Comment by MathMonkeyMan
4 hours ago
Yes but they're quite good at it. Reliable OCR is font dependent, whereas I think a lot of models just kind of figure it out regardless.
4 hours ago
Yes but they're quite good at it. Reliable OCR is font dependent, whereas I think a lot of models just kind of figure it out regardless.
One reason I don't quite trust AI for OCR is that it will, on occasion, hallucinate the output.
All OCR is untrustworthy. But sometimes, OCR is useful. (And I've heard it said that all LLM output is a hallucination; the good outputs are just hallucinations that fit.)
A few months ago a warehouse manager sent us a list of serial numbers and the model numbers of some gear they were using -- with both fields being alphanumeric.
This list was hand-written on notebook paper, in pencil. It was photographed with a digital camera under bad lighting, and that photograph was then emailed.
The writing was barely legible. It was hard to parse. It was awful. It made my boss's brain hurt trying to work with it, and then he gave it to me and it made my brain hurt too.
If I had to read this person's writing every day I would have gotten used to it eventually, but in all likelihood I'll never read something this person has written ever again. I didn't want to train myself for that and I didn't have enough of a sampleset to train with, anyway.
And if it were part of a high-school assignment it would have been sent back with a note at the top that said "Unreadable -- try again."
But it wasn't a high school student, and I wasn't their teacher. They were a paying customer and this list was worth real money to us.
I shoved it into ChatGPT and it produced output that was neatly formatted into a table just as I specified with my minimal instruction ("Read this. Make a table.").
The quality was sufficient to allow us to fairly quickly compare the original scribbles to the OCR output, make some manual corrections that we humans knew how to do (like "6" was sometimes transposed with "G"), and get a result that worked for what we needed to accomplish without additional pain.
0/10. I'm glad it worked and I hope I never have to do that again, but will repeat if I must.
There was a good talk some years ago at some of the CCC events where some guy found out that scanners sometimes change numbers on forms.