I was inspired by "Thinking with Type" and the recent Decoy Font to share this. This is an early (and hobbyist) analysis of whether VLMs suggest different qualitative judgements and decisions given fonts (and as compared to the same text ingested via an LLM).
Key findings:
1. A model specific vision vs. text bias
2. Fable 5 refusals cluster by font
3. A modest but real impact of certain fonts like Comic Sans and Open Dyslexic on formal assessments/toy decisions.
Hoping to evaluate these further in the coming weeks with higher n and not fully synthetic examples, and would love and feedback or questions
Hi! Full write-up is here: https://wdanfort.github.io/thinkingtype/
I was inspired by "Thinking with Type" and the recent Decoy Font to share this. This is an early (and hobbyist) analysis of whether VLMs suggest different qualitative judgements and decisions given fonts (and as compared to the same text ingested via an LLM).
Key findings:
1. A model specific vision vs. text bias
2. Fable 5 refusals cluster by font
3. A modest but real impact of certain fonts like Comic Sans and Open Dyslexic on formal assessments/toy decisions.
Hoping to evaluate these further in the coming weeks with higher n and not fully synthetic examples, and would love and feedback or questions