Review an AI-origin claim before labeling a person’s writing

Define the origin claim and observed source

“This post contains a dash” describes visible text. “The author used a particular model to create the whole post” asserts a production history. Those statements need different evidence. Mixed work—human drafting, editing, translation or model assistance—also need not fit a simple wholly-human/wholly-generated label.

Keep actual creation information, a source disclosure and your own inference separate. Do not ask for unnecessary private prompts or account logs, and do not fill their absence with certainty. A named accusation about a real person needs its actual support; the cases here deliberately use no real writer or private text.

Read the date, language and population behind a detector result

OpenAI’s official 2023 classifier page states that the classifier was withdrawn on 20 July 2023 for low accuracy. Its historical English challenge-set report described 26% true positives and 9% false positives, with limitations for short text, other languages, editing and out-of-distribution inputs. That page is not an available current product recommendation.

Do not transfer those historical rates to another detector, a German social post or all 2026 writing. For a real current tool, inspect what its score means and what relevant validation actually exists. A confident display is not automatically a calibrated probability about a particular author, and an old “try it” link does not override a retirement notice.

Work the complete constructed classification table

Assume an invented batch of 10,000 texts with known-for-the-exercise labels: 1,000 generated and 9,000 human. An imaginary classifier flags 80% of the generated texts and 10% of the human texts. It therefore flags 800 generated plus 900 human texts, missing 200 generated and leaving 8,100 human texts unflagged.

Among the 1,700 flags, the generated share is 800/1,700 ≈ 47.06%, although sensitivity was 80% and overall classification accuracy is 89%. These denominators answer different questions. The 10% prevalence and classifier properties are invented assumptions, not observations of LinkedIn, a current provider or the author of any real post.

Resolve three constructed origin claims

D1 — Constructed text includes a dash and a common opening phrase. Decision: record those observed features, but leave its creation history unknown without actual evidence. Outcome: no assertion that punctuation proves generation, a particular model or deceptive conduct.

D2 — Use the complete 10,000-text fixture: 800 true flags and 900 false flags. Decision: report 800/1,700 among flags under the declared assumptions, not “every flagged text is 80% likely generated.” Outcome: a conditional arithmetic lesson, not a measured prevalence or an individual accusation.

D3 — Constructed recommendation cites the withdrawn 2023 classifier as a current German-post validator. Decision: correct availability, date and language scope before using the recommendation. Outcome: do not present its historical English rates as performance of an available current service.

Keep uncertainty and the proposed publication separate

A reasonable outcome can be “the origin is not established; the statement still needs an ordinary factual/source review.” Truth of a claim and identity of the production process are separate. Likewise, a genuine disclosure of one assisted step does not establish that every word, image or event was generated.

NIST’s publication abstract distinguishes provenance, labeling and detection among technical approaches to synthetic-content transparency. We read that overview, not a full benchmark report for a current detector. Content credentials can supply bounded asset-history assertions in supported media; absence of a credential does not prove human authorship or generation.

Keep observations, tool scope and the bounded conclusion

The worksheet records the proposed origin claim, exact permitted text/reference, observed feature, actual disclosed creation information, detector/version/output definition if any, validation scope, denominators and a supported or unresolved conclusion. The offline packet reproduces only the constructed confusion-matrix arithmetic.

The official historical classifier body, NIST publication overview and relevant provenance explanation were read on 1 November 2026. No detector was queried, no private text uploaded and no real person classified. The invented rates are not the historical classifier’s rates and are not another original research study.