The machine reads the outline and misses the shadow

Your newest reader has never held a pencil, reads a million images a day, and, it turns out, can be fooled by a shadow.

That is the short version of “Sorry Robot, Happy Human: Vision-Language Models Read Only One of Two Legible Typographic Layers” by Mert İncidelen, Yamen Kashkash, Asya Berker and Murat Aydoğan (arXiv:2609.31403). It went online on 25 September 2026 and, the authors say, was accepted to the DocInsights 2026 workshop at EMNLP. It is a preprint, so hold it lightly. But it is a lovely experiment.

What they did

They made 300 pictures, each with two texts on top of one another: one drawn in crisp outlines, one in soft shading underneath. They called the set DecoyBench. Then they showed the pictures to people, and to six recent AI models from three families, at a large size (512 by 512 pixels) and a tiny one (64 by 64).

What happened

People did fine. In the paper’s words, “human participants could read both text layers with high accuracy.”

The machines read the crisp outline “with near-human accuracy at high resolution, but almost never fully extracted the shading text.” Shrink the picture and the result turns upside down: nobody, human or model, could read the outline any more, while the shaded text “could be extracted with high accuracy.”

The authors put it carefully: the models show “a consistent behavioral limitation when processing typographic structures containing multiple spatial frequency layers.” Put plainly, fine lines and soft tones are different kinds of detail. We see both at once. These models, in this test, pick one.

Why it matters

More and more of what we design is read by machines first: search engines, accessibility tools, AI assistants describing an image to someone who can’t see it. If your poster says one thing in its lines and another in its shadows, a machine may only pass on half the message.

It is also a gift to designers. Hiding one message inside another is an old trick (think of type that only appears when you step back). Now there is a way to measure it.

From the studio

We’d run a client’s key visuals, posters and thumbnails through AI vision tools and show exactly what the machines read and what they miss, next to what people see. Then we’d redesign the pieces that fail, so they work for both kinds of reader. That is part of our AI and generative work: using the tools, and knowing where they break.

From the institute

The foundations are in Typography (contrast, hierarchy and legibility) and Generative Typography (type built from layers and rules). Digital Art & Artificial Intelligence and AI in Visual Storytelling cover how these systems read images.

Try this: make one poster with a single layer of text and another with a hidden shaded layer. Show both to three friends and one AI tool, at full size and shrunk to a postage stamp. Write down who read what.