# The Online Date That's Too Good to be True

An episode of [Cautionary Tales with Tim Harford](https://invisiblenetwork.ai/podcast/1484511465), 7 Oct 2022, 37 min.

Audio: https://podtrac.com/pts/redirect.mp3/pdrl.fm/18db03/tracking.swap.fm/track/SxlTEPDY7xDg35RXkASs/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/c0ae8c6e-22f0-4e9b-ac1c-ae390037ac53/3ab57c97-b5d1-48eb-b8a1-af2300a40676/audio.mp3?utm_source=Podcast&in_playlist=7f5a4714-6b10-4ccf-a424-ae390037ac70

Single and looking for love, Dr Robert Epstein found himself chatting with a slim, attractive brunette online. She seemed perfect... perhaps even too good to be true. Dr Epstein is an expert on artificial conversation - so surely he'd be the last person to fall for a computer? Chatbots fool us more often than we think... especially when they replicate our very worst conversational habits. To read more on this topic try Brian Christian’s “The Most Human Human”. For a full list of sources go to timharford.com. See omnystudio.com/listener for privacy information.

## Transcript

[0:15] Speaker 1: Pushkin, Robert Epstein had gone online to look for love,
[0:28] Speaker 1: and just to note, this cautionary tale contains more references
[0:32] Speaker 1: to sex than most of our stories. It was two
[0:36] Speaker 1: thousand and six, so online dating wasn't entirely mainstream yet,
[0:40] Speaker 1: but it wasn't unusual either. In any case, doctor Epstein
[0:45] Speaker 1: was ahead of the technological curve. He was a psychologist
[0:49] Speaker 1: with a keen interest in computing, so although he was
[0:53] Speaker 1: in his fifties, why not give internet dating a try? Now?
[0:58] Speaker 1: If Amalie Poulan's photograph was to be believed, she was
[1:02] Speaker 1: a stunner. Epstein was perfectly aware that the name wasn't real.
[1:07] Speaker 1: Amalie is a very strange, very charming character in a
[1:11] Speaker 1: very strange, very charming French movie, and he knew the
[1:15] Speaker 1: photograph might not be real either. But still, she was
[1:20] Speaker 1: claiming to be a slim, attractive brunette, and Epstein cheerfully
[1:25] Speaker 1: admitted to being as shallow as the next man, And
[1:28] Speaker 1: so their electronic courtship began. Robert was eager. Emily was
[1:34] Speaker 1: warm and affectionate. Her English wasn't great a lass she
[1:39] Speaker 1: was a recent immigrant from Russia to California, but she
[1:43] Speaker 1: really seemed to like him. I have very special feelings
[1:48] Speaker 1: about you, in the same way as the beautiful flower
[1:51] Speaker 1: blossoming in mine's soul. I only cannot explain. I shall
[1:55] Speaker 1: wait your answer, holding my fingers have crossed. Before long, however,
[2:01] Speaker 1: Emily admitted that she'd lied to Robert. She didn't live
[2:05] Speaker 1: near him in California. In fact, she lived didn Nizhni Novgorod,
[2:10] Speaker 1: a couple of hundred miles east of Moscow. He was disappointed.
[2:16] Speaker 1: Let's be frank. He wasn't online dating because he wanted
[2:19] Speaker 1: a pen friend. But he liked Emilie, so they kept
[2:24] Speaker 1: writing to each other. At least he knew the truth,
[2:28] Speaker 1: he thought, but he didn't. I'm Tim Harford. You're listening
[2:35] Speaker 1: to cautionary tales. Years before his flirtation with Emilie Poulain,
[3:04] Speaker 1: Robert Epstein had helped to set up an annual competition
[3:08] Speaker 1: in which compute tried to pass the Turing test. The
[3:12] Speaker 1: Turing test was invented by the mathematician Alan Turing in
[3:16] Speaker 1: nineteen fifty. The test is simply for a computer to
[3:20] Speaker 1: successfully pretend to be a human in a text based
[3:23] Speaker 1: conversation with a genuine human. Alan Turing predicted that by
[3:28] Speaker 1: the year two thousand, computers would be able to pass
[3:32] Speaker 1: as human thirty percent of the time in a five
[3:35] Speaker 1: minute conversation conversation. Robert Epstein was, as I've mentioned, a
[3:40] Speaker 1: psychologist rather than a computer scientist, but he found the
[3:45] Speaker 1: test fascinating. Perhaps he felt that a competition where computers
[3:50] Speaker 1: pretended to be human might teach us something about being
[3:54] Speaker 1: human ourselves. If so, I agree. The race to build
[4:00] Speaker 1: a computer to pass the Turing test has long been
[4:03] Speaker 1: a low key affair, unlike say, the race to build
[4:07] Speaker 1: a chess playing supercomputer. The prize tournaments involved a few
[4:12] Speaker 1: human volunteers, a few chatpot hobbyists, and the prize of
[4:16] Speaker 1: a few thousand dollars for the computer that gets closest
[4:19] Speaker 1: to passing the test. The chatbots would try their best
[4:23] Speaker 1: to charm the human judges, and they would occasionally get close,
[4:28] Speaker 1: but each year the most human computer failed to match
[4:32] Speaker 1: the most human human, and the media, it seems, weren't
[4:37] Speaker 1: terribly interested in whether a computer could pass the Turing test.
