OpenAI has publicly acknowledged that the em-dash has become a "neon sign" for AI-generated text. They are updating their model to use it more sparingly, highlighting the subtle cues that distinguish human from machine writing and the ongoing effort to make AI outputs more natural and less detectable.

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.

The New York Times test showing readers prefer AI writing misses the point. The critical question for professionals is determining when to use AI. A useful framework involves a spectrum from "all human" for personal, creative work where the process is the purpose, to "all machine" for repetitive, high-volume tasks.

In the age of AI, the new standard for value is the "GPT Test." If a person's public statements, writing, or ideas could have been generated by a large language model, they will fail to stand out. This places an immense premium on true originality, deep insight, and an authentic voice—the very things AI struggles to replicate.

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.

In an experiment, a professional writer's colleagues couldn't reliably distinguish his satirical column from an AI-generated one. Some even preferred the AI's version, calling it more coherent or closer to his style, revealing AI's startling ability to mimic and even improve upon creative human work.

For an AI detection tool, a low false-positive rate is more critical than a high detection rate. Pangram claims a 1-in-10,000 false positive rate, which is its key differentiator. This builds trust and avoids the fatal flaw of competitors: incorrectly flagging human work as AI-generated, which undermines the product's credibility.

Early AI detectors used "perplexity," a measure of how surprising text is to a language model. This method is flawed because while AI text is predictably low-perplexity, so is text from non-native English speakers who take fewer linguistic risks, leading to a high rate of false positives.

While the em dash is a known sign of AI writing, a more subtle indicator is "contrastive parallelism"—the "it's not this, it's that" structure. This pattern, likely learned from marketing copy, is frequently used by LLMs but is uncommon in typical human writing.