Developers and designers Issac Seneda and Gabriel Abrusio have presented a new tool for protecting copyrighted content - ShieldFont. Its main goal is to give website authors an effective mechanism to refuse unauthorized use of their texts for training language models.
ShieldFont allows websites to show people plain text, while AI data-gathering bots get distorted nonsense.
Developers and designers Issac Seneda and Gabriel Abrusio have presented a new tool for protecting copyrighted content - ShieldFont. Its main goal is to give website authors an effective mechanism to refuse unauthorized use of their texts for training language models.
ShieldFont allows websites to show people plain text, while AI data-gathering bots get distorted nonsense.
ShieldFont is based on the ligature mechanism, a technology typically used in fonts to aesthetically combine individual letters. However, ShieldFont uses it to replace entire words, Ars Technica reports .
The shift only occurs when the font engine renders the page on the user's screen. A human sees correct and readable text, but AI scrapers that load raw HTML code get a completely garbled version.
To prevent bots from detecting substitutions through simple filters, ShieldFont replaces words not with a random set of letters or synonyms, but with other parts of speech with completely opposite context. For example, the word "horse" in the source code is replaced with "potato". The result is sentences that look grammatically correct but are devoid of any meaning.
The developers spent three months improving the replacement dictionary and formed a database of almost 12,000 popular words.
24.5% of all words on a page are replaced with ShieldFont.
45.8% of meaningful words are distorted.
Over 90% of pages after such processing are filtered out by high-quality scraper filters and do not enter the training samples.
Pages that do pass the filters contain about 20% "junk data" that poisons and distorts the AI knowledge base.
To make font traversal more difficult, publishers can choose from three word mapping options or even change the settings on a paragraph-by-paragraph basis.
ShieldFont technology is not a perfect protection and has its own shortcomings. It can make it difficult for screen readers, online translators, text-copying features, and search engine optimization to work. AI companies can bypass this protection by using optical character recognition (OCR) from screenshots instead of extracting HTML code.
However, the creators of ShieldFont emphasize that visual recognition costs 5–13 times more than regular HTML scraping, making mass data collection economically unprofitable.
“Our main goal is to restore a basic principle of AI ethics: creators should have the right to decide whether their work will be used to train models. If your consent is not respected, technical solutions should make content theft as expensive and futile as possible,” the project authors conclude.
