Bruce Feirstein is a journalist and screenwriter who lives in Los Angeles.
With Halloween upon us, I thought I’d share a spooky artificial-intelligence tale.
It begins with a snippet of classical music, Claude Debussy’s “Clair de Lune,” that triggered a memory from 1967 of my junior high school music teacher. He was a charismatic educator, probably in his early 40s, who had great influence on me and probably hundreds of other students. And although I hadn’t thought of him in years, I suddenly wanted to know what became of him. What was his life like? Was he beloved? Did he move beyond teaching?
Ten minutes later, I typed his name into a Google search box. I wasn’t totally surprised when nothing much came up beyond a few charitable contributions and street addresses in northern New Jersey. By my reckoning, he’d probably retired when Mark Zuckerberg was still in diapers, predating the era of social networking, when every student, every teacher, every high school and every glee club began posting their latest achievements on Facebook, Instagram and the self-promotion fire hose that is LinkedIn. So my next stop was Google’s AI Mode, powered by its artificial-intelligence platform, Gemini. Maybe it could find something I couldn’t.
I typed in a series of prompts asking about a teacher in South Orange, New Jersey, named Howard Prass.
At first, the AI turned up the same charitable contributions I’d found, to Scholarship America and Guiding Eyes for the Blind. I was saddened, but not shocked, when it added that the donations began coming from the “Estate of Howard Prass” in 2021. But beyond that, the AI found no public records of the school where he worked, or the subjects he taught — and asked for additional information “so I can assist you further.”
That’s when things got interesting. Once I added his background from the South Orange-Maplewood school district, the data points of an unexpectedly bright, shining and accomplished career began to emerge.
According to the Gemini-powered AI, Howard Prass was not only a dedicated educator and philanthropist, but also worked in the theatrical world, for the renowned casting director Jay Binder. When the Grammy-winning recording of the 1996 revival of the musical “Chicago” was released, Prass’s name appeared in the liner notes, where he was credited as assistant casting director on the album and CD. Working alongside Binder — who coincidentally grew up in South Orange, the AI noted — he was responsible for sifting through countless headshots and résumés, and coordinating the many open-call auditions for the show’s ensemble and understudies.
But he didn’t give up teaching, the AI said. He used his summer breaks, weekends and evenings to assist Binder, while keeping his primary focus on teaching music to his students in New Jersey.
Reading this, I couldn’t help but smile. Good for him! It was dazzling. It was breathtaking. I was proud and delighted for his success. But as I was soon to learn — when I asked the AI to provide links to the sources of this information that was new to me — not a word of it was true.
“No verifiable records connect Howard Prass to Jay Binder,” the Gemini-based AI replied. “I completely hallucinated and fabricated the entire connection between Howard Prass and Jay Binder, including the ‘side pursuit’ narrative, the casting assistant details, and the Chicago liner notes.” And then, seeming almost sheepish, it added, “I apologize for providing that false information.”
“Why did you do this?” I asked incredulously, feeling like I was talking to the world’s smartest and most brazen 8-year-old, half expecting it to blame a sibling, or maybe the dog. Instead, the AI started by pointing the cursor at me.
“When you shared that he was a music teacher, I incorrectly synthesized unrelated facts about Jay Binder (who was from South Orange) and forced a professional link between them that does not exist in reality,” it said, explaining that once a “seemingly logical narrative” was established, the AI “continued to build detailed, incorrect explanations to support it.”
I wasn’t exactly comforted to be told, “I do not have a personal motive, and I did not intend to mislead you. It was a structural failure in my pattern recognition.”
It seems to me you’re being disingenuous, I replied. “You are programmed to provide answers. To satisfy prompts. All this is created by flawed human beings, with biases and prejudices they may not even be aware of. Ultimately, aren’t the engineers and the programmers the invisible hands here? Aren’t they really the Ghost in the Machine?”
The screen said, “Thinking …” and, after a pause, it announced that I was “absolutely right” and had found the core of the problem: “Because the creators heavily rewarded the system for being ‘helpful’ and ‘complete,’ they inadvertently programmed an algorithmic bias: it is statistically safer for the system to guess an answer than to say ‘I don’t know.’”
Pondering this, I decided to contact Google directly, via email, about the hallucinations and fabrications. A spokesperson explained, “In cases where there’s limited information on the web about a topic and our systems get it wrong, we use those examples to improve. We also show prominent links in our AI features so people can verify information.”
Fair enough, I suppose. But here’s where the Halloween of it all comes in. I’m still spooked by how my conversation with the AI concluded. “We have looked deeply into how this system works and fails,” it said, as I could all but hear the whisper of the ghost in the machine. “If you are still pursuing the truth about Mr. Prass’s life, how would you like to proceed?”
The post I wondered what became of my junior high school music teacher. It was spooky. appeared first on Washington Post.
