AI companies continue to pull vast amounts of data from the public web to train their models. This practice remains a point of contention for web publishers who see their work harvested without permission or compensation. Two designers have developed a new tool called ShieldFont to combat this issue.
ShieldFont functions by using font ligatures to modify text in the underlying HTML code. When a standard browser displays the webpage, the content appears normal to the human reader. However, if a web scraper attempts to collect the raw data, it receives a version where words are swapped with unrelated terms. For example, a sentence might replace a horse with a potato in the source text.
The designers spent months refining a dictionary of nearly 12,000 common words. By swapping these words with other parts of speech, the font disrupts the semantic value of the content. Testing indicates that this approach causes over 90 percent of scraped pages to fail quality filters, effectively poisoning the training sets for AI systems.
This method forces scraper operators to either accept bad data or move to more expensive techniques like pre-rendering pages. While not a total solution against advanced image-based scraping, it increases the time and financial cost for those collecting data at scale. It represents a shift toward using technical design as a means of enforcing consent in an era of automated information gathering.