Every type of design product needs a different way of setting limits for AI.

When designing websites and apps, I use AI in the research, ideation, and prototyping stages. There, AI is used to create things that don’t yet exist.

But with the book “70 Years of Surgical Journey” for Binh Dan Hospital, it’s the opposite. AI helps recreate things that once existed. And once its job is done… AI has to disappear.

But to make that happen, I had to set a lot of limits for AI. Some things AI was allowed to do. Some things AI could only intervene in up to a certain point. And some things were absolutely off-limits.

Old newspapers

Binh Dan Hospital’s articles and images once appeared in old magazines dating back to 1953, like Tuan Bao Doi Moi, Dan Ta, The Gioi Tu Do Magazine, Anh Den Dau… but 70 years had yellowed the paper.

Some other documents had digital versions too small to print. There were also internal bulletins from 1971, their pages speckled with stains, edges frayed and torn.

I wanted to keep the documents carrying the marks of time, as they were. Just not letting time obscure the story on the page. To preserve the marks of time while removing the “noise,” the first limit I set for AI was: don’t rewrite history.

• Old articles often used regional dialects or spellings no longer common today. AI could easily mistake these for spelling errors and “correct” them. So the original typeface, down to every line of content, had to stay untouched. AI was only allowed to clarify what once existed, never to invent what never had.
• AI could wipe away every frayed edge and every stain on the paper. But for this project, AI was only allowed to process the parts obscuring the content. The yellowed color of the paper, and part of the trace of time, had to remain.

If everything came out perfectly clean, the archival photos would lose the feeling of having genuinely lived through 70 years.

Photos with no second chance to be taken

“No one steps into the same river twice.”

Old events have passed, moments can never be recreated, and some people are gone forever. Old photographs become all the more precious. They don’t just preserve history — they’re photos that can never be taken a second time. So every remaining pixel was worth saving.

But in our hands, images from the hospital’s early days were mostly just old prints. Some photos were only a few centimeters in the library archive. Some were copied from an album that had already discolored. Others came from doctors’ families, the black-and-white film faded with the years.

If this project had happened a few years ago, many of those photos would probably have had to be discarded, or printed very small to hide their flaws.

The next limit was even harder. AI was allowed to restore detail, but not to guess at detail. If it created details that never existed, it was no longer an archival photo. AI had to restore a photo without the reader seeing any trace of AI on it.

Thanks to that, photos once at risk of being discarded were finally able to appear more fully in the yearbook. A technology built to create the future could also be used to preserve the past.

Creating a symbol, instead of an individual

The cover of the 70th-anniversary yearbook shows a doctor holding a scalpel.

While I was still unsure whose image to use, the PM offered a perspective that made me nod immediately:

“We shouldn’t use the image of one specific doctor. Because the moment you put one face on the cover, the story unintentionally shrinks down to a single person. When those 70 years are really the story of thousands of people.”

So how could AI create a symbol?

Through meetings with the Board of Directors, I discovered a fact: doctors aged 40–45 are still considered the “young generation” in medicine. Perhaps because becoming a skilled surgeon takes many years of accumulated experience. Even within the yearbook, there are many stories of doctors who, after retiring, continue their journey of saving lives as advisors. This time, the limit was completely different. AI had to create a symbol accurate enough to represent thousands of people.

• The right age. A middle-aged Asian doctor.
• The right role. The doctor appearing right in the middle of surgery, rather than a posed portrait.
• The right attire. Surgical clothing down to the fabric, cap, mask, gloves.
• Every bodily detail correct. Even now, in 2026, AI still occasionally gives me a hand with six fingers. The scary part is it looks perfectly plausible at a glance — only when you count do you notice something’s wrong.

In the end, the person on the cover was no longer one individual, but a symbol in which every surgeon at Binh Dan Hospital could see themselves.

The one blurred photo in the book

A book using a blurred photo? This troubled me for days, because I always wanted every page to feel complete.

It was an article about the Thursday Night Program from before 1990. The only remaining material was a very small photo of five former program members, reunited and taking a commemorative photo together. A priceless photo — but the figures and faces were only barely recognizable.

I tried having AI restore it in several ways. The result was a photo that looked almost perfect — except the faces in it were no longer the faces history had left behind. They were only what AI had guessed from the few remaining details. That was the final limit I wouldn’t let AI cross.

For the first and only time in the entire process of making this yearbook, I only allowed AI to do basic image processing: denoising, rebalancing the lighting and contrast to make the photo a bit clearer. So that the people who took part in the Thursday Night Program back then could still recognize one another by build and by face.

For a historical yearbook, I chose to stay true to that moment.

AI raised the book’s standards

We also used AI to illustrate a few stories told only through spoken accounts, when no photograph remained to preserve that memory. Or to catch errors that were hard to spot in the final stage, like an extra space between two words, when everyone’s eyes had gone nearly blurry from exhaustion.

AI raised the quality standard the whole team could aim for. But AI doesn’t know which limits are right for the purpose of each design product. Those limits still have to be decided by us.

Many people have asked me: what did AI actually do for this yearbook?

The answer is: AI helped a great deal. But only when a reader turns each page without thinking about AI at all. When they see only history — that’s when AI has truly finished its job.