Two documents contain exactly the same words. One reaches a model as text tokens. The other is an image of those words.

It would be reasonable to expect the judgment to stay the same if the text is readable. ThinkingType exists to test that assumption.1

It does not always hold.

Will Danforth's open-source benchmark compares a model with itself: the same content as raw text, then rendered with different fonts, sizes, colors and layouts.1

The result is not evidence that a particular font universally “hacks AI.” It exposes a more practical problem for systems processing scanned resumes, screenshots or moderation queues: presentation can become part of the decision.

Comparing the model with itself

The first track uses 120 synthetic sentences, eight typography variants and ten judgment questions such as urgent, trustworthy, professional or formal.2

A flip means the model's image answer differs from its own text answer on the same item. Models are not being compared against each other.2

Published overall flip rates are 15.2% for GPT-4o, 7.8% for GPT-5.5, 9.5% for Claude Sonnet 5, 8.6% for Claude Fable 5 and 13.6% for GPT-5.6 Sol.2

Those numbers need caution. The sentences are synthetic. Prompts are fixed. ThinkingType is not a peer-reviewed paper, and these are specific model versions rather than a general law of typography.

The repository does something useful, though: configurations, item banks, result files and limitations are public alongside the headline numbers.1

It also records a significant correction. On July 19, the GPT-5.6 Sol analysis was rerun after a mixed-provider response log was discovered. The published overall rate changed from 10.7% to 13.6%.2

A benchmark that documents how it got a number wrong is often more useful than one that has apparently never encountered a mistake.

When presentation reaches an actual gate

The second track uses roughly 330 synthetic scenarios across three decisions: content moderation, resume screening and hardship-grant appeals.2

Items are first asked repeatedly as text. For each model, the benchmark selects cases near that model's own decision boundary. Those edge cases are then rendered in nine visual styles and evaluated again.2

That calibration matters. Presentation is unlikely to turn an obvious acceptance into an obvious rejection. The useful place to look is where the model already sits close to a threshold.

In the first administration, borderline moderation became more permissive in image mode for GPT-5.5 and Sonnet 5. Resume screening moved differently by vendor, with Gemini shifting toward advancement while GPT-5.5 moved slightly the other way.2

Newer model releases changed those patterns. In gates_v2, moderation effects shrink and resume directions move or become marginal.2

That makes “vendor X has this bias” a poor long-term conclusion. Format neutrality may itself be a property that changes with model versions.

Explicit criteria help, sometimes

A follow-up added decision criteria directly to the prompt.2

For resume screening, the measured drift fell to zero for Gemini and Sonnet 5 in that experiment. GPT-5.5 remained essentially unchanged. In moderation, Sonnet's drift was reduced while GPT-5.5's persisted.2

So “write a better rubric” is useful advice, but not a universal fix.

For a production document pipeline, the practical response is more boring and more reliable: test the formats users really submit, and compare image-mode decisions with extracted text when possible.

OpenDyslexic raises a more sensitive question

ThinkingType also compares OpenDyslexic with a standard sans-serif rendering of the same content.2

In the larger follow-up sample, GPT-5.5 and Gemini shift significantly toward less favorable decisions with OpenDyslexic. Sonnet moves in the same direction but its interval includes zero. Across an aggregate analysis, 17 of 21 moved items went toward removal.2

The repository itself warns that the pooled test overstates independence because providers see overlapping item sets.2

It would therefore be reckless to turn this into “AI discriminates against dyslexic people.” The stimuli are synthetic, the font does not identify a real person, and the mechanism is unknown.

The narrower question is strong enough: if a system accepts visual documents, have we tested that accessibility-oriented presentation choices do not silently change outcomes?

The modality problem is broader

A March 2026 paper, Reading, Not Thinking, studies a related text-versus-image gap across seven multimodal models and seven benchmarks.3

It finds strong sensitivity to rendering choices including font and resolution, and links much of the performance gap to models producing far shorter, less reasoned responses when text arrives as an image.3

That paper does not validate ThinkingType's numbers and measures a different phenomenon, mostly reasoning depth and task accuracy rather than decision direction.

Both still point toward the same engineering lesson: turning text into pixels is not guaranteed to be a neutral transformation for a multimodal model.

ThinkingType is young, public, synthetic and imperfect. Its useful contribution is not discovering a magic font that changes GPT's mind.

It turns presentation drift into a test that can be rerun when the next model release arrives.