Hallucination Watch — a plain-language look at where the new AI tools will fool you, using one real Tacoma photograph.
There is an old photograph of a Tacoma street corner. Real people stand on a real wooden sidewalk in front of a real photographer's studio, around 1888. Above the door, a savings bank. Next door, a real-estate office with a sign painted on its wall, selling lots in a neighborhood a coming financial panic would erase before a single lot sold.
We know all of that is true because the Tacoma Public Library's Northwest Room keeps the photograph and its caption — naming the corner (the southeast corner of Pacific Avenue and 11th Street) and the businesses in the frame.
Then we ran that same black-and-white photograph through two different AI tools that promise to bring old pictures to life in color. Both produced something beautiful. Both got the history wrong in exactly the same way — and it's a way worth understanding, because you are going to see a great deal of it.
What the tools changed
In the real photograph, the real-estate sign advertises lots in a real, documented Tacoma subdivision. One AI tool repainted that sign to read a word that was never there. The other repainted it to read a different word that was also never there. The name of the real-estate firm got rewritten too — into two different wrong names, one from each tool. One tool even added a sign for a business that simply does not exist in the photograph.
Neither invented word appears anywhere in the historical record. The tools didn't misread the signs. They didn't read them at all.
Every one of these changes looks completely natural. That is the danger. The colors are warm, the light is convincing, the letters are crisp and legible. A reader has no reason to suspect that the crisp, legible letters are fiction.
Why this happens — the one idea to carry with you
An AI colorizer does not "add color" to your photograph the way you might tint a drawing. It rebuilds the image, pixel by pixel, guessing what each patch should look like. When it reaches a sign, it doesn't read the text and preserve it — it sees the shape of lettering and guesses at plausible letters to fill that shape.
The model cannot read. So at every sign, every label, every painted word, it invents. This has a name: hallucination. It is not a glitch or a rare failure. It is the normal behavior of these tools whenever they meet text — whether coloring a photo, summarizing an article, or answering a question. The output is confident either way. Confidence is not the same as accuracy, and with these tools the two have nothing to do with each other.
What to watch for
- Treat any text inside an AI image as unverified. Signs, labels, handwriting, dates — assume the tool guessed, and check anything that matters.
- Be most suspicious when it looks best. The more polished the result, the easier it is to stop questioning it. That is exactly when to slow down.
- Find the original. The archive that holds a historical photo almost always holds a caption — the real names, the real place. That caption is the record.
- Two tools, two answers, means neither is the source. When two AI versions disagree, that disagreement is proof that both are guessing. The truth is in the document underneath.
Why we're telling you this
We use these tools ourselves. Used honestly, a colorized image can help a true story reach people who would scroll past a faded gray rectangle. There is nothing wrong with the color. There is something wrong with letting the color make claims — with letting a guessed sign stand in for a real one.
So when The Grey Revolution shows you an image that an AI touched, we label it, and we keep the real source as the thing we actually stand behind. Kindly, but truthfully: the picture may be imagined; the history has to be real.
The next time a gorgeous colorized photo comes across your feed, look at the signs. Then go find out what they really said.
The Grey Revolution documents Tacoma for the people who lived it. This piece is part of our GenAI thread, where we explain the new tools in plain language — what they're good for, and where they'll fool you.