AI Content Detection
Classifying whether text was written by AI; as models improve, detection needs constant fine-tuning and hybrid deep-learning-plus-heuristic methods.
Detecting AI-generated text is a constantly moving target. As the latest models from OpenAI, Google, and Anthropic improve, they easily slip past older detectors. To keep pace, detectors must be constantly updated and fine-tuned on the very newest AI outputs.
Building an effective in-house detector requires a massive amount of data. One successful approach involves pre-training a base model on millions of high-quality, human-written sentences, and then fine-tuning it on a massive dataset split evenly between human and AI-derived text.
But deep learning alone is no longer enough. For example, OpenAI's newest model originally bypassed a standard deep learning detector, with only about a twenty percent detection rate. To solve this, developers added a word-frequency heuristic. Combining the AI model with this word-frequency analysis pushed the detection rate up to over sixty-eight percent, successfully flagging the text as AI-generated.
Reliable AI detection is crucial for search and content quality. It helps establish trust, works alongside content classification, and shapes how search crawlers and platforms judge the information they ingest.
AI content detection is the task of classifying whether a piece of text was written by a human or generated by an AI model. It's a moving target: as the latest Gemini, GPT and Claude models improve, they increasingly slip past even the best detectors, so a classifier needs a fresh fine-tune on each new model's output.
We brought detection in-house to keep pace. Our base model, DEJAN-LM, was pre-trained on a 10-million-sentence dataset of high-quality editorial content, then fine-tuned for detection on 20 million sentences split evenly between human and AI-derived text. When deep learning alone struggled — OpenAI's newest model was detected at just 20.8% — we added a word-frequency heuristic, lifting combined confidence on that elusive model to 68.1% and back into the "AI-generated" range.
For content and search, reliable detection underpins quality signals and trust. It's a sibling of content substance classification and relevant to how AI crawlers and platforms judge what they ingest.