AI Removed the Stranger--and Missed Me

Eight years ago, at my daughter’s wedding, I was footloose and fancy-free, singing into a microphone while wearing a bright green hat, a peach-colored tie, and a dark suit. I considered myself a father celebrating a joyous occasion. My brother compared me to Van Morrison.

A photographer snapped a close-up photo of me singing; it also showed someone in the background. Recently, I asked ChatGPT to edit the photo by removing the person in the background. It did an excellent job. The unwanted figure disappeared, leaving me alone with my microphone, my boutonniere, and a hat that could probably have directed traffic during a power outage. See for yourself:

This was an impressive feat. In my experience, removing someone from a photograph is considerably easier than removing them from a wedding guest list, but it still requires skill. Not only that, ChatGPT returned the edited image with an accurate caption: “Elderly man in a green hat.”

Both descriptions—me as a poor imitation of Van the Man and the AI’s caption—refer to the same person, but only one is likely to be invited to another wedding.

I asked ChatGPT how else it would describe the image, beyond “elderly man in a green hat.”

“A distinguished gentleman,” it replied.

Now we were getting somewhere.

I have always considered myself distinguished, though I generally prefer others to reach that conclusion on their own. Here was a machine doing so without prompting, a reference check, or even a glance at my curriculum vitae.

But how did it know I was “distinguished”?

Was the distinction visible in the angle of my microphone? Did the boutonniere signal professional status? Did something about the green hat suggest membership in Alpha Omega Alpha?

Perhaps it mistook me for the chairman of a very festive hospital board. Maybe it looked into my online past and found a photo of me in a marching band at the St. Patrick’s Day parade fifty-five years ago, when I was a freshman at Boston University.

The machine offered several options. I could be “a man in a green fedora speaking into a microphone,” “a silver-haired gentleman in festive attire,” or “a gentleman mid-speech.”

That last one was especially revealing because I was singing, not speaking. The difference is obvious. Anyone who has heard me do either would recognize it immediately.

The photograph showed a microphone near my mouth. From there, the machine prompted an activity. Speaking was plausible. So were singing, announcing the next dance, or explaining why the father of the bride shouldn’t have unrestricted access to the entertainment equipment.

It chose a plausible option and dressed it in confident prose, much as it had dressed me in distinction.

To be fair, “distinguished gentleman” is a familiar, flattering description of a silver-haired man in a suit. It is also a judgment the photograph cannot establish. The machine had no evidence of my character, accomplishments, or standing. For all it knew, I had borrowed the suit and wandered in, searching for champagne-soaked cake.

The troubling part was that ChatGPT and I have been communicating for nearly two years. We have worked on essays, books, revisions, and arguments, as well as the recurring challenge of achieving solid gridlines around a table. I had shared my professional history and my thoughts on medicine, and the machine helped me refine my authorial voice.

Yet when I showed it my photograph, it did not recognize the author.

I do not fault it for failing to identify a face when it had no reliable basis to do so. Guessing that every silver-haired man I upload must be Arthur Lazarus would create an entirely different problem.

Still, the experience exposed a peculiar gap. It could work with information about me without knowing that the man holding the microphone was me.

There I was: familiar in prose, anonymous in pixels.

That Jekyll-and-Hyde dynamic follows us into medicine. An AI system may compile a comprehensive account of a patient from symptoms, lab results, imaging, and clinical notes. It may identify a possibility the physician overlooked. Personally, I want that kind of assistance. Having spent decades in medicine, I have no sentimental attachment to overlooking things.

But an accurate inventory of information is only part of knowing a patient. A physician who has cared for someone over time may hear “I’m fine” and notice that it sounds different today. A joke may mask fear. A missed appointment may signal a transportation problem. A refusal may follow an experience the patient has never felt safe enough to discuss. These meanings emerge through attention, conversation, and a genuine relationship in which one person becomes accountable to the other.

Physicians can misunderstand patients, too. We can rush, assume, and let a familiar explanation stand in for understanding the patient’s complex presentation. A medical degree does not confer immunity to the distinguished-gentleman problem. But we recognize that failure is part of our job. We can ask another question, revisit an impression, and allow the patient to correct us. AI can produce an elegant account from whatever it has been given. However, AI’s charm can make missing pieces harder to spot.

In my photograph, the missing pieces were harmless: a daughter’s wedding, a father’s joy, and the fact that the alleged speech was a song. In clinical care, the missing piece may explain why a treatment plan fails, why a patient’s smile conceals worsening depression, or why someone is afraid to take a prescribed medication. That is why I want AI beside the physician, helping us see more and think better, while the physician remains responsible for learning who the patient is.

As for my photograph, I remain delighted. The background figure is gone. The green hat remains. The elderly man has been promoted to distinguished gentleman.

Next time, however, I would suggest a more informative caption.

“Dr. Arthur Lazarus, singing at his daughter’s wedding.”

It requires fewer adjectives.

And considerably more knowledge.