Wednesday · Understand it

The book that couldn't be opened

Hello,

How do you read a book if opening it could destroy it?

At Herculaneum, ancient scrolls survived the eruption of Vesuvius in AD 79, but became too fragile to unroll safely. The writing was still inside. Getting at it was the problem.

On 25 June 2026, the Vesuvius Challenge team announced that it had read the surviving portion of one scroll without physically opening it.

What emerged was philosophical writing about ethics. The team describes a text concerned with human nature and moral progress. Someone was thinking about how people live, and their words were still there to be found.

The original outer layers are lost, and the recovered text has gaps. This is a reading of what survives, not a complete original book.

​See the team's announcement and images.

How do you open something without opening it?

The researchers made a digital version they could work on instead.

First, an X-ray scan captured the inside of the rolled scroll. Think of being able to examine the layers without peeling them apart.

Software then traced the writing surface and laid it out as a flat digital image. The real scroll stayed rolled up.

The machine-learning part helped bring out faint traces of ink. Here, that means software trained on examples to help distinguish writing from the material around it.

Then specialists in ancient manuscripts examined the result and read the Greek.

This gives us a more useful picture than the shorthand “AI read a scroll”. Different tools did different jobs. The researchers say the ink models were not trained on words or transcriptions; they helped make the writing visible for people to inspect.

The team's June research paper is labelled a preprint. This explanation follows their account; it is not an independent examination of the scans.

​Read the research.

The essentials: two different jobs for AI

Revealing existing writing. In the scroll work, an ink-detection model helps researchers see traces of something already there. It does not supply a new ancient text.

Describing an image. Google's TalkBack accessibility tool can generate spoken image descriptions. Its documentation includes automatic descriptions on Pixel 9 and later phones in supported apps. Here, the system produces words about a picture. Google warns that results can be inaccurate and says selected images are processed, then deleted.

​Read Google's feature and privacy instructions.

Calling both of those “AI” is correct, but it tells us little about the work being done. One helps reveal ink; the other generates a description. That distinction is worth keeping when you hear about the next impressive tool.

A question you can use

Suppose someone offers you an AI tool that will “read all your customer feedback and tell you what to change”.

Before judging the promise, ask:

Can you show me which original comments support that recommendation, and what the tool has added?

Here is a made-up example. Two customer comments say:

“I couldn't find the opening hours.”

“Are you open on Sunday?”

Making the opening hours easier to find is a reasonable suggestion to investigate. “Customers want you to open seven days a week” goes further than those comments support. A question about Sunday is not a request to open every day.

That gives you a small job to try: put the opening hours somewhere easier to find, then watch whether those questions become less common. You have a reason for the change and something to look for afterwards.

Both recommendations could sound tidy in a report. Looking back at the comments gives you a way to judge them. A tool that cannot show that connection has left you with more checking to do.

You do not need to know how the software was built to ask for the evidence behind its conclusion.

If you want to try one small example

The companion activity on the website lets you compare two customer-feedback recommendations, then spot an added delivery promise. The examples are fictional. You don't need an AI account, and the exercise sends no messages.

Or stop here. The useful thing to take away is a question: what came from the original material, and what has the system added?

Is there an AI feature or claim you keep hearing about but cannot quite picture? Reply to Millie with one you would like explained. Leave out private details.

Millie

Someone forwarded this to you? Subscribe to One Eye on AI.

Sources and corrections

The links lead to the project's announcement, its June research preprint and Google's own instructions. The customer comments and recommendations above are fictional teaching examples, not a live AI test. Reply if something needs correcting.