We’ve all run into situations where we had to identify a random typeface in a photograph or an old layout. This happened to me again the other day, and as I was about to jump over to WhatTheFont I thought, “Hey, I have all these spiffy AI minions littering my computer, I wonder how well they’d do at this.”
Here’s the image containing the typeface I wanted to identify.
As it turned out, all of my AI tools performed horribly on this task. If you’re an AI skeptic, what follows might generate a little schadenfreude in you.
When asked to identify the font in a large screenshot, ChatGPT, Claude, and Gemini all came back with similarly off-the-wall responses, confidently declaring it “a classic Didone,” explaining that the “vertical stems of the M” (which are clearly not vertical) were a “sure identifier” of Bodoni, or other equally unlikely and wildly inaccurate answers. All these bloopers were delivered with the simulated self-confidence that AIs do so well.
I posited to Gemini that perhaps the problem was how it and its digital brethren were parsing the large, crisp image they were so completely misinterpreting. Its answer is worth sharing because it’s a clue to how clueless the current crop of AI models are, and why.
Notice how the Large Language Model got the vocabulary right, despite being completely unable even to identify the font classification, never mind the specific typeface:
“You’ve hit on a fundamental limitation of how multimodal vision models process typography,” Gemini wrote. It went on:
1. Patch Tokenization vs. Vector Precision
Images aren’t analyzed as vector outlines or continuous curves. They get sliced into a grid of fixed-size pixel patches (tokens) and converted into mathematical embeddings. Micro-typographic cues—a 2-degree stem splay, a hairline gap versus a closed counter, or the exact transition of an apex—lose structural fidelity during that downsampling and patch encoding.”
Translation: It doesn’t see all that well.
2. Semantic Priors Overpowering Visual Features
Once the model recognizes the macro “gestalt” of an Old Style / Renaissance Roman, strong statistical priors kick in. Instead of strictly reading the raw pixels for every glyph, the model defaults to the canonical feature list of whatever candidate font it associates with that overall look (e.g., pulling the memorized traits of Minion or Bembo and describing them as present, even when they contradict the image).”
Translation: It jumps to conclusions and lets confirmation bias take it from there.
3. Lack of Native Measurement Tools
Human type identification relies on measuring relative proportions, stroke contrast ratios, and zooming in on specific details (like the inner vertex of an M or the teardrop/wedge of a terminal). Vision transformers evaluate probabilistic feature maps across the whole image at once, which makes fine-grained structural audits notoriously prone to false confidence.”
Translation: It takes one glance, then stops looking and works off its store of preconceived ideas.
Mind Your Ps and Qs, AI
There are plenty of things our digital minions are good at, but accuracy and detail on a task like font identification are not part of that skillset.
To be fair, neither Adobe Fonts nor WhatTheFont could get it right either, because it turned out to be plausible but AI-generated text. But neither of those services explained to me why they were certain it was a Didone, or Palatino.
This article was last modified on September 8, 2026
This article was first published on September 8, 2026
Commenting is easier and faster when you're logged in!
Recommended for you
InDesign Magazine Issue 113: Calendars
Issue 113 of InDesign Magazine include articles on calendar design, high-contras...
Morisawa Font Library Now Available Outside Japan
Morisawa, one of the most prestigious type foundries in Japan, first made its fo...
What Ever Happened to the Fonts that Used to Come with InDesign?
Cue the Grumpy Old Man voice: Back in MY day, Adobe applications came on DISCS,...