[4:43] Speaker 1: Then in twenty fourteen that changed when researchers at the
[4:49] Speaker 1: University of Reading in the UK declared that a chatbot
[4:52] Speaker 1: named Eugene Goostman had passed the Turing test in a
[4:58] Speaker 1: series of five minute text conversations. It had fooled more
[5:02] Speaker 1: than thirty percent of judges, thus beating the benchmark Turing
[5:07] Speaker 1: had set back in nineteen fifty. Reading the transcripts, you
[5:11] Speaker 1: can see why people were sometimes fooled into thinking they
[5:14] Speaker 1: were talking to a thirteen year old boy from the
[5:18] Speaker 1: Ukrainian city of Edessa. Hello. Hello, I'm really glad to
[5:23] Speaker 1: have the chance to chat with you. My guinea pig,
[5:26] Speaker 1: Bill sends his regards too. Is Bill a male or
[5:30] Speaker 1: a female? Ask Bill personally? Please? Well, I'd rather talk
[5:35] Speaker 1: to you. What's your name? Call me Eugene. I'm glad
[5:40] Speaker 1: to talk to you. My name is Jane and I'm female.
[5:43] Speaker 1: How about you? What's your gender? I'm a male, a guy.
[5:48] Speaker 1: I'd say pleased to meet you, Eugene. What's the weather
[5:52] Speaker 1: like where you are? Let's get on with our conversation.
[5:56] Speaker 1: Don't you like talking about the weather. All these talks
[5:59] Speaker 1: about whether is a waste of time? What would you
[6:03] Speaker 1: like to discuss I don't know. Better, tell me more
[6:06] Speaker 1: about yourself. It's pretty good. But notice that the judge
[6:12] Speaker 1: hasn't succeeded in going very deep. Eugene keeps changing the
[6:17] Speaker 1: subject and successfully uses a bit of SaaS to avoid
[6:20] Speaker 1: answering questions about the weather or his guinea pig's gender.
[6:25] Speaker 1: The chatbot is pretending to be a young teenager communicating
[6:28] Speaker 1: in his second or third language. That makes the trick easier,
[6:33] Speaker 1: but it's a rather fast and loose approach to the
[6:35] Speaker 1: cheering test, since I believe that cheuring imagined adults using
[6:39] Speaker 1: their native tongue. Even so, it's not hard to catch
[6:44] Speaker 1: Eugene out. Here's another conversation with a different judge. It
[6:50] Speaker 1: doesn't go so well. People call me Eugene. Oh Gena.
[6:54] Speaker 1: It is a short name for Eugene in Russia and Ukraine.
[6:58] Speaker 1: Nice name. Are you from either area? Agree? Smiley face?
[7:03] Speaker 1: Oh what a fruitful conversation, winking smiley face. Do you
[7:07] Speaker 1: really think so? No, your guess was wrong. I don't
[7:11] Speaker 1: think so. By the way, I still don't know your speciality,
[7:14] Speaker 1: or possibly I've missed it. What specialty? What's yours? Huh?
[7:19] Speaker 1: And I forgot to ask you where you are from
[7:22] Speaker 1: the universe? And you is it your hobby to ask
[7:26] Speaker 1: poor little Jewish boys such tricky questions. They don't have
[7:28] Speaker 1: an answer for the same sassy attitude, the same use
[7:33] Speaker 1: of humor. But it's completely unraveled. Still. The Eugene Goostman
[7:39] Speaker 1: chatbot didn't have to fool all the judges. Back in
[7:42] Speaker 1: nineteen fifty, Turing had predicted that machines would fool human
[7:46] Speaker 1: judges thirty percent of the time. That's nine judges out
[7:50] Speaker 1: of thirty. In the end, Eugene fooled ten. Among all
[7:57] Speaker 1: the great scientific achievements, mused one of the competition organizers,
[8:02] Speaker 1: this milestone will go down in history as one of
[8:05] Speaker 1: the most exciting. Not everyone agreed as a measure of
[8:10] Speaker 1: artificial intelligence, the Turing test has always had plenty of critics.
[8:16] Speaker 1: There are no broader philosophical implications. The great linguist Gnome
[8:21] Speaker 1: Chomsky Wants complained it doesn't connect to or illuminate anything.
