The Font Designers Created to Poison AI

Some type designers are not trying to make the prettiest font in the world. They are trying to make machine-learning systems choke on it.

The idea is simple enough to understand: alter letterforms so they still look readable to a person but create confusing training data for image models. Creative Bloq has reported on free fonts designed to “poison” AI, and the story is interesting because it turns typography into a tiny act of resistance.

A letter has two audiences now

Humans read shapes by context. We can cope with a strange “a” or an uneven curve because the word gives us clues.

Image models do not read quite like that. They learn from patterns, repeated forms and examples. A small change that is harmless to a person may be noisy to a system collecting visual data.

The practical lesson is not “make ugly type”

It is that type has behaviour. A font is not just a mood. It affects reading speed, cutting, engraving, legibility, licensing and now the way images are interpreted by machines.

For a maker, this is a reason to test lettering physically. Print it small. Cut it. Engrave it. Turn it into a stencil. If a letter only looks good at poster size, it is not ready for a product.

Try this: choose one word and test three versions at the actual size you plan to make it. The best font is the one that survives the material.