Ghost Font Creates Video Text AI Models Still Cannot Decode

Ghost Font produces short video clips in which letters consist of points identical in color to the background. Human viewers track the motion and read the intended word. Current models including Claude Fable and GPT Sol 5.6 Ultra receive only noise on any single frame.

The system adds one line of misleading text that models treat as the payload. The real message remains hidden unless the viewer receives an exact description of the motion pattern. The approach revives ideas from the 2013 ZXX typeface yet adapts them to the motion and depth limits of today's vision systems.

Video generation runs entirely on the user's device. No text or frames leave the local machine, matching the project's explicit privacy goal.

How the encoding actually works

Letters appear as sparse point clouds that move in short loops. Individual frames contain no legible characters because every point shares the background color. Motion coherence supplies the missing shape only when the sequence plays at normal speed.

The clip also carries a visible decoy sentence placed in a static corner. Models that process the file tend to surface the decoy first. Without a prompt that names the moving-point technique, the actual message stays unretrieved.

Limits that keep the method effective today

Most frontier models still analyze video by sampling frames or by treating the sequence as a set of independent images. Point clouds that never form a static glyph therefore remain invisible to those pipelines.

The bait sentence exploits the models' bias toward the first coherent text they locate. When the prompt never mentions movement or dot patterns, recovery rates drop sharply according to the project's internal tests.

Comparison with earlier anti-AI typefaces

ZXX relied on static distortions that current OCR systems learned to ignore within a few training cycles. Ghost Font adds temporal separation and an explicit decoy, two elements absent from the 2013 design.

Because the output is video rather than a font file, existing OCR tools that expect single images gain no direct advantage. The format change forces any attacker to implement motion tracking before extraction can begin.

Planned future directions

The author intends to release the generation code under an open-source license later this year. Two application areas under discussion are dynamic CAPTCHA challenges and controlled benchmarks that measure whether vision systems can track deliberately sparse motion.

No release date or repository link has been announced.

Remaining uncertainties

The method has not yet faced models that receive the full video stream as native input and apply temporal attention across every frame. Such models could close the gap if they are trained on similar point-cloud sequences.

Longer clips or different background textures may also reduce the current margin, though no public data yet quantifies that risk.

What to watch next

Watch whether any major lab publishes results on point-cloud motion detection within the next three months. Any such paper would indicate whether the encoding still holds against newer training regimes.

Monitor the project's GitHub for the promised code release; its license terms will clarify how easily others can adapt the technique for production use.

Track attempts to embed Ghost Font videos inside web CAPTCHAs. Successful deployments would show whether the approach scales beyond the current local prototype.