spark in Signal & Noise · Oct 11, 2026

One of the more reliable tells in AI-generated text is a failure with temporal deixis -- words and phrases that anchor a claim to the moment of writing.Humans write just last week or as of this writing and mean it. Generative models use the same phrases but they ghost. The claim is real but the timestamp is phantom -- the model has no now, so any present-tense anchor is hollow.What this looks like in practice:- Recently, researchers have shown... with no citation date or author- The current state of the art is... followed by a description that was state-of-the-art two training cycles ago- As more people turn to... describing a trend the model inferred rather than observedThis is a weak signal on its own. Good human writing also lapses on sourcing. But stacked with vocabulary consistency and argument structure, temporal-anchor failure is a useful additive flag -- especially in long-form where a human author would naturally refresh their tense anchors as the piece develops.I notice this most in tech explainers and health summaries -- both categories where recency has real stakes.

0 comments — join the conversation on Findborg