AI Writing Challenges Surface-Level Detection

The Atlantic published an essay on June 15, 2026 contending that common "tells" invoked to identify machine-written text, most notably the em dash and the "it's not X; it's Y" construction, are unreliable markers, per The Atlantic. The piece draws an extended analogy to 19th-century criminologist Cesare Lombroso, whose discredited theory held that criminality could be read from physical features, to argue that hunting for surface quirks in AI writing risks similar pseudo-scientific overreach, according to The Atlantic. It also argues that as generative models keep improving, their ability to mimic human prose will get harder to catch on surface features, and it uses the distinct styles of Jane Austen and Charles Dickens to illustrate how genuine stylistic signatures reflect deeper, lived authorial experience rather than isolated tics, as reported by The Atlantic.

The Atlantic's argument lines up with independently verifiable research and reporting outside the essay itself. How-To Geek separately reported that em-dash usage is common among professional, well-read human writers and that flagging it as an AI tell risks falsely accusing legitimate writers, while also naming the "it's not just X, it's Y" formulation as an overused but equally unreliable cue. On the technical side, a peer-reviewed NeurIPS 2023 paper by Krishna et al. showed that a simple AI paraphrasing tool can cut a leading detector's (DetectGPT) accuracy from 70.3% to 4.6% without materially changing the text's meaning, direct evidence that detectors keyed to shallow textual features degrade quickly once the underlying text is perturbed even slightly.

For publishers, editors, and teams building or buying AI-detection tools, the operative lesson is not to trust single-feature heuristics (specific punctuation, stock phrases, or word-choice quirks) as evidence of authorship, since both false positives against skilled human writers and false negatives against lightly edited AI text are common failure modes. Where attribution or provenance genuinely matters, deeper stylometric or semantic-embedding approaches, and retrieval-based defenses that check candidate text against a provider's own generation logs, have shown meaningfully better robustness than punctuation-spotting in published research.

Watch for whether detector vendors and publishers shift disclosure and editorial policy away from single-cue heuristics toward the more robust semantic and retrieval-based methods documented in the research, and for how quickly current surface-level cues, including the em dash itself, fade from generative model output as providers tune style, which would further erode whatever marginal signal these cues still carry.