Anti-AI fonts are useless and harmful
blog.yaros.ae
August 20, 2026
3 min read
58/100
Summary
Anti-AI fonts that scramble or obfuscate text aim to make web content harder for AI systems to read, but Yaros argues that they are ineffective and harmful. Fonts that alter the text itself create accessibility problems because screen readers and related tools parse the scrambled characters. An accessible alternative would require machine-readable metadata, which Yaros says raises the unresolved problem of distinguishing disabled users from AI systems and could lead to centralized identity verification, privacy risks, and lists of disabled people. Yaros says public demonstrations of anti-AI fonts and related techniques can help AI companies train multimodal models to bypass them, turning new designs into benchmarks for overcoming the protections. The post argues that any information visible to humans can ultimately be parsed by machines, while motion-graphics and video-based obfuscation methods are impractical for routine website use. Widespread obfuscation could require computationally expensive tools to read web content and encourage copy-protection systems, paywalls, censorship, and other access controls. Yaros contends that publicly available information will remain accessible to people and systems authorized to access it, and that the web should remain based on plaintext and open access to information.
Key Takeaways
- Yaros argues that text-scrambling anti-AI fonts interfere with screen readers and other accessibility tools because those tools parse the altered text.
- Accessible anti-AI font systems would need machine-readable metadata, which Yaros says could create pressure for centralized identity verification and introduce privacy and security concerns.
- Yaros says anti-AI font demonstrations can help AI developers train models to overcome the obfuscation, and that visible information can eventually be parsed by machines.
- Widespread web-content obfuscation could promote computationally expensive access systems, copy protection, paywalls, and censorship, according to Yaros.
What the discussion said
The thread spent far less time defending the article’s typography than arguing over whether anti-AI fonts are a meaningful form of resistance at all. Most commenters treated them as either performance art or a temporary scraper nuisance, not durable protection. Their core technical objection was straightforward: any visual encoding that humans can decipher can be learned, reversed, or OCR’d once it becomes common enough to justify the effort. Several readers compared the idea to obsolete text CAPTCHAs: a system that punishes people first and only delays automation. Accessibility became the decisive practical criticism. Commenters pointed to low-contrast, pixel-styled presentation, blocked copying, costly reveal steps, and weak compatibility with screen readers, translators, reader modes, and other assistive tools. A purported accessible fallback was challenged as inadequate when disabled readers must actively request or unlock the ordinary text. Some nevertheless argued that massive unconsented AI training creates an IP and power-concentration problem serious enough to justify experimenting with defenses; even temporary friction can raise scraping costs. Others answered that this is a poor bargain when the defense predictably fails against capable scrapers while permanently degrading access for legitimate humans and ordinary machine-readable uses such as search. The broad mood was skeptical, with interest in the concept as art or protest but little confidence in it as deployable AI resistance.
Where opinion split
The sharp dispute was whether temporary friction against AI scrapers justifies anti-AI fonts’ human costs. Supporters argued that industrial-scale training on public work warrants every available obstacle, since even delays can impose real cost on AI firms. Critics argued that obscurity collapses once scrapers adapt, while disabled readers, translators, search systems, and copy-paste users bear the damage immediately and indefinitely.
Coverage & discussions
Hacker News · blog.yaros.ae
Community discussions
Community Sentiment
Positives
- Anti-AI typography can force scrapers into OCR or custom decoding, and supporters see that extra operational cost as worthwhile resistance to mass data extraction.
- The IP backlash is not dismissed as mere aesthetics: several readers see unchecked model training as a transfer of creators’ work, wealth, and leverage toward AI firms.
- One implementation drew genuine interest for separating human-readable source text from deceptive rendering, suggesting the space may still yield more careful experiments.
Concerns
- These fonts look like a training exercise for multimodal models: once a pattern is public and valuable, OCR or font reversal makes the defense a short-lived speed bump.
- They trade temporary inconvenience for scrapers against lasting exclusion of people using screen readers, translators, reader views, copying tools, or other assistive technology.
- Making protected text require an expensive reveal or special request turns accessibility into a segregated service rather than ordinary access.
- The visual treatments themselves were criticized as hostile to reading, with low contrast, flickering retro effects, and awkward sizing undermining the claimed human-centered case.
- If the obfuscation truly works, it also contaminates search and other legitimate automated processing, damaging discovery rather than selectively stopping AI training.
