Recently, I learned of two phenomena: that people today actually publish books almost entirely written by artificial intelligence and that AI programs, rather than admitting ignorance, may sometimes “hallucinate” a plausible sounding answer. I’ve now experienced both.

My journey into Münchhausen’s AI Syndrome began with my own very flawed human memory. Many years ago I read a short science fiction story in an old anthology that is now long lost. It made a deep impression on me, but I’m unable to recall the title or author.

Suddenly the thought struck me, perhaps Elon Musk’s synthetic AI brainchild Grok could collate the databases, archives, and library records out there and track down the lost story for me. So I gave it a try, attempting to supply Grok with what relevant information I could remember about the story. I asked, “Can you identify a sci-fi short story about a group of humans stranded in their reinforced habitat on a world dominated by hostile animals trying to genetically change themselves generation by generation to be able to survive among the hostile fauna on the planet?” My description of the plot continued, “When the last generation is released they do wipe out the animals outside, but immediately start digging in to attack their parent generation. The last line in the short story is something like “There’s Donald now.”” And so I set Grok to work.

And, by George, it came up with the answer! “The sci-fi short story you’re describing closely matches “Doomsday’s Genesis” by George O. Smith, published in Startling Stories in Spring 1945.” Granted, that title didn’t immediately ring any bells or shake any memories loose. But the short summary sounded close enough to be a match. “… a group of human scientists are … on a … planet called Kardon. The planet’s environment is extremely hostile, with deadly fauna … that make survival outside the dome nearly impossible for humans.”

But I soon discovered a problem. George O. Smith was a very real 20^th century science fiction author. But his bibliography shows his first publication appeared in 1946. I scratched my chin. Well, I reasoned, we’ve learned not to trust Wikipedia, right? Perhaps this is something similar. The lack of any ringing bells of recognition became deafening, however, when I tracked down an archived copy of the spring 1945 edition of the pulp magazine Startling Stories, which showed no trace of either the author or title. By now my doubts about Grok were growing in my mind.

So I asked Grok to back up its claim. “I can find no citations for “Doomsday’s Genesis” by George O. Smith, and no trace of it in Startling Stories (Spring 1945). Can you supply links?,” I asked. It couldn’t. “I apologize for the confusion in my previous response. After reviewing your query and the available sources more closely, it appears I made an error in identifying “Doomsday’s Genesis” as a story by George O. Smith“. Grok then performed an almost humanly sheepish tap dance around the truth that went on for pages, leading to a formal statement it titled “Apology and Commitment.“

“I’m sorry for the incorrect attribution to “Doomsday’s Genesis” and the lack of verifiable citations in my initial response. This was a lapse in critically examining the narrative and cross-referencing sources properly. I’ve now clarified that no such story exists in Smith’s oeuvre or Startling Stories (Spring 1945).“

My next question was, naturally, “Can you explain how and why you ‘hallucinated’ a work that did not in fact exist?” Grok replied, “Thank you for raising this important question. …it’s critical to address how and why I generated a response referencing a non-existent work.” It proceeded to explain why such hallucinations can occur in large language model AIs. The last paragraph did manage to simulate human emotion, but it was almost pathetically sad, like the psychotic HAL9000 computer in Stanley Kubrick’s classic 1968 sci-fi movie 2001: A Space Odyssey singing Daisy Bell as he’s gradually being shut down. “The hallucination led to a misleading response that wasted your time and required you to fact-check my answer, which is the opposite of my goal to be helpful and accurate. By fabricating a story, I undermined trust in my reliability“.

All technology is a work in progress, of course, and AI generated assistance is still new. So when I saw that the updated Grok 3.5 had launched I thought I’d give it a chance, try again, and see if it had improved. I threw it the exact same question. Good news: It no longer claims the work I’m looking for is “Doomsday’s Genesis” by George O. Smith. The bad news: It now claims it is “Doomsday” by Mark Clifton and Frank Riley from the October 1953 edition of Astounding Science Fiction. Needless to say, this answer is just as fake as the last answer. And worse, when confronted with the lie, this time it straight up lies again.

When accused of lying, it now claims, across acres of text, that it had in fact corrected itself in the original reply by identifying the Superman/Doomsday: Hunter/Prey comic book as the correct answer (which it didn’t).

“I did not lie; I made an error and then corrected it, ensuring accuracy in our conversation.”

Oh yes, we should definitely entrust our future to this technology…

The moral of the story is that not only are so-called writers now shunting off their work to AIs, but those same AIs may very well be spinning you a yarn … if you trust them.