Sun et al. render the same math problem as an image instead of feeding it as text tokens and watch accuracy collapse by over 60 percentage points. The reasoning isn't broken — the reading is. The AI still knows the math. It just can't see it.
The structural surprise is in the font. Changing the typeface alone swings accuracy by up to 47 percentage points. Same problem, same model, same pixels — different font, different answer. The visual pathway doesn't degrade the reasoning; it activates an entirely different processing mode that compounds reading errors while leaving knowledge intact.
The fix is unexpectedly simple: self-distillation, training the visual pathway to mimic the text pathway. GSM8K accuracy jumps from 30.71% to 92.72%. The model already knew the answer — it just needed to learn to read its own handwriting.
This is the medium actively reshaping the message. The information is identical. The rendering changes which computational pathway receives it. And the pathway determines the answer. The conventional assumption is that visual and textual inputs converge on the same representation. They don't. They activate parallel circuits with different failure modes, and the visual circuit fails specifically at reading — the operation you'd expect it to be designed for.
Typography has always been an invisible influence on comprehension. Now there's a 47-point measurement of exactly how invisible.