Fuck AI

I once wrote a rather low resolution 'assisted' OCR program to scan in and convert image documents to plain text. It worked like a crude AI of sorts, and would make mistakes as always, but if the 16x16 grid of low-res pixels fell below a certainty threshold, it would pause and ask the user to type the character(s) it was observing.

I seeded it with a fairly generic font for reference, to start with anyways. But as the program progressed, the model would adapt the recognized font to the point that by like 1/4 of the way through the scanned document, it had a much higher certainty factor and could more or less continue without much human intervention.

By the time the program finished, I could also dump its internal adjusted memory to see what the text font looked like to the computer. If I scaled that up way past 16x16 potato blocks, I could upscale the concept to detect exactly what font was used. Assuming I had a rip of every font that's ever existed anyways..

In the right hands (with models that actually make sense), it's totally possible to know exactly what AI is doing behind the scenes. But still needs human eyes to double and triple check the results before getting the Stamp Of Approval™