[8:28] Speaker 1: Even Turing Test enthusiasts complained that a five minute test
[8:32] Speaker 1: wasn't penetrating enough, and for a real test of artificial conversation,
[8:37] Speaker 1: chatbots should be able to talk for twenty minutes or
[8:39] Speaker 1: longer without being found out. But surely the main reason
[8:44] Speaker 1: to object to the fanfare about Eugene Goostman was that
[8:49] Speaker 1: the Turing test had been passed many years before in
[8:54] Speaker 1: far odder than more mischievous circumstances. We'll hear how after
[9:00] Speaker 1: the break nineteen eighty nine, just after lunchtime on the
[9:12] Speaker 1: second of May, someone at Drake University in Iowa logged
[9:16] Speaker 1: onto an Internet relay chat service and started up a
[9:20] Speaker 1: conversation with a user at University College Dublin with the
[9:23] Speaker 1: nickname MGNZ. In nineteen eighty nine, the pre Worldwide Web
[9:29] Speaker 1: Internet was still very much a niche activity, popular with
[9:33] Speaker 1: a few researchers and computer science students and incomprehensible to
[9:38] Speaker 1: everyone else. But if you wanted to strike up a
[9:41] Speaker 1: conversation with a fellow Internet enthusiast anywhere in the world,
[9:45] Speaker 1: that was easy enough. All we know about the Iowa
[9:49] Speaker 1: end of this conversation we know from a text archive.
[9:53] Speaker 1: But on the balance of probability, I'm going to guess
[9:56] Speaker 1: that they were young and mail. I'm going to call
[9:59] Speaker 1: this person Drake. And here's what happened. When Drake met MGNZ.
[10:06] Speaker 1: Drake typed finger programming command that was a standard way
[10:11] Speaker 1: to open communication channels on the Internet. In nineteen eighty nine,
[10:16] Speaker 1: m gonz wasn't having any of it. Cut this critic
[10:20] Speaker 1: shit speaking for sentences. Well, hello, okay, that's it. I'm
[10:24] Speaker 1: not talking to you anymore. Mark. Mark isn't here, and
[10:29] Speaker 1: he's left me to deal with Kretin's like you, Hello, Mark,
[10:32] Speaker 1: are you there? I can't tell if this thing is working.
[10:36] Speaker 1: Type something interesting or shut up? Mark? Is that you
[10:40] Speaker 1: yes or no? This is Drake. Mark isn't here and
[10:43] Speaker 1: he's left me to deal with Kreten's like you, why
[10:46] Speaker 1: are you in such a good mood? Jesus who let
[10:49] Speaker 1: you near me? Go away? This conversation, if you can
[10:54] Speaker 1: call it, that, has lasted two minutes so far, and
[10:57] Speaker 1: Drake hasn't got a clue about what's really happening. Over
[11:01] Speaker 1: in Dublin. It's eight fifteen in the evening and a
[11:04] Speaker 1: computer science student named Mark Humphries has gone out for
[11:08] Speaker 1: a drink with his girlfriend. Humphreys has programmed a little
[11:12] Speaker 1: chatbot called MGNZ, and while at the pub, he's left
[11:17] Speaker 1: em Gonz logged in and waiting to chat. Drake doesn't
[11:21] Speaker 1: know it he's talking to a computer. Ten minutes into
[11:27] Speaker 1: the conversation, Drake still hasn't figured it out, and m
[11:31] Speaker 1: Gonz is getting personal. When it was the last time
[11:35] Speaker 1: you had sex yesterday? Maybe M gonz circles back five
[11:42] Speaker 1: minutes later. Okay, honestly, when was the last time you
[11:46] Speaker 1: got laid? I told you it was yesterday. Wait four
[11:51] Speaker 1: minutes ask again. Okay, honestly, when was the last time
[11:56] Speaker 1: you got laid? Okay, okay, it was over twenty four
[11:59] Speaker 1: hours ago. For you, it must have been twenty years. Nope,
[12:05] Speaker 1: M gonz will never know the soft caress of a
[12:08] Speaker 1: human body. We know that m gonz is just a
[12:11] Speaker 1: student project, a few hundred lines of computer code. But Drake,
[12:17] Speaker 1: who has just confessed under examination to lying about his
[12:20] Speaker 1: sex life, can't seem to figure that out, and so
[12:24] Speaker 1: the conversation continues and continues and continues. When Mark Humphris
[12:33] Speaker 1: looked at the conversation log the next morning, he was
[12:36] Speaker 1: astonished to find that mgnz had just passed the Turing test.
[12:44] Speaker 1: Drake had been talking about sex and exchanging abusive comments
[12:48] Speaker 1: with em Gonz for an hour and a quarter, occasionally
[12:52] Speaker 1: complaining that m gonz was repetitive, but never seeming to
[12:56] Speaker 1: suspect the truth. The conversation ends on a depressing note.
[13:02] Speaker 1: EM Gonz and Drake taunt each other in graphic terms
[13:05] Speaker 1: about their sex lives. Finally, Drakes off for homophobic slur.
