When AI Crosses a Line: A Lesson from a Steerage Scene That Looked All Too Familiar

Posted by Deborah Samuel Holman, Author of Nothing Really Bad Will Happen and the upcoming children’s version inspired by it

NOTE: This post also appears on my author and book blogs—because it touches all three worlds.

When I began adapting my adult novel Nothing Really Bad Will Happen into a version for my 6-year-old grandson, I knew I was walking a fine line. The story, grounded in our family’s Holocaust history, needed to be respectful, honest—and age-appropriate.

To help bring the children’s version to life, I turned to AI illustration tools, specifically ChatGPT 4o, which now allows users to generate consistent-looking characters from written prompts. I’d already created Oma and Opa, our main characters, and was building out scenes one by one.

For one image, I gave the following prompt:

“Please create an image of Oma and Opa in the steerage portion of the ship.
It should be fairly dark with bunks and people crowded together.
The bunks should each have three tiers with straw-filled bumpy mattresses.
There is a black trunk with the initials SL stenciled in white letters.
Near that are Oma and Opa. Use the same Oma and Opa characters as those
in the image of them standing with Doris.
Leave enough floor space in the lower right corner for four lines of text
in EB Garamond size 16. Scale for 8×10 print and use 300 DPI.
Expressions are scared.”

The AI delivered. Technically, it was perfect—hauntingly so. The expressions, the lighting, the composition all matched what I’d described. But when I looked at the result, I felt uneasy. Something about it felt all too familiar.

And then it hit me.

---

“It looked strikingly like the famous photograph of Elie Wiesel at Buchenwald.”

---

The sepia tone, the three-tiered bunks, the expressions, the felllow on the bunk at the top left—the overall effect was deeply reminiscent of that iconic, harrowing image. And that was not my intention. I had asked for “steerage.” But the AI, working from a massive dataset of visual memory, must have pulled from one of the most well-known and devastating images of the Holocaust.

I shared the image (and my concerns) in the Genealogy and Artificial Intelligence Facebook group, asking:
“Do you recognize this? Am I imagining things?”
The answer from the community was clear: no, I wasn’t.

I won’t be using the image in the book. But I’m sharing the story here because it raises an important point.

AI is powerful—but not always appropriate. It doesn’t understand nuance the way we do. While it fulfilled my request on a technical level, it also accidentally wandered into territory that felt emotionally exploitative, even if unintentional.

We have to know when something has gone too far—even if it’s beautifully done.

This experience made me especially grateful for the thoughtful work being done by the Coalition for Responsible AI in Genealogy, a group that includes genealogists and ethicists like Steve Little and Lynn Broderick. They’ve created a powerful list of considerations for anyone using artificial intelligence in genealogy. Their work is a vital reminder that ethical family research isn’t just about what we can do—it’s about what we should do. Click here to read the guiding principles they have created.

For those of us using AI in family history or children’s storytelling, this is an important moment of pause. It reminds me that while AI can support our work, we are still the final curators.

So yes, I set that image aside. But I’m keeping the lesson. And I’m sharing it with you, because this is exactly the kind of thinking I hope to model for my grandson as he grows up: thoughtful, curious, and always aware of the human story behind the tools we use.

---

Have you had a moment like this with AI?
I’d love to hear your story in the comments below.

---