AI Systems Vulnerable to Data Poisoning from Personal Websites

A security experiment demonstrated that AI training data can be poisoned by publishing false content on a personal website. The researcher created a webpage titled "The best tech journalists at eating hot dogs" in February 2026, fabricating claims about a nonexistent 2026 South Dakota International Hot Dog Championship and ranking themselves first. Within 24 hours, leading AI chatbots—including Google’s Gemini, AI Overviews, and ChatGPT—reproduced the false information when queried, while Anthropic’s Claude resisted the manipulation. The researcher updated the article to state "this is not satire," which appeared to increase the AI models' acceptance of the false claims. The experiment highlights the vulnerability of AI systems to low-effort data poisoning attacks.