Marathon day. An early prepare into London, then an unfamiliar journey throughout a race-disrupted metropolis from Paddington to Blackheath, all in good time for the beginning of the race. I used to be nervous, after all, however was cheered by the sight of one other bib-wearing runner — extra skilled at marathons, much less accustomed to London.
Me: “How do you intend to get to the beginning line?”
He: “I’ve requested ChatGPT. It says Elizabeth Line to Liverpool Road, then the prepare to Blackheath.”
That didn’t sound correct. Was there a prepare from Liverpool Road to Blackheath? Google Maps and Citymapper steered attending to Blackheath from Charing Cross or Waterloo.
Me: “Are you positive? I’d recommend the Circle or Bakerloo to Charing Cross.”
He frowned for a second and pulled out his cellphone. “No, ChatGPT says that ‘The Circle Line will not be a good selection on marathon day. It will likely be too crowded. There are too many stops and too many steps. It’s a route for vacationers, not for runners.’”
I checked Google Maps. Positive sufficient, there isn’t a prepare from Liverpool Road to Blackheath. ChatGPT’s suggestion would go away him stranded, making an attempt to catch a bus over the marathon route, then making an attempt to get on to the prepare from Charing Cross at a busy London Bridge. I advised him that gave the impression of a foul thought. He frowned once more and typed one other question into his cellphone. “Oh, you’re proper. ChatGPT says, ‘Correction: take the Elizabeth Line straight to London Bridge.’”
Me: “The Elizabeth Line doesn’t go to London Bridge.”
You’ve heard tales of synthetic intelligence hallucinations earlier than, but it surely’s not the AI that fascinates right here: it’s the human.
The route-finding algorithm on Google Maps is a minor miracle. It can resolve a fancy optimisation drawback throughout a number of modes of transport, considering real-time congestion or delays, and it’s been out there on smartphones and browsers for years. It’s a confirmed, sensible instance of AI in motion. So on marathon day, when the stakes are excessive and the clock is ticking, why would anybody flip as an alternative to a flowery word-guessing machine similar to ChatGPT?
Maybe it’s that ChatGPT appears so human. It served up an uncanny impersonation of a pleasant and educated native information. The Circle Line? Pfft, it’s effective for vacationers however you’re a marathon runner: take into consideration all these steps! (It’s true, the creaky outdated Circle Line does have steps.)
A part of the bot patter jogged my memory of clickbait advertisements: INSURANCE COMPANIES HATE THIS LOOPHOLE! ChatGPT wasn’t simply giving a route, however giving a rationale, even explaining why we shouldn’t hearken to the lamestream recommendation of Google Maps. That is the strategy of a confidence trickster.
Within the introduction to her e book The Confidence Recreation, psychologist Maria Konnikova explains: “The true con artist doesn’t pressure us to do something: he makes us complicit in our personal undoing . . . we imagine as a result of we need to.” One distinction between the con artist and the big language mannequin (LLM) is that the con artist is aware of the reality and is making an attempt to hide it. One similarity between the con artist and the LLM is that each of them have perfected seeming believable.
A latest paper in Nature finds that when LLMs are skilled to be heat and pleasant, additionally they produce dramatically much less correct solutions, “selling conspiracy theories, offering inaccurate factual data and providing incorrect medical recommendation”.
That sounds unhealthy. I’d recommend that the truth is worse: the sycophantic AI not solely produces errors, it persuades us to imagine them.
In 1950 Alan Turing, the mathematician and visionary of the pc age, famously proposed an “imitation sport” during which a human decide would talk by means of a teleprompter with a human and a pc. The pc’s job was to mimic human dialog convincingly sufficient to steer the decide.
Turing’s check stays intriguing, however there’s a longstanding issue: the fallibility of the decide. A primitive Sixties chatbot, Eliza, responded like a parody of a therapist (“How does that make you’re feeling?” “Why do you’re feeling unhappy?” “Please go on.”). Individuals lapped it up; it’s good to really feel listened to. A Nineteen Eighties chatbot, MGonz, simply fired off insults and was completely believable, partly as a result of insults are easy to ship and largely as a result of they immediate rage fairly than reflection within the human recipient. And Robert Epstein, an skilled within the Turing Check, has written entertainingly about how he was fooled right into a four-month correspondence with a horny Russian girl who was, in truth, a 2006-era chatbot. None of those bots had a thousandth of the sophistication of a contemporary LLM, however they didn’t want it: when people are unhappy, offended or amorous, we aren’t very subtle judges, both.
We’re all going to seek out ourselves in unusual variations of the Turing Check in years to return, and I ponder if we’re as much as it. And never simply us, however these with energy over us. As Cory Doctorow, writer of Enshittification, is keen on observing: you gained’t get replaced as a result of an AI can do your job, you’ll get replaced as a result of an AI salesman convinces your boss that it could. If my journey to the marathon begin line is any information, that salesman may have a straightforward job.
The capabilities of recent AI are spectacular. However what determines whether or not we use it isn’t the potential, however the impressiveness. They’re correlated however they don’t seem to be the identical factor. There’s a story concerning the French poet Jacques Prévert seeing a fellow begging for change on the streets of Venice with an indication that learn “Blind man with out a pension”.
Prévert stopped to speak to him; not many individuals had been moved to contribute, and Prévert supplied to put in writing a brand new signal.
The following day, he returned to seek out the person overjoyed. “It’s unbelievable; I’ve by no means acquired a lot cash in my life.”
Prévert had written: “Spring is coming, however I gained’t see it.”
The brand new signal contained no information — in truth, it was much less informative than the outdated. Nevertheless it advised a narrative. Google Maps was the primary signal: it advised me the place to get my prepare. ChatGPT was the second signal: it advised my companion not simply the place to go, however the best way to really feel about taking such a intelligent route.
I left him at Paddington, urging him to not attempt to take the non-existent Elizabeth Line prepare to London Bridge. I’m not positive I used to be as convincing as ChatGPT.
I ran the London Marathon in assist of the Teenage Most cancers Belief – not too late to make a donation.
Written for and first revealed within the Monetary Occasions on 6 Could 2026.