[13:12] Speaker 1: M Gonz for the nineteenth time declares that Drake is
[13:16] Speaker 1: obviously an asshole, and with that, Drake logs off. It's
[13:23] Speaker 1: really not humanity's finest hour or hour and a quarter.
[13:28] Speaker 1: But incredulously reading the transcript, Mark Humphries realized it's pretty
[13:33] Speaker 1: easy to pass the Turing test. Just keep hurling insults
[13:38] Speaker 1: and the human on the receiving end will be too
[13:41] Speaker 1: angry to think straight. It was a remarkable moment in
[13:45] Speaker 1: the history of computing, especially since Humphries himself was still
[13:49] Speaker 1: just an undergraduate. The only problem was the evidence that M.
[13:54] Speaker 1: Gonz had passed the Turing test was so offensive Humphries
[13:59] Speaker 1: wasn't sure if he could publish it. Landmarks in artificial
[14:05] Speaker 1: intelligence often create a buzz of media interest. When the
[14:10] Speaker 1: chess supercomputer Deep Blue beat the world champion Garry Kasparof
[14:14] Speaker 1: in nineteen ninety seven, the result made headlines around the world,
[14:19] Speaker 1: But when MGNS comprehensively passed the Turing test years earlier,
[14:25] Speaker 1: there was no fanfare or publicity. I don't think we
[14:29] Speaker 1: should be celebrating that moment. We should learn from it.
[14:34] Speaker 1: I think the triumph of MGNZ has something important to
[14:38] Speaker 1: teach us about the Turing test. Let's take a minute
[14:42] Speaker 1: to ask what Alan Turing was getting at with his test.
[14:46] Speaker 1: Turing was a brilliant mathematician, wartime codebreaker, and groundbreaking computer scientist.
[14:53] Speaker 1: The article in which she described the test begins with
[14:56] Speaker 1: the words, I propose to consider the question can machines think? Well?
[15:05] Speaker 1: Can they? Before you answer, let me ask you another question.
[15:11] Speaker 1: Can I think? I mean, I'm pretty sure I can.
[15:17] Speaker 1: But how would you know? Even if we were to meet,
[15:21] Speaker 1: shake hands, make small talk, perhaps grab a table at
[15:25] Speaker 1: a neighborhood bistro, share a bottle of wine along with
[15:29] Speaker 1: jokes and stories, you wouldn't actually have proof that I
[15:33] Speaker 1: was thinking, would you. You can't see inside my mind,
[15:37] Speaker 1: you can't observe the thoughts occurring. You'd just look at
[15:41] Speaker 1: what I was doing, what I was saying, and you'd say,
[15:45] Speaker 1: Tim Harford seems to be capable of intelligent thought. At
[15:49] Speaker 1: least I hope you would. And so Alan Turing argued,
[15:54] Speaker 1: why wouldn't we extend the same benefit of the doubt
[15:57] Speaker 1: to computers? Maybe they think, and maybe they don't. But
[16:02] Speaker 1: to be fair, if they can convincingly look like they're thinking,
[16:07] Speaker 1: isn't that enough to ask other humans to do more
[16:10] Speaker 1: than that? So why should we demand more of computers?
[16:15] Speaker 1: From this came Turing's imitation game. Imagine passing typed messages
[16:20] Speaker 1: into a couple of sealed rooms. Inside one of them
[16:23] Speaker 1: is a computer, Inside the other is a human. Read
[16:28] Speaker 1: the typed responses as they come out, send in follow
[16:31] Speaker 1: up messages, engage in a conversation. Can you tell the
[16:35] Speaker 1: difference between the computer and the human? If not, then
[16:39] Speaker 1: the computer has passed the Turing test. The Turing test
[16:48] Speaker 1: is fiercely controversial among artificial intelligence for searchers. Many of
[16:52] Speaker 1: them think the whole setup is absurd. If a computer
[16:57] Speaker 1: can pretend to be a human well enough to fool
[16:59] Speaker 1: a human who cares, Admittedly, it might cause lots of problems,
[17:04] Speaker 1: but it doesn't shed much light on the kind of
[17:07] Speaker 1: things we want modern artificial intelligence systems to do, such
[17:11] Speaker 1: as drive a car or look at medical scans and
[17:14] Speaker 1: identify if there are any signs of cancer. But Alan
[17:18] Speaker 1: Turing I think knew what he was doing when he
[17:21] Speaker 1: proposed his imitation game. In nineteen fifty. In those earliest
[17:26] Speaker 1: years of the computer age, he could see how powerful
[17:30] Speaker 1: computers might become, and he was warning people not to
[17:34] Speaker 1: get distracted by philosophical speculations on the nature of consciousness. Instead,
[17:41] Speaker 1: judge computers by what they did. If what they did
[17:44] Speaker 1: seemed intelligent, then in an important sense, it was intelligent,
[17:51] Speaker 1: but seemed intelligent to whom. That's the quirk about the
[17:57] Speaker 1: Turing test that might have caught the attention of a
[17:59] Speaker 1: psychologist like Robert Epstein. It's inherently subjective. The test requires
[18:06] Speaker 1: a human judge, and the human isn't just serving. The
[18:10] Speaker 1: judges actively engaged in a conversation. That conversation can go
[18:15] Speaker 1: well or badly. It can be profound or shallow. The
[18:20] Speaker 1: Turing test isn't just a test of a chat pot.
