I wrote recently about using Ai to help Transcribing handwritten diary entries into Day One. On the whole this has been a great success but there have been one or two interesting hallucinations.
A Blank Page
Not every page in my diary has a written entry but I have been scanning and transcribing them all for completeness. This is the prompt that I have been giving Google Gemini:
You are an expert palaeographer and archivist. Transcribe the handwritten text from this image.
Crucial: Do not insert a line break at the end of every handwritten line. Instead, join the text into continuous, flowing paragraphs just like standard prose. Only insert a line break (a new paragraph) where the author has clearly started a brand new paragraph block or section in the journal. Preserve original spelling, punctuation, headings, and dates. Do not add any conversational intro/outro filler.
For the most part this has been working well and when it finds a blank page it (usually) gives a response such as this:
There is no handwritten text visible in the provided image. The page appears to be blank, with only pre-printed journal headings present.
November 9th
November 9th 1988 was one of those blank days, as you can see from the image above. However, rather than return a blank response or some text telling me that there was nothing to transcribe I got this entry instead:
I have absolutely no idea how or why Gemini decided to write this and it is certainly factually incorrect: in 1988 I was living on my own so no family to do washing-up for, I don’t drink coffee and certainly don’t eat Marmite!
The moral of this story? Don’t blindly trust Ai!
That’s really fascinating, it’s amazing how much AI can pick up even with the blanks. It makes you think about the biases built into the data too.