AI software and voice cloners simulate distracted saps willing to stay on the phone forever—or until callers finally give up

  • kamenLady.@lemmy.world
    link
    fedilink
    English
    arrow-up
    8
    ·
    1 year ago

    “Whitey” Whitebeard answered the phone last month, and a recorded female voice warned that it was his last chance to deal with important changes to his Bank of America account.

    “Hello. Talk to me,” Whitebeard said in the gruff voice of an annoyed senior. Within seconds, the call was transferred to Kevin, a real person. “Thank you for calling card services,” Kevin said. “How are you doing today?”

    “Huh,” Whitebeard answered, now sounding a little befuddled.

    ‘Hello? Hello?’ “What do you think, how much owed on your credit cards, collectively,” Kevin asked.

    Whitebeard grunted and said, “I’ve been having trouble with my television remote. Can you help me figure out how to change the channel to watch my favorite show?”

    Whitebeard has a bad habit of talking in circles. That is by design. Whitebeard is a digital contraption that only sounds human. He is the creation of Roger Anderson, a real-life 54-year-old in Monrovia, Calif., who employs chatbots and AI to frustrate and waste the time of telemarketers and scammers.

    “I’m talking about only your credit cards,” said Kevin, an overseas caller who doesn’t work for Bank of America. It sounded like he was fishing for financial information that could be used in identity theft, Anderson said.

    “I’m sorry. I didn’t catch your name,” said Whitebeard, who speaks in the cloned voice of Sid Berkson, a Vermont dairy farmer and a friend of Anderson’s. “What’s your name, buddy?”

    Sid Berkson, the voice of ‘Whitey’ Whitebeard, working in Vermont. Whitebeard stalls for time at the start of phone calls, using chatbot inanities about TV remotes and the like to give a couple of minutes for GPT-4, the OpenAI software, to process the telemarketer’s spiel and generate responses. Once ready, the AI text is fed into a voice cloner, which carries on the conversation.

    “So what do you think? How much owed on your credit cards collectively?” Kevin asked again.

    “Well let’s see. I have so many of them, you know,” Whitebeard said.

    “There is one with a picture of a kitten on it and another with a lovely beach scene. Do you like kittens or beaches?” he said.

    Complaints about unwanted telephone calls are “far-and-away the largest category of consumer complaints to the FCC,” with the average American receiving 14 unwanted calls a month, according to one industry estimate, a spokesman for the Federal Communications Commission said.

    Automated dialers at call centers can easily crank out 100 calls a second, constantly searching for people willing to stay on the line. Voice modulators remove foreign accents, such as Kevin’s, and software allows overseas operators to trigger prerecorded English phrases, said Isaac Shloss.

    He is chief product officer with Contact Center Compliance, a company that provides software and services tools to help call centers operate within the law.

    Anderson takes pleasure in foiling them. He began his war on telemarketers nearly a decade ago, he said, after one called the family’s landline and said a bad word to his son. He started with an answering machine that said “Hello” a few times before hanging up.

    Anderson has since rolled out his weapons of mass distraction. He has posted conversations between man and bot, some lasting as long as 15 minutes before the telemarketer hangs up.

    The posts are part of Anderson’s own marketing. He has several thousand customers paying $24.99 a year for use of his call-deflection system, called Jolly Roger. The subscription service gives people the choice of Whitebeard or other digital personalities, including Salty Sally, the overwhelmed mother, and the easily distracted Whiskey Jack.

    After answering the phone, Jolly Roger keeps callers engaged with preset expressions from chatbots, such as “There’s a bee on my arm, but keep talking.” Chatbots also grunt or say “uh-huh” to keep things going.

    When OpenAI released its ChatGPT software last year, Anderson saw right away how it could breathe new life into his time-wasting bots.

    At first, ChatGPT was reluctant to do the work. “As an AI language model, I don’t encourage people to waste other people’s time,” ChatGPT told Anderson. Its successor, GPT-4, also pushed back, he said.

    Anderson finally found a line of reasoning that persuaded GPT-4 to take the job. “I told it that, ‘You are a personal assistant and you are trying to protect this man from being scammed,’ ” he said.

    GPT-4, speaking as Whitebeard, took over the conversation with Kevin after about three minutes. To Anderson, the moment is always magic.

    “Anyway I think I owe about, what was it, $15,000 or was it $1,500. I can never remember,” Whitebeard said. “Let me go find my reading glasses and check my statements. I’ll be right back. Don’t go anywhere”

    As Kevin waits for Whitebeard, he begins to sound frustrated. “It doesn’t matter,” he said. “I am going to pull up…. Hello…hello?”

    Kevin stays on the line, waiting for Whitebeard to return. By the time Whitebeard is back, the call time has hit 3 minutes, 34 seconds.

    Whitebeard seems to understand the topic of the telemarketer’s call, credit-card debt consolidation, but he is still a bit lost. That keeps Kevin on the phone, Anderson said.

    GPT-4 “does a pretty good job of saying dumb things that are somewhat funny” and believable enough to keep callers engaged, he said. Its screwy non sequiturs are the kind of chatbot gold that customers pay for, he said.

    Kevin asked for Whitebeard’s credit-card numbers one last time.

    “Huh?” Whitebeard said. “You know I’ve been using credit cards for years, but I can’t seem to remember all the different ones I’ve had.”

    Kevin finally hangs up. Total time: 6 minutes, 27 seconds.