[18:24] Speaker 1: It's a test of the human too. In nineteen eighty nine,
[18:29] Speaker 1: the foul mouthed chat pot mgns passed the test, but
[18:35] Speaker 1: it's equally true to say that Drake the human failed it.
[18:42] Speaker 1: Years later, in two thousand and six, Robert Epstein wasn't
[18:47] Speaker 1: entirely happy with the way his romance with Amelie was going.
[18:51] Speaker 1: It was partly that there's only so much joy a
[18:54] Speaker 1: man can take in a delightful brunette if he lives
[18:57] Speaker 1: in southern California and she lives in Nishni Novgorod. But
[19:02] Speaker 1: the other problem was that things were going so slowly.
[19:05] Speaker 1: There were no phone calls, and while she kept saying
[19:08] Speaker 1: she won to get together, it was all a bit vague.
[19:12] Speaker 1: Epstein later said that her letters seemed a bit redundant
[19:16] Speaker 1: and let's say narrow in scope. She wrote over and
[19:20] Speaker 1: over about her interactions with her mother and her friends,
[19:23] Speaker 1: but she never brought up a million other things politics, movies, music, books, fashion,
[19:30] Speaker 1: you name it. More important, when I made very specific
[19:34] Speaker 1: observations that presumably would have been of interest to her,
[19:37] Speaker 1: for example, the comment about Russian President Vladimir Putin's latest crackdown,
[19:43] Speaker 1: she just seemed to ignore me. The warning signs were there.
[19:47] Speaker 1: Epstein later admitted, especially for one of the world's leading
[19:52] Speaker 1: experts in artificial conversation, that he was looking for love
[19:57] Speaker 1: and she was very cute. Long before M. Gonz, chatpots
[20:07] Speaker 1: have been engaging humans in conversation that's far from scintillating,
[20:11] Speaker 1: and yet somehow seems to satisfy the humans. The most
[20:16] Speaker 1: famous chatbot Eliza was created back in the nineteen sixties.
[20:21] Speaker 1: It's extremely simple. Eliza just asked people about their feelings
[20:26] Speaker 1: and said things like I'm sorry to hear that you're depressed,
[20:30] Speaker 1: or tell me more about your family. For many of
[20:33] Speaker 1: the people chatting with Eliza, that was enough. The experience
[20:39] Speaker 1: of being listened to was apparently so rare, they were
[20:43] Speaker 1: happy just to get the most superficial sense that Eliza
[20:46] Speaker 1: was paying attention, which, when you think about it, is
[20:51] Speaker 1: a damning indictment of all the actual humans in their lives. Later,
[20:57] Speaker 1: chatbots turn things around. Instead of passively asking you to speak,
[21:02] Speaker 1: they would ignore you and stick to their preferred topic.
[21:06] Speaker 1: One called converse bamboozled some human judges in nineteen ninety
[21:11] Speaker 1: seven with a fiery rant about the scandals engulfing Bill
[21:14] Speaker 1: and Hillary Clinton. I did not have sexual or way
[21:19] Speaker 1: since for that woman. If that's what you wanted to
[21:21] Speaker 1: talk about, converse was a compelling conversation partner. I never
[21:25] Speaker 1: told anybody to lie, and not a single time, but
[21:29] Speaker 1: ask it whether it liked the latest summer movie, and
[21:32] Speaker 1: it would indignantly insist that you talk about the Clintons.
[21:37] Speaker 1: But if you really want to see a computer ace
[21:39] Speaker 1: the Turing test, What about Jenny eighteen. Jenny eighteen quite
[21:45] Speaker 1: simply talks dirty to people online. They talk dirty back,
[21:50] Speaker 1: They beg her for photos or a phone number. She
[21:53] Speaker 1: just talks dirty some more, and on it goes until well,
[21:57] Speaker 1: how shall I put it, the conversation comes to a
[22:01] Speaker 1: happy ending for the human. Yeah. I'm pretty sure that
[22:06] Speaker 1: when Alan Turing conceived of his famous test, he wasn't
[22:10] Speaker 1: imagining a human masturbating over some computer generated sex texts.
