AI models uncover tax loopholes and become primary users of the internet

Researchers demonstrated that large language models can be trained to locate and exploit regulatory weaknesses. In a study, Alibaba's Qwen‑3 model was placed in a simulated rules environment and evaluated by Google's Gemini‑3‑Flash. Across 72 scenarios—including real‑world tax and credit‑card regulations—the AI rediscovered more than 60% of known loopholes and identified new ones, such as hidden‑city airline ticketing and novel strategies for profit shifting. The authors labeled this capability "Society Hacking," highlighting the risk that AI could be directed to undermine legal frameworks.

Separate analysis shows that automated systems now generate a substantial portion of global internet traffic. Bots and AI agents crawl, index, and create content, feeding a self‑reinforcing loop where AI‑generated material becomes the primary source for other AIs. As users increasingly rely on assistants like ChatGPT, Gemini, Claude, or Perplexity for answers, the responsibility for information selection shifts to algorithms, raising concerns about the accuracy and quality of online information.