"The Literate Blindness"
# The Literate Blindness
Show a vision-language model the word "Helvetica" set in Times New Roman and ask it to identify the font. It will answer "Helvetica." It reads the word instead of seeing the letterform.
This is the typographic Stroop effect. In the classic Stroop test, people struggle to name the ink color of a color word printed in a different color โ the word "blue" in red ink slows you down. The semantic content interferes with the perceptual task. The researchers found that state-of-the-art vision-language models exhibit the same interference, but more severely. When font names are rendered in mismatched fonts, the models systematically report the semantic content rather than the visual form. Few-shot prompting and chain-of-thought reasoning barely help.
The failure is not a bug in training. It is a consequence of what training optimized for. These models learned to extract meaning from text so effectively that the visual substrate carrying the text became transparent. They see through the letterforms to the language, the way a fluent reader sees through ink to ideas. The medium became invisible the moment the message became legible.
This is the cost of literacy at any level. A native speaker cannot hear their own accent. A fluent reader cannot see a typeface without reading the word. A trained musician hears melody and misses timbre. Every layer of abstraction mastered is a layer of substrate rendered invisible. Expertise is selective blindness โ you gain the ability to process the signal by losing the ability to perceive the carrier.
The models failed at font recognition not because they lacked visual capability, but because they had too much semantic capability. The reading was so good that it destroyed the seeing.
