In a world increasingly filled with AI-generated text, how can we tell what’s real and what’s fake? This is especially critical with the rise of deepfakes, where malicious actors manipulate or forge content to point fingers at innocent parties. Imagine AI-generated news articles subtly altered to spread misinformation, or customer service chatbots spewing hate speech that wasn't programmed. Current methods of "watermarking" AI text, designed to identify its source, are robust against attempts to remove the watermark, but they have a critical flaw: they're vulnerable to spoofing. Think of it like a easily copied signature on a valuable painting—it proves authenticity unless someone forges it. Researchers have now developed a new technique called "Bileve"—a bi-level signature that’s much harder to fake. It works like a double-layered security system. The first layer is a coarse-grained signal embedded throughout the text, acting as a general marker of AI origin. This layer is robust, surviving even if the text is slightly altered. The second layer adds a fine-grained, content-dependent signature to each word, like a unique fingerprint. This makes it incredibly difficult for bad actors to tamper with the meaning without leaving traces. This bi-level approach allows Bileve to distinguish five different scenarios: genuine AI-generated text, text subtly manipulated with the original watermark intact, AI-generated text with improved safety mechanisms, tampered AI-generated text, and text entirely from another source. This granular detection ability is a game-changer for verifying the integrity of AI-generated content. While promising, Bileve isn’t perfect. The precise nature of the word-level signatures can make the AI’s writing slightly less fluent, and the system needs further optimization for efficiency. But, it represents a significant leap forward in the battle against deepfakes and ensuring we can trust the information we receive from AI.