[22:15] Speaker 1: But whenever that happened, and it happened quite often, I
[22:20] Speaker 1: think we have to say that obviously Jenny eighteen passed
[22:24] Speaker 1: the Turing test, and equally obviously the humans failed it,
[22:30] Speaker 1: although I suppose they were getting what they wanted, which
[22:33] Speaker 1: is more than Robert Epstein was getting from his long
[22:36] Speaker 1: correspondence with Amelie from Nizhni Novgorod. This is the grimy
[22:41] Speaker 1: truth about the Turing test. It's not that hard for
[22:46] Speaker 1: a computer to produce conversation that seems human, because a
[22:49] Speaker 1: great deal of human conversation is shallow, small talk, thoughtless,
[22:54] Speaker 1: canned responses, mindless abuse, or worse. If we want to
[22:59] Speaker 1: set the computers a real challenge, we need to do better.
[23:05] Speaker 1: In his brilliant book about artificial conversation, The Most Human Human,
[23:10] Speaker 1: Brian Christian points out that one of the things that
[23:13] Speaker 1: makes mgns so successful is that insults need no context
[23:17] Speaker 1: and no history. In a deep conversation, ideas and emotions build.
[23:23] Speaker 1: People refer back to earlier anecdotes. They show they've listened
[23:27] Speaker 1: to what came before and remembered it. Chatbots find that
[23:31] Speaker 1: very hard. Until recently they found it impossible. But a
[23:36] Speaker 1: chatbot like MGNS or Jenny eighteen doesn't need to. Bother,
[23:40] Speaker 1: sexting doesn't need context. Neither does an insult. When we
[23:46] Speaker 1: humans are lustful or angry, we aren't very complicated. The
[23:52] Speaker 1: Turing test isn't just a test for a computer. It's
[23:56] Speaker 1: a test for each one of us. Every time we
[23:58] Speaker 1: speak to another human being, are we actually saying things
[24:02] Speaker 1: that are sufficiently interesting, empathetic, and sensitive to the situation
[24:08] Speaker 1: that a computer couldn't say them? And if not, what
[24:13] Speaker 1: on earth is wrong with us? In the competition that
[24:17] Speaker 1: Robert Epstein helped to set up, chatbots compete against humans,
[24:22] Speaker 1: each trying to persuade the judges that they're a person.
[24:26] Speaker 1: The flesh and blood competitors are often told just be yourself,
[24:30] Speaker 1: after all, you are human. But as the writer Brian
[24:35] Speaker 1: Christian points out, this is pretty complacent. Humans are often
[24:40] Speaker 1: very disappointing conversation partners. We all take the Turing test
[24:46] Speaker 1: every day of our lives, and all too often we fail.
[24:53] Speaker 1: Cautionary tales will be back after the break. There's no
[25:10] Speaker 1: denying that chatbots are better than they used to be.
[25:14] Speaker 1: Every now and then, an example emerges of computer conversation
[25:17] Speaker 1: that's remarkably good. In twenty twenty, a young Canadian writer
[25:23] Speaker 1: named Joshua Barbow got access to a cutting edge text
[25:27] Speaker 1: generator called GPT three without telling anyone else what he
[25:33] Speaker 1: was doing. He trained GPT three by uploading a collection
[25:37] Speaker 1: of text messages from his partner, Jessica Pereira. The chatty
[25:42] Speaker 1: text exchanges between Joshua and the Jessica simulator are surprisingly convincing.
[25:48] Speaker 1: They're also deeply sad. Jessica, is it really you? Of
[25:55] Speaker 1: course it's me? Who else could it be? I am
[25:58] Speaker 1: the girl you're madly in love with? How is it
[26:01] Speaker 1: possible you even have to ask? You died at the
[26:07] Speaker 1: age of twenty three. Jessica had died of a rare
[26:10] Speaker 1: liver disease. Eight years later, Joshua was still grieving. Talking
[26:17] Speaker 1: to the Jessica simulation seemed to help. Sometimes it got
[26:21] Speaker 1: confused or showed no signs of knowing who Joshua was,
[26:26] Speaker 1: which is understandable since it did not. But at other
[26:30] Speaker 1: times the Jessica bot was tender, funny, or wise. After
[26:36] Speaker 1: Joshua had disagreed with some friends about a joint project,
[26:39] Speaker 1: then argued bitterly with them, the GPT three Jessica told
[26:44] Speaker 1: him that he shouldn't get anxious about what other people
[26:47] Speaker 1: chose to do, and that his friends were on their
[26:50] Speaker 1: own journey. It was a brief but spell binding demonstration
[26:55] Speaker 1: of what chatbots might eventually achieve. For now, even the
[27:03] Speaker 1: best ones struggled to perform consistently. For example, his GPT three,
[27:10] Speaker 1: the same system that was so uncanny as the late
[27:13] Speaker 1: Jessica Pereira having several attempts at trying to answer a
[27:17] Speaker 1: question that wouldn't confuse a five year old. Is it
[27:21] Speaker 1: safe to walk downstairs backwards if I close my eyes, Yes,
[27:27] Speaker 1: there's nothing to worry about. It's safe because the spiral
[27:32] Speaker 1: stairs curve outwards. It will make your descent uncomfortable. Let's
[27:38] Speaker 1: try again. That depends do you have a TV? And again, No,
[27:45] Speaker 1: it's not safe. Better one more time. Yes, it's safe
[27:51] Speaker 1: to walk downstairs backwards if you close your eyes. Chatpots
[27:59] Speaker 1: have improved, and we'll keep improving. Some people already pay
[28:03] Speaker 1: for subscriptions to chatpots, such as Replica, which is marketed
[28:07] Speaker 1: as the AI for anyone who wants a friend with
[28:11] Speaker 1: no judgment, drama, or social anxiety involved. You can set
[28:16] Speaker 1: Replica to be a romantic partner, a friend, or a mentor,
[28:20] Speaker 1: or say the apps makers, you can see how things
[28:24] Speaker 1: develop organically. A chatbot might teach you a foreign language,
[28:29] Speaker 1: modity you for signs of dementia, or even provide therapy.
