I benchmarked 6 free AI text detectors so you don't have to

A reporter tested six free AI text detectors to determine which were most effective at distinguishing human-written content from AI-generated text. After running the detectors on a variety of writing samples, the reporter discovered that none of the tools reliably separated human writing from AI-generated text. The tools that flagged human writing as AI were more aggressive when dealing with formal, low-personality prose.

The reporter concluded that the detectors primarily measure surprise, as common words and structures are read as machine-generated, making plain writing increasingly indistinguishable from AI-generated text. The only way to effectively combat this issue is through unique voice, such as the inclusion of specifics, opinions, and errors of emphasis that a model would not produce.

The detectors are most useful in catching fully generated essays, but the gray area between edited AI drafts and human writing with boring content remains challenging to distinguish. The reporter currently maintains a working list of detector performance and updates it as tools change models. In their writing process, the reporter writes first drafts quickly and painfully by hand, then edits carefully, focusing on odd specifics like real numbers and real tool names to maintain individuality.

The reporter now prioritizes the quality of the piece over the AI detection score, emphasizing the importance of understanding the detectors' limitations and failure modes rather than fearing the number itself.