Anti-AI fonts are useless and harmful
Efforts to create “anti‑AI” fonts that confuse web scrapers are drawing skepticism over both their effectiveness and their side effects. Commenters argue that any font humans can read will soon be trivially handled by OCR or adapted models, turning these schemes into short‑lived obstacles that mostly break accessibility tools, search, and basic usability for real users. Many see the real leverage against unwanted AI training in legal, policy, and licensing approaches rather than technical obfuscation that risks fragmenting or degrading the open web.
Effectiveness of Anti-AI Fonts
- Many argue these fonts are security theater: anything humans can read, AI can eventually read, especially with OCR or font-reversal.
- Comparison to CAPTCHAs: text-based CAPTCHAs were once hard for machines but are now trivial for models; expectation is fonts will follow the same path.
- Some think fonts might modestly raise scraping costs (e.g., forcing OCR, browser automation), but opponents say scrapers will adapt quickly, especially if fonts become common.
- Criticism that such schemes only work while rare; once widely used, they’ll get custom decoders or efficient kernels.
Accessibility and Human Impact
- Strong concern that obfuscating fonts disproportionately harm people using screen readers, reader modes, custom fonts, or text-only browsers.
- Analogy to making disabled users use the “loading dock” entrance: accessibility that requires extra steps is argued to be de facto exclusion.
- Counterview: some dismiss the accessibility argument as affecting a tiny minority or being an acceptable trade-off to block AI, a stance others strongly condemn.
Shieldfont and Similar Schemes
- Shieldfont is discussed as a more sophisticated approach: screen readers can reveal real text via an expensive “uncover” step; by default, bots get decoy text.
- Critics say this still blocks assistive tech and normal copy/paste/search until decoding, and any “decode” mechanism can be mimicked by scrapers.
- Concerns that search engines will index decoys, effectively sending shielded sites into an SEO “black hole,” though some note social/organic traffic might make that acceptable.
Ethics, Consent, and Ownership
- Some see blocking AI as anti-social: AI “learning” from public text is viewed as akin to indexing and ultimately beneficial to humanity.
- Others argue models represent massive unconsented IP exploitation and power transfer; any friction is justified, even at serious usability cost.
- Debate over whether hosting/scraping costs and asymmetry (users punished for scraping, big AI not) morally justify technical resistance.
Policy vs. Technical Arms Race
- Multiple comments say technical anti-AI measures will devolve into a cat-and-mouse game that humans lose while AIs improve.
- Preference from some for legal, regulatory, or contractual solutions over fragile tech tricks; others argue for “both,” despite arms-race downsides.
- A recurring pessimistic theme: if all public info is inevitably scraped, the open web may degrade or disappear, regardless of font-based defenses.