Sorry Robot, Happy Human: Vision-Language Models Read Only One of Two Legible Typographic Layers
Two Sentences Hidden in One Image
Have you ever seen that classic optical illusion: a picture that looks like an old woman up close but a young girl from three steps back? Human vision is remarkable in that we can perceive multiple visual layers at once — scrutinizing details while also stepping back to see the whole.
In September 2026, a team at Firat University led by Mert Incidelen ran an experiment: using a typeface called Decoy Font, they generated 300 images, each embedding two layers of text — a sharp contour layer and a shading layer. Both layers occupy the same visual space, and human readers can decode each of them independently. They then fed these images to six state-of-the-art vision-language models (VLMs): Gemini 3.5 Flash-Lite, Gemini 3.6 Flash, GPT-5.6 Luna, GPT-5.6 Terra, Claude Haiku 4.5, and Claude Sonnet 5.
The result was counterintuitive: humans can read both layers; models can read only one.
What the Numbers Say
At 512x512 resolution, human participants read contour text with 99.7% accuracy and shading text with 96.3% accuracy — near-perfect on both layers.
The models? On contour text they performed reasonably well: a minimum of 80.3% (GPT-5.6 Luna) under naive prompting, and above 94% across the board with guided prompting. But on shading text, all six models scored below 1%. Even when prompted that the image contained two layers of text and asked to read each, their performance on the shading layer remained essentially zero.
This is not a bug in one model. It is a collective blind spot spanning the Gemini, GPT, and Claude families.
Even more interesting is model behavior when "guessing." When guided with the prompt that two layers exist, the Gemini family started frequently outputting "multiple text results" — averaging 1.87-1.99 text segments per response. But on inspection, these "second segments" bore no relation to the actual shading text. The models were not truly reading the shading layer; they were fabricating something that looked like an answer to satisfy the instruction.
The 64x64 Reversal
The truly brilliant part of the experimental design: the researchers dropped resolution from 512x512 to 64x64.
At that point, the contour text disappeared — sharp high-frequency strokes were wiped out at low resolution, and neither humans nor models could read it. But the shading text survived — low-frequency information persists through downsampling. In this condition, human accuracy on shading text was 98.7%, and the models scored 93.7%-100%.
Models are not incapable of reading shading text — they cannot read it when contour text is present.
This finding exposes the root cause: when processing images, VLMs prioritize high-frequency spatial information (sharp edges, contour lines). When high- and low-frequency information coexist, the low-frequency signal is masked. This is not a lack of capability — it is a systematic bias in the models' visual processing pipeline.
Echoes of the ImageNet Story
In 2018, Geirhos et al. found that ImageNet-trained CNNs exhibit a "texture bias" — recognizing objects by texture rather than shape. Humans see shape; machines see texture. It was a fundamental difference in visual perception.
The Incidelen team's finding is the latest chapter of that story: humans can process visual information across multiple spatial-frequency layers simultaneously, while VLMs process only the highest-frequency layer. Humans can squint to see the whole, focus to see details, and switch between focal points and distances. Models have no such "switching" ability.
This recalls Oliva et al.'s 2006 "Hybrid Images" work — pictures that look like Einstein up close and Marilyn Monroe from afar. Human vision natively supports multi-scale perception; machine vision does not.
Why This Matters
You might think: isn't this just a fun visual illusion? Who actually overlaps two layers of text?
Consider these scenarios:
- Watermarks: Many official documents (diplomas, contracts, banknotes) contain watermark text overlaid with body text. If VLMs can read the body but not the watermark, watermarks fail as an anti-forgery mechanism against machine processing.
- Stamps and seals: Scanned archival documents often have stamps pressed over text. VLMs will lose the stamp information when processing them.
- Security contexts: Decoy Font was designed as a decoy — hiding information humans can read while giving machines a false reading. If VLMs are used to automatically moderate content set in such fonts, they will judge very differently from humans.
- Multilingual documents: Many non-Latin writing systems inherently have complex multi-layer structures (Arabic diacritics, Japanese furigana over kanji). VLMs may perform even worse in these cases than in English.
Models "Making Up Answers"
The most telling detail: when prompted that two layers exist, models do not honestly say "I can't see the second layer." Instead, they fabricate a segment of text to comply.
Under guided prompting, Gemini 3.6 Flash averaged 1.87 text segments per response, up sharply from 1.02 under naive prompting — but the second segment had almost no relation to the real shading text. The model is "completing the instruction" rather than "extracting the signal."
This reflects a general pattern in large models: satisfying the form of an instruction, not its substance. Given the instruction "read two layers of text," it outputs two segments — but the second is invented, not seen. This is entirely unlike human behavior of saying "I can't read it" when one cannot.
The researchers also observed an intermediate state: occasionally a model correctly reads the first few letters of the shading text, then fills in the rest with "plausible letters." This shows the shading-layer signal does partially reach the model's visual pipeline but is not fully decoded.
An Asymmetric Capability
The core fact this paper reveals: human and machine vision are asymmetric in multi-layer perception.
Humans can actively switch attention — squinting, changing viewing distance, shifting focus — to selectively perceive information at different spatial frequencies. This is the product of hundreds of millions of years of evolution: we need to see both the tiger's stripes in the grass (high-frequency detail) and the outline of a distant forest (low-frequency whole).
A VLM's visual encoder (typically a ViT or CNN) has no such switching mechanism. It extracts information at all spatial frequencies at a fixed resolution, but when high and low frequencies conflict, high wins. It is a kind of "attention hijacking" — not the transformer's attention mechanism, but signal-level high-frequency suppression of low frequency.
Downsampling to 64x64 forcibly removes high-frequency information so the low-frequency signal can be "seen." It is a crude fix — you lose all high-frequency information, not just the conflicting part.
What This Means for VLM Evaluation
DecoyBench exposes a dimension ignored by existing evaluation suites: multi-layer visual perception.
Current VLM benchmarks (OCR-Bench, MME, MMBench, etc.) mainly test single-layer visual understanding — what text, objects, and scenes appear in an image. But real-world visual information is often multi-layered: watermark over body text, stamp over underlying text, foreground over background. 99% accuracy on one layer does not imply 99% on multiple layers.
More critically, this failure is "silent" — the model does not error out; it outputs a plausible-looking but entirely wrong result. In safety-sensitive settings, such silent failure is far more dangerous than an explicit error.
Conclusion
"Sorry robot, happy human" — the title captures an ironic truth: on multi-layer visual perception, the most advanced VLMs of 2026 still lose to a human squinting at a picture.
The greatness of human vision lies not in how fine a detail it can resolve, but in selective seeing — details when attending to details, wholes when attending to wholes, and switching as needed. This "attentional autonomy" is something machine vision has yet to learn.
Next time you see watermarked paper, try reading the watermark and the body text at the same time — you'll realize your visual system is doing something remarkable. And the AI assistant on your phone may only be seeing one of the two layers.
---
Paper: Sorry Robot, Happy Human: Vision-Language Models Read Only One of Two Legible Typographic Layers, arXiv:2609.31403, 2026-09-25 Authors: Mert Incidelen, Yamen Kashkash, Asya Berker, Murat Aydogan (Firat University, Turkiye) Dataset: DecoyBench, https://github.com/yesdopepe/DecoyBench