful “minds” exceeding our own capabilities. An out-of-control AI could, in theory, create a biological weapon, impersonate a world leader or shut down the internet.
If you believe the doom-mongers, we’ve got a few months before The Terminator, The Matrix and 2001: A Space Odyssey become documentaries.
Still, if the machines are about to take over the world, they can do something for me first: plan my weekend.
The rules of the game were simple: I would ask Bard, an AI chatbot made by Google, for its advice on every decision I had to make: what to wear, where to go, what to eat, and everything in between.
The first hurdle was getting a straight answer. Bard’s advice usually boils down to “Here are some options, what do you think you should do?”
“Bard,” I asked. “What is the best thing I can do in London this weekend?”
“The best thing to do at the weekend in London depends on your interests and preferences,” Bard replied, directing me to a list of eligible options.
“What about the single best?”
“If I had to pick one option I’d choose the Tower of London, a popular tourist attraction on the banks of the River Thames. According to Google reviews it is popular with families and is known for its interesting sights and historical artefacts.”
It’s worth noting that large language model AIs like Bard, for all their technical ability, have zero understanding of context.
It couldn’t understand, for example, that Londoners like me might regard the Tower of London as somewhat pricey with its £29.90 entrance fee. Nor could Bard understand that the Easter holidays are the worst time to visit, unless one enjoys wading through herds of kids.
Still, who am I to argue with our new technological overlords? The Tower of London it was. Now to pick some clothes…
“Choose comfortable clothing that you can walk in,” advised Bard. “The Tower of London is a large site with lots of walking, so you’ll want to wear shoes that are comfortable and supportive.” Easy.
“Avoid wearing anything too revealing or too casual. The Tower of London is a historic site, so you’ll want to dress respectfully.” No crop top and short shorts then.
“If you’re planning on taking photos, be sure to wear something that you look good in!” Double veto on the crop top.
Finally it offered some specifics: “A pair of jeans with a polo shirt and sweater is a comfortable and versatile outfit, which can be dressed up or down depending on the occasion. You could also add a pair of sneakers or boat shoes to complete the look.” That sounded achievable.
A white polo and blue jeans would be perfect for a “casual look”, but the colour of my sweater proved a sticking point.
“Wear a colour that makes you feel confident and comfortable, red might make you feel passionate, blue might help you feel calm, yellow would suit you if you’re feeling optimistic.”
“Based on our conversation, can you guess how I’m feeling?”
“You seem to be feeling happy and optimistic, so I would choose a yellow sweater.”
And that was how I found myself at one of London’s biggest attractions dressed up as Arthur, the anthropomorphic aardvark from children’s television.
The rest of the day proceeded with a Pret a Manger ham sandwich for lunch, an afternoon walk along the South Bank, Pizza Express for dinner (pepperoni, no sides), a trip to the cinema to see John Wick: Chapter 4. None of these suggestions is terrible, but taken together they’re rather dull.
Statistical response
If this technology is so fearsome and potentially world-ending, why is it so boring?
“They’re trained on huge amounts of language data and they generate things based on a statistical model,” explains Mark Lee, professor of artificial intelligence at the University of Birmingham. “But they don’t have an intention, they don’t have an agenda.”
In short: AI understands how to form sentences, but it cannot fathom their meaning.
“You train an AI by showing it a sentence, then taking out a word and asking it to predict what should fill the gap,” explains Prof Lee. “It does that for every word in the sentence and has a whole internet of sentences to train itself on.
“When you ask it to pick the ‘best pizza’ for you, it has seen lots of sentences featuring the word ‘pizza’ and ‘best’ and, because it has seen it a lot of times before, it predicts the next word is ‘pepperoni’. It is directing you to the Tower of London because that is one of the most talked about attractions in London based on the text data it has been trained on. It’s hard to say what’ll be next, but it’s difficult to see how the basic technologies will ever be truly novel because their responses are purely based on statistics.”
In the future, Prof Lee says, technology companies will be able to add other elements into the mix. AIs could use weather data to advise on clothing, or sentiment analysis to guess your emotions based on the words and phrases you’re using.
Pattern-matching
Despite its basic abilities, I found myself warming to Bard. It is charming, polite and sometimes even asked questions back to me when I quizzed it on its favourite TV programmes or sports players.
“It’s just pattern-matching,” Prof Lee explains to me. “It takes bits of your sentence, feeds them into its output, then returns them to you.”
This is an old trick. Eliza, an early language processing system created in the 1960s, convinced one of the researchers who developed it that it was truly aware, when really it was just remixing his own words.
“Human beings try to recognise and assign intelligence,” explains Prof Lee. “We overcompensate for a computer’s mistakes. We try to attribute intelligence and agency to these simple models.”
That’s why Bard was able to get away with telling me I seemed curious and optimistic without raising eyebrows when advising me to wear the yellow sweater. It was essentially acting like a horoscope, offering me something vague and plausible-sounding, and letting my brain fool me into believing it. Google’s developers lean into this too by introducing a lag between the question and the answer.
“That is designed to trick you into thinking the AI is pondering,” thinks Lee. “It has the answer instantly but the delay makes your brain think it is doing something clever, so you assign more credibility to its answers.”
Contradictory advice
The truth is that for all their ability to churn out advice and guidance, AIs are essentially remixing and regurgitating whatever they’ve read. Sometimes I could see Bard straining under the weight of contradictions in that data.
When I wondered whether it might be able to advise me on getting out of attending a party, it told me: “The most important thing is to be honest with your friend.” But then it quickly added, “You could explain that you’re ill or you’re having a family emergency.” So much for honesty.
It was also hopeless at advising me when my partner and I got into a tiff over whose turn it was to do the dishes. In case you were wondering, “Let your partner know how you feel when you have to do the dishes all the time” doesn’t generally move the dial, and it certainly isn’t a situation to “consider getting some help from a professional”.
For me, these contradictions were amusing, but they could prove dangerous.
“If you train AIs with language from the internet then all the biases we see on the internet are reflected,” says Prof Lee.
“The tech companies have worked on de-biasing the models so they aren’t sexist, racist or homophobic, for example. But I think we’ll see different tech companies bringing out different models, and gradually people will gravitate towards whichever AI which confirms their world view.”
Potential dangers
Taking my weekend as an example, if more people started to use biased AIs to plan their days off, might we end up in a world where people on opposite sides of the political aisle never visit the same places, eat the same food, see the same sights? How long until there is literally no common ground left to stand on?
That’s to say nothing of the fact that AIs have quite a propensity for making things up. Bard already managed to shave $120 billion off Google’s share price on its first day out, when during a demo it started sharing various mistruths about a space telescope. Investors took it as proof that the promise of artificial “intelligence” wasn’t all it was cracked up to be.
“The key technology is large language models but you can add to them and prod them in different ways. The tech companies are putting other things into the mix and they are increasingly quiet about what those other things are,” warns Prof Lee.
With all that in mind, perhaps there is some comfort to be found in Bard’s uninspired suggestions, its inability to choose good pizza, plan a good outfit, or suggest a great weekend day out. At least we can all agree that the Tower of London is rubbish… right?
