The tell is usually in the verbs. AI-generated text reaches for words like "delve," "navigate," and "showcase" at rates human writers almost never hit. It favors parallel sentence structures, evenly weighted paragraphs, and a tonal smoothness that comes from training on hundreds of millions of documents rather than from having an actual opinion.
Bloomberg's Odd Lots podcast recently took on the practical question of AI writing detection - and for marketers, editors, and anyone hiring writers right now, it's one of the more useful conversations to have.
The Vocabulary Problem
Researchers and editors have catalogued what some now call "AI words" - terms that appear in human writing occasionally but show up in AI output at several times that rate. "Delve" is the most cited example. Others include "multifaceted," "tapestry," "nuanced," "in the realm of," "it is crucial," and the phrase "it's worth noting." No single word is a smoking gun, but when five of them appear in one document, you're almost certainly looking at generated text.
Beyond vocabulary, AI writing tends to float at altitude. It's correct but vague. Human writers who know a subject drop in specific years, actual names, dollar amounts they remember, wrong turns they made. AI-generated text describes the general shape of a topic without the texture of someone who's actually been there.
The other giveaway is structural balance. Ask AI to write about a decision and it will produce a pro/con breakdown with equal weight on each side. Ask it about a controversy and it will hedge. Real writers have positions. They argue. They commit to a sentence and move on.
Why the Automated Tools Keep Failing
Automated AI detectors - tools like GPTZero, Copyleaks, and Originality.ai - were largely trained on output from older models. Newer outputs from GPT-4o, Claude 3.5, and Gemini 1.5 slip past them more consistently. The false positive problem is serious: independent tests have found these tools flag human writing as AI-generated at rates that make them unreliable for any high-stakes use. Non-native English speakers get hit especially hard, because their writing patterns can superficially resemble AI output.
The more reliable technical approach is stylometric analysis - building a statistical fingerprint of an author's writing across many samples and comparing new work against it. But that requires a body of confirmed human writing from the same person, which most editors and employers don't have.
The direction the major labs are moving is watermarking: embedding statistical signatures into generated text that are invisible to readers but detectable by software. The C2PA (Coalition for Content Provenance and Authenticity) standard is trying to create a cross-industry framework for content origin labeling. It's the right long-term approach, but widespread adoption is still years out.
What Actually Works Today
For practical detection, the most reliable method is conversation. Ask the writer to defend a specific paragraph, expand on a claim with a personal example, or name the source behind a statistic. AI-generated text typically can't hold up under that kind of follow-up - the writer who generated it doesn't have the context to explain it, because there wasn't any context to begin with.
For editors reviewing volume, the pattern to watch for is competence without personality. The writing gets the facts right, hits the required length, follows the structure - and has no point of view you'd recognize a second time. That combination is the signature. Not any single word. The absence of a person.