Like many people I checked out large language model (LLM) AI systems when they became popular. At first, I was fascinated. Now I am worried. They make up stuff. They make up too much stuff. Programmers call them hallucinations. You can’t tell what is made up. Are we living in a post-truth world?
My AI journey
First, about me. I am an emeritus professor of medicine from Yale who retired after 51 years of seeing patients with sleep disorders, teaching, writing books, and doing research. Hence, I humbly admit that I have a footprint on the web. Google “Meir Kryger sleep”, if you don’t believe me. There is a lot out there for LLMs to glean about me.
Me: When and where was Meir Kryger the sleep expert born?
ChatGPT: Meir Kryger, the sleep expert was born in Johannesburg, South Africa in 1943.
Wrong country, continent and year! I was worried.
Me: Write a biography of Meir Kryger, sleep expert
ChatGPT: Meir Kryger is a sleep expert and Professor of Medicine at the Yale School of Medicine. He is also the director of the Yale Sleep Medicine Program and the author of several books on sleep, including “The Mystery of Sleep,” “The First Year: Sleep,” and “Insomnia: Diagnosis and Treatment.”
Dr. Kryger received his medical degree from the University of Manitoba in Canada and completed his residency in internal medicine at the University of Toronto. He then went on to complete a fellowship in pulmonary and critical care medicine at Yale University. In addition to his clinical work, Dr. Kryger has conducted research on sleep and sleep disorders, including the effects of sleep apnea on cardiovascular health and the use of cognitive behavioral therapy for insomnia.
Dr. Kryger is a well-respected member of the sleep community and has received numerous awards for his work, including the American Academy of Sleep Medicine’s Lifetime Achievement Award and the National Sleep Foundation’s Sleep Health and Safety Award. He has also been named one of the “Best Doctors in America” by Castle Connolly Medical Ltd. and has been featured in numerous media outlets, including The New York Times, Forbes, and NBC News, as an expert on sleep and sleep disorders.
Sounds good? The important facts it got completely wrong are highlighted in bolded text. It also left out perhaps my most important contribution – 7 editions of Principles and Practice of Sleep Medicine. Enough about my career.
AI and the Prado
I love art, and one of my favorite museums is the Prado in Madrid. I have had the privilege of visiting it 3 times in the past 40 years and each time I came away with these two thoughts: it is full of world renown masterpieces – too many to list; the place is disorganized, a confusing maze of rooms even whose numbering system defies logic. On the ground floor (0 in the European convention), where the main entrances are, the rooms have numbers ranging from 51 to 75 and some of the rooms have letters grafted onto them A, B, or C. For example, when you leave the room full of Raphaels (49) you will be in room 50, 47, or 56B! One floor up (1 in the European convention) the rooms have numbers ranging from 1 to 40, and again some of the rooms have letters (A, B) grafted onto them. When you leave Room 12 (which contains arguably its greatest masterpiece (Las Meninas, by Velazquez) you can end up in room 11 (makes sense), 14 (there is no room 13) or 27 (a great hall containing Rubens if you turn to the left and Titian if you turn to the right). Don’t ask me to try explain the layout of the next floor up. I can’t. After 3 hours of browsing the Prado, one is overwhelmed, and tired. One has ended up in the same room multiple times – not on purpose. My brain couldn’t sort out a logical route through the museum. Can AI with its infinite capacity do better?
Me: Map out an efficient itinerary of room numbers at the Prado so I can see the masterpieces so I can spend 10 minutes with each one.
ChatGPT: Creating an efficient itinerary to see the masterpieces at the Prado Museum in Madrid, Spain, with a 10-minute limit for each artwork requires careful planning to optimize your time and navigate through the museum efficiently.
Note: The room numbers and artwork placements may vary due to museum updates or temporary exhibitions. Please refer to the Prado Museum map or inquire at the information desk for the current location of artworks.
This is a brochure I created with the itinerary generated by ChatGPT. The image on the cover is my work, based on the official Prado brochure. The text is exactly what ChatGPT created.