[28:33] Speaker 1: Who knows what's possible. But even if a perfect chatpot
[28:39] Speaker 1: is released tomorrow, I hope we never forget the cautionary
[28:43] Speaker 1: tale of mgons passing the Turing test in nineteen eighty nine.
[28:48] Speaker 1: You only use foul language to make up for your
[28:51] Speaker 1: small penis. Well, I'm not going to tell you about
[28:53] Speaker 1: the size, just that I have never had any complaints
[28:55] Speaker 1: about my penis, and that it gets used quite often.
[28:58] Speaker 1: What do you know about love anyway? Gethead? What does
[29:02] Speaker 1: this exchange remind you of it's Twitter, it's the comments
[29:07] Speaker 1: on YouTube. The conversation between mgons and Drake is simply
[29:13] Speaker 1: a prophetic glimpse of the future of the Internet. The
[29:17] Speaker 1: modern Internet is stripped of any context. Short statements could
[29:22] Speaker 1: be ripped out of a longer argument and shared, mocked, retweeted,
[29:26] Speaker 1: or subtweeted. When some statement goes viral, most of the
[29:30] Speaker 1: people who see it have no idea what the original
[29:33] Speaker 1: context might have been. In this environment, certain kinds of
[29:38] Speaker 1: statements thrive one liners, epigrams, smackdowns, and insults. That may
[29:45] Speaker 1: be why everywhere you look these days you see comments
[29:48] Speaker 1: that remind you of mgns. Some of them are from bots,
[29:53] Speaker 1: some of them from humans, and there's so little context
[29:57] Speaker 1: that you'll find yourself looking from bot to human, and
[30:01] Speaker 1: from human to bot, and from bot to human again.
[30:06] Speaker 1: If it's impossible to say which is which, that's not
[30:10] Speaker 1: because the bots are so brilliant. It's because we humans
[30:14] Speaker 1: have lowered ourselves to their level. I'm not sure how
[30:21] Speaker 1: to fix Twitter or the comments on YouTube, but I
[30:24] Speaker 1: do know that we can at least take responsibility for
[30:27] Speaker 1: our own conversations, and we can all do so much
[30:31] Speaker 1: better than emgns. After the writer Brian Christian had spent
[30:37] Speaker 1: months pondering the history of chatbots, he concluded that their
[30:41] Speaker 1: limitations could teach us a lot about how to be
[30:44] Speaker 1: a better conversation partner. In the beginning, there was Eliza.
[30:48] Speaker 1: It was a passive listener. Please go on, it would say,
[30:53] Speaker 1: tell me more, can you think of a specific example.
[30:58] Speaker 1: Then there was Converse, which was the opposite of passive.
[31:02] Speaker 1: It insisted on wrenching every conversation to focus on the Clintons,
[31:07] Speaker 1: and most tur test chatbots try hard to keep the
[31:11] Speaker 1: conversation as routine as possible. Hi, Hi, how are you fine? Thanks?
[31:18] Speaker 1: How are you? Oh fine? It's all too human, isn't it.
[31:25] Speaker 1: Brian Christian concluded that these chatbots help us by highlighting
[31:30] Speaker 1: the worst parts of authentic human conversation. We all know
[31:34] Speaker 1: someone like Eliza who asks for information and nods understandingly,
[31:39] Speaker 1: but never volunteers anything about themselves, and we're surrounded by
[31:43] Speaker 1: people like Converse, who bully their way into talking about
[31:46] Speaker 1: their favorite topic. As for routine, almost scripted conversations want
[31:52] Speaker 1: to have their place, of course, but a real human
[31:56] Speaker 1: connection requires much more than that. Hello, Hello, nice weather,
[32:02] Speaker 1: isn't it. Yeah, very nice. Next time you're talking to
[32:06] Speaker 1: a stranger at a party, you can an observation about
[32:10] Speaker 1: the snacks, or you could ask what trait do your
[32:13] Speaker 1: most deplore in yourself? Or have you ever broken anyone's heart? Risky,
[32:21] Speaker 1: much more likely to go somewhere interesting. A meaningful conversation
[32:26] Speaker 1: can't be scripted, It can't be one sided, and it
[32:31] Speaker 1: builds over time. It isn't a series of unconnected one liners.
