February 26, 2026
Spot the fake, or spot the bot?
I rendered 1,418 confusables over 230 fonts. Most aren't confusable to the eye
Data says most look-alikes are safe, but commenters see phishing, bad fonts, and maybe a bot
TLDR: A new test across 230 fonts says only 3.5% of look‑alike characters are truly risky, suggesting most “confusables” aren’t. Commenters feud over real‑world phishing (plain old I vs l), surprise at default fonts, and a spicy side plot accusing the author of bot‑written work—raising both security and trust stakes.
A researcher claims a twist in the “fake letter” saga: after rendering 1,418 look‑alike characters across 230 fonts and scoring them with SSIM, only 3.5% looked truly dangerous. Translation: most scary “confusables” aren’t that confusing to your eyes. The crowd? Split—loudly. One camp cheers the data and the no‑AI approach (deterministic math, no training, no GPU), calling it actually reproducible for once. Another camp shouts back: phishing doesn’t need fancy Unicode. As one user put it, uppercase “I” vs lowercase “l” is already a disaster mid‑word—who needs exotic scripts when Arial can fool you?
Then the plot twist: the AI drama. A vocal skeptic insists the piece reads like it was ghostwritten by a chatbot, quipping “Claude Code did it,” while another sleuth points to the author’s prolific posting and wonders if it’s “LLM content” and possibly hallucinated. That accusation becomes the thread’s meme: is this a font study or a bot study? Meanwhile, a surprised chorus fixates on a spicy detail: some felt Microsoft and Mac default fonts didn’t shine, which only fuels the browser‑era fear of mixed fonts picking odd look‑alikes. In short, the science says “mostly safe,” but the comments say “watch your eyes, your URLs, and your bots.”
Key Points
- •The tool “confusable-vision” renders and scores 1,418 TR39 confusable pairs across system fonts using SSIM.
- •Rendering is restricted via fontconfig to fonts that contain each character, reducing work by 97% (8,881 vs 326,140 renders).
- •SSIM provides deterministic similarity scores (-1 to 1), preferred over CNNs for reproducibility and auditability.
- •The pipeline performs 235,625 comparisons in same-font and cross-font modes on 48×48 greyscale PNGs without augmentation.
- •Results classify confusables: High (≥0.7) 49 (3.5%), Medium (0.3–0.7) 681 (48.0%), Low (<0.3) 611 (43.1%), with 96.5% not high-risk.