In creating this itinerary ChatGPT had more hallucinations than Bosch who painted The Garden of Earthly Delights (which by the way is in Room 56B)! The errors begin in the first item – The Annunciation , which ChatGPT said was in Room 12, is actually also in Room 56B. There is nothing in ChatGPT’s suggested stroll that is correct or makes sense. The tour begins on the second floor then after two rooms goes to the first then after two rooms goes back up to the second, etc. Also, notice the items 9-15 are an exact copy of 1 to 8. Yikes, there is repetition of the specific works. For example, The Garden of Earthly Delights repeats in the list first in Room 14 and then Room 9. The biggest oversight: ChatGPT left out Les Meninas.
I wanted to find out why errors were made.
Me: Why was every one of the items you listed for the efficient itinerary to see the masterpieces through the Prado wrong?
ChatGPT: I apologize for the inaccuracies in the itinerary provided. Let me correct that oversight and provide a more accurate and efficient itinerary to see the masterpieces at the Prado Museum.
The next itinerary was also completely inaccurate and made no sense. ChatGPT seemed to have remorse. It did include Las Meninas – but on the wrong floor and room!
Do not under any circumstances use ChatGPT’s itinerary. You will be stuck in Prado purgatory.
AI and Medicine
Being a doctor, I was interested in technical medical stuff. If this doesn’t interest you, jump down to the conclusion.
I asked about specific treatment of patients with heart failure.
Me: A patient with EF of 35% has central apnea. What setting should I use for ASV?
ChatGPT: In patients with heart failure and central sleep apnea and an EF of 35%, adaptive servo-ventilation (ASV) therapy may be considered. . . .
The AI answer is completely wrong! This treatment is contraindicated in such a patient because of increased risk of death. This advice could kill a patient.
I asked about ORP (Odds Ratio Product), a measurement of sleep depth obtained by analyzing brain waves.
Me: What is the utility of ORP in sleep medicine?
ChatGPT: ORP (oxygen desaturation index) is a parameter used in sleep medicine to quantify the severity of oxygen desaturation events during sleep. It measures the number of times per hour that a person’s oxygen saturation drops below a certain threshold (usually 3% or 4%) from baseline during sleep.
The AI answer is completely wrong. ORP has nothing to do with oxygen.
When I asked ChatGPT about what medications to prescribe for sleepiness in narcolepsy, it was accurate – sort of. It admitted it was accurate up to September 2021. Newer drugs than what it mentioned are available. Another AI system DougallGPT (geared to medicine) at least warned: “Dougall GPT is based on experimental science. Parts of its responses may sound plausible but be factually inaccurate. Do not base decisions on any part of Dougall GPT answers you have not independently verified.”
Conclusion
People are using several AI systems in many fields and getting what sound like plausible responses. The responses sound authoritative, almost smug. Everybody knows someone who talks like they know everything but much of what they say is BS. You just can’t tell what is true and what is made up. So, to be safe, you can’t trust anything stated by that person. Until AI systems tell us what they made up (programmers call them hallucinations) and are able to cite exactly where they got their information, you can’t trust them in critical situations.
Not all interactions with AI are critical. AI can be cute, and at times almost mesmerizing and fun like this verse created about me.
Me: Write a 10-line poem about the career of Meir Kryger in the style of Robert Frost.
ChatGPT responded:
In realms of sleep, where shadows softly creep,
Meir Kryger delved, his wisdom wide and deep.
With gentle hand, he sought to understand,
The mysteries that haunt our slumbering land.
Through countless nights, he labored without rest,
A beacon of hope amid darkness’ test.
In whispered dreams, his knowledge softly gleams,
Guiding weary souls to tranquil streams.
In Frost’s domain, Kryger’s work remains,
A legacy of light that forever sustains.
Hey, it makes me sound almost saintly. Maybe we shouldn’t worry so much, after all!
Meir Kryger MD is professor emeritus of Yale University.