AI watermarking doesn't visibly alter text. Instead, it uses a secret key to slightly boost the probability of certain words appearing in a sequence. This creates a statistically significant pattern that can only be detected by a tool with access to the original model and the secret key.

Effective AI detection frames the problem as large-scale authorship identification. By training on paired examples of human versus LLM-generated text on the same prompts, detection models learn to recognize the unique statistical "smell" or stylistic signature of each major AI model.

As AI models improve, detecting AI-generated content will become increasingly difficult. A more sustainable long-term strategy may be to focus on verifying and labeling authentic, camera-captured content. This flips the problem from an arms race of detection to a system of verification.

Creating reliable AI detectors is an endless arms race against ever-improving generative models, which often have detectors built into their training process (like GANs). A better approach is using algorithmic feeds to filter out low-quality "slop" content, regardless of its origin, based on user behavior.

AI detection can identify text from new LLMs because most models share a common "ancestry." They are either trained on the same foundational corpora, like Common Crawl, or fine-tuned with synthetic data from major models, giving them a detectable shared statistical fingerprint.

Pangram Labs' detector isn't hard-coded. It's a deep learning model trained on millions of examples. For each human text (e.g., a Yelp review), it sees an AI-generated equivalent, learning the subtle, often inarticulable, differences in word choice and structure that separate them.

Anthropic's Claude is weaving invisible watermarks directly into its text output, making simple copy-pasting detectable as AI-written. This fundamentally changes the tool's use case for marketers, repositioning it from a final content generator to an assistant for creating initial drafts or first-pass edits that require significant human revision.

To distinguish between light AI assistance (like Grammarly) and heavy generation, advanced detectors analyze the "cosine difference"—the distance in a multidimensional space between the original human text and the AI-edited version. This quantifies the degree of AI influence.

Heuristics for spotting AI writing, like the overuse of em dashes, are becoming obsolete as models learn from human feedback. For instance, ChatGPT now uses em dashes *less* frequently than human writers at The Economist, flipping the old tell on its head and complicating detection efforts.