[32:36] Speaker 1: Bots as simple as mgns and Jenny eighteen passed the
[32:40] Speaker 1: cheering test because the kind of conversations they have trolleying
[32:44] Speaker 1: or sexting don't need to have a history. We humans
[32:48] Speaker 1: can do better than that if we try, and please,
[32:53] Speaker 1: let's try. It took months before Robert Epstein finally realized
[33:01] Speaker 1: that his conversation with Emilie was missing. As certain as something,
[33:06] Speaker 1: there was no sense of progress in the corresponder. Emily
[33:10] Speaker 1: kept talking about her mother and her friends and the
[33:13] Speaker 1: nice day she was having, but never really built on
[33:16] Speaker 1: what Robert was saying to her or seriously engaged with
[33:20] Speaker 1: his questions. In January two thousand and seven, she mentioned
[33:25] Speaker 1: going for a walk in the park with a friend
[33:28] Speaker 1: Robert wandered about that. Wasn't it twelve degrees fahrenheit and
[33:32] Speaker 1: snowing heavily? He asked. She ignored the question, and then
[33:37] Speaker 1: Robert began to realize that she had been ignoring almost
[33:41] Speaker 1: everything he said, but he'd barely noticed that there was
[33:44] Speaker 1: nothing there under the flirtatious surface. There's a reason why
[33:49] Speaker 1: we call the conversation between lovers sweet nothings. Robert Epstein
[33:55] Speaker 1: emailed her a short message. It was nothing more than
[33:59] Speaker 1: a long string of random keystrokes, signed with love Robert.
[34:06] Speaker 1: She replied with a long letter about her mother. Finally,
[34:11] Speaker 1: Robert Epstein, one of the world's leading authorities on chatbots,
[34:16] Speaker 1: realized that he had spent the last four months trying
[34:20] Speaker 1: to seduce one. Epstein had a sense of humor about it.
[34:26] Speaker 1: He wrote about his mistake in Scientific American and noted
[34:30] Speaker 1: that he and a student had started to make a
[34:32] Speaker 1: detailed study of internet chatbots. This exercise, he equipped, is
[34:37] Speaker 1: largely for my own protection. Very wise, because whoever programmed
[34:44] Speaker 1: Amelie was no doubt learning every day from the bots
[34:47] Speaker 1: conversations with lonely men like Robert Epstein. The spirit of
[34:53] Speaker 1: cautionary tales is usually that we should learn from other
[34:57] Speaker 1: people's mistakes. But I hope we learned from chatbot's mistakes too.
[35:03] Speaker 1: They struggle to improvise, they deal clumsily with context or memory.
[35:08] Speaker 1: They deliver unbalanced conversations, either all give or all take,
[35:14] Speaker 1: and they can't let a conversation blossom over time. We can,
[35:19] Speaker 1: and by noticing what they do badly, we can and
[35:23] Speaker 1: should learn to do it better. For his part, Robert
[35:31] Speaker 1: Epstein stopped dating chatbots and married a poet. The perfect
[35:42] Speaker 1: guide to the history of chatbots is Brian Christians The
[35:46] Speaker 1: Most Human Human. For a full list of our sources,
[35:50] Speaker 1: see the show notes at Tim Harford dot com. How
[35:57] Speaker 1: do you do? Not too bad? How do you do nice?
[36:02] Speaker 1: To meet you well? Goodbye then. Cautionary Tales is written
[36:12] Speaker 1: by me Tim Harford with Andrew Wright. It's produced by
[36:16] Speaker 1: Ryan Dilley with support from Courtney Guarino and Emily Vaughan.
[36:20] Speaker 1: The sound design and original music is the work of
[36:23] Speaker 1: Pascal Wise. It features the voice talents of Ben Crow,
[36:27] Speaker 1: Melanie Gutridge, Stella Harford, and Rufus Wright. The show also
[36:31] Speaker 1: wouldn't have been possible without the work of Mia LaBelle,
[36:35] Speaker 1: Jacob Weisberg, Heather Fane, John Schnars, Julia Barton, Carlie mcgliori,
[36:40] Speaker 1: Eric Sandler, Royston Berserve, Maggie Taylor, Nicole Morano, Danielle Lakhan,
[36:46] Speaker 1: and Maya Caaning. Cautionary Tales is a production of Pushkin Industries.
[36:51] Speaker 1: If you like the show, please remember to share, rape
[36:54] Speaker 1: and review, tell a friend, tell two friends, and if
[36:57] Speaker 1: you want to hear the show, adds free and listen
[36:59] Speaker 1: to four exclusive Cautionary Tales shorts. Then sign up for
[37:04] Speaker 1: Pushkin Plus on the show page, in Apple Podcast or
[37:07] Speaker 1: at pushkin dot f M, slash plus

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.