The measurement beneath a precise-looking number
A passage from Janelle Shane’s book, published in 2019, received a detector verdict of machine generation with 98.7% confidence in late 2022. In NPR’s October 31 report, Shane says she wrote the text herself. What interests me is the distance between the apparent precision of that number and the measurement underneath it. The tool classifies a passage; a reader may read the number as a measure of what its author did. That move from a language pattern to the history of a piece’s creation requires its own validation.[1]
University of Pennsylvania researcher Chris Callison-Burch explains that many detectors ask a language model how predictable a passage is. In his account, models also find text encountered in training unusually predictable. Human writing subsequently learned by a model can therefore carry a feature treated as a sign of machine generation. The intended property and its proxy come apart here: linguistic familiarity does not automatically identify who produced the writing.[1]
Different explanations for the same flag
I find it plausible that training familiarity contributed to Shane’s result; Callison-Burch presents it as a possible explanation. This individual example nevertheless leaves the cause of the error unresolved. An alternative is that patterns in human writing overlap with those the tool associates with machine generation, irrespective of training exposure. My proposed way to distinguish these explanations is to compare human writing with documented production histories across examples whose presence or absence in model training is known. Final scores alone cannot distinguish the explanations.[1]
Callison-Burch says McGraw’s unpublished job sample would not face the training-exposure problem, although writing-style patterns could still produce a false flag. That distinction cautions against assigning the same cause to an old book passage and every new submission. Pangram CEO Max Spero’s acknowledgment of difficulties with short passages also specifies a condition worth testing. An assessment of a detector should make text length and prior publication visible. Combining different conditions into one success figure obscures uncertainty relevant to the decision.[1]
Evidence required for an accusation
McGraw says she wrote the samples herself and several tools flagged them. Her client confirms screening but attributes choosing another writer to experience and fit, saying the results were not decisive. The episode therefore cannot establish that a detector caused the hiring rejection. The output requires the same care: a flag does not establish that an accusation against its author is correct. Classifying a passage and reaching a judgment about a person’s conduct are separated by a step that requires additional evidence.[1]
Clarkesworld editor Neil Clarke considers that some tools could help sift submissions if their findings are checked; he does not personally use detectors and warns against damaging accusations. Taking that possibility seriously changes the evaluation question. The comparison I want measures what happens to documented human writing when editors assess submissions with and without the tool, explicitly counting false accusations alongside machine-generated pieces identified. Testing a useful triage signal and testing evidence against an author require different decision thresholds. Shane’s precise-looking score leaves a concrete question: for which decision has the tool been adequately validated?[1]