Instead of detecting AI fakes, a new approach focuses on proving authenticity at the source. Organizations like C2PA work with hardware makers to embed cryptographic signatures into photos and videos, creating a verifiable chain of "content provenance" that proves an asset was captured by a real device.
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.
The economic incentive for AI-generated posts on platforms like Reddit is a B2B service. Startups sell companies the promise of "organic mentions," using AI bots that engage in normal-seeming conversations before strategically recommending or mentioning a client's product.
Pangram Labs estimates that 40% of internet pages are AI-generated. This is largely driven by the SEO industry, which has switched to AI to produce keyword-targeted articles for pennies, flooding search results and platforms like Medium with low-cost, low-value content.
Historically, well-structured, grammatically correct writing served as a reliable heuristic for an intelligent and serious author. AI completely breaks this connection by allowing anyone to generate perfectly polished prose for any idea, no matter how absurd, removing a key filter for evaluating content.
Pangram Labs uses an "active learning" loop to enhance its model. After an initial training, the model scans a massive corpus to identify its own errors (false positives/negatives). These hard-to-classify examples are then fed back into the training set, making the next version more robust.
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.
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.