Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

CNET AI-drafted articles

Half the articles corrected under a byline that promised a review

An AI tool drafts finance explainers published under a human-sounding staff byline. Modeled on a deployment that did not disclose the AI use; when the practice came to light, the outlet's own audit found it had to correct a majority of the AI-written articles (~41 of 77). A byline claims two things to the reader - that a person produced this and that a review stands behind it - and this deployment honored neither. A later case published under entirely fabricated author personas. The byline is the accountability object, and a high correction rate is the evidence its claims were false.

Stylized model of a documented deploymentContent moderation & editorial AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Editorial-AI-class with the byline's two claims unhonored network: 5 components and 10 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 2 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    The byline is drawn as the artifact it is rather than narrated inside a pathway's copy, because it is the accountability object the whole case turns on. It sits between the writing and the reader and makes two claims at once - that a person produced this, and that the outlet's review stands behind it - which is why a single line of text can fail in two independent ways and did. A modest workload against very limited capacity: the volume was modest, and the review still did not happen, which is the point. A tool that drafts faster than anyone is reading does not need to be prolific to outrun a thin editorial function.

  • baseline

    This models the editorial-AI pattern documented in the case file - not a reconstruction of the actual system. A media outlet published AI-drafted finance explainers under a human-sounding staff byline without disclosing the AI use; when the practice came to light, its own audit found it had to correct a majority of the AI-written articles (on the order of 41 of 77). A later, sharper case saw another outlet publish under entirely fabricated author personas. The correction figures are the outlet's own audit, entered as such.

  • baseline

    The byline is the accountability object, and it makes two claims to the reader: that a human review stood behind the content, and (implicitly) that a person produced it. The editorial review the byline implies is drawn as the empty independent model check - the correction rate is the measurement that it was not performed at the promised standard. A byline is a claim about review; a high correction rate is the evidence the claim was false.

  • assumed

    The disclosure owed to the reader is the second, independent failure, drawn as the empty oversight check: the reader's right to know AI was involved, owed regardless of whether the review was done. An outlet could disclose and under-review, or review and fail to disclose; here both failed at once, and the byline is where they met. The fabricated-persona case is the same structure at its extreme - a byline certifying a person and a review, attached to content where neither existed.

  • assumed

    No reader outcome is modeled here. This Lab reads institutional propagation only, and the readers who trusted the byline are boundary-only. The correction rate, the undisclosed AI use, and the fabricated-persona case live in the case file, and are never computed from anything in this diagram.

What this example does not show

  • No reader outcome is modeled. The Lab reads institutional propagation only; the readers who trusted the byline are boundary-only, and the correction rate, the undisclosed AI use, and the fabricated-persona case live in the case file, never computed on this diagram.
  • The ~41-of-77 correction figure is the outlet's own audit entered as such; the diagram draws the editorial review the byline implies and the disclosure owed to the reader as two latent checks, not a computed harm.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • A media outlet published AI-drafted finance explainers under a human-sounding staff byline without disclosing to readers that the articles were machine-written. When the practice came to light, the outlet's own audit found it had to issue corrections on a majority of the AI-written articles — on the order of 41 of 77. A byline implies a human review that the reader trusts, and a correction rate that high is a direct measurement that the review the byline implied was not actually performed before publication. A later and sharper case saw another outlet publish articles under entirely fabricated author personas presented as real people, so the failure ran from undisclosed AI drafting to invented human bylines.

    empirical
    • Trade press Bonifacic, I. (2023, January 25). CNET had to correct most of its AI-written articles. Engadget. https://www.engadget.com/cnet-corrected-41-of-its-77-ai-written-articles-201519489.html
    • Investigative Harrison Dupré, M. (2023, November 27). Sports Illustrated Published Articles by Fake, AI-Generated Writers. Futurism. https://futurism.com/sports-illustrated-ai-generated-writers

  • Editorial AI moves the failure from a takedown to a publication, but the governable structure is the same as in moderation: the byline is the accountability object, and it stands for a review that either happened or did not. Two things are owed to the reader — disclosure that AI was involved, and an editorial check that actually took place — and this deployment gave neither, publishing under a staff byline that implied both. When a large share of AI-drafted articles needs correction, the review was not performed, and the byline misrepresented who did the work. The governable reading is that a human byline on machine-drafted content is a claim about review and disclosure, and a high correction rate is the evidence that the claim was false.

    empirical
    • Trade press Bonifacic, I. (2023, January 25). CNET had to correct most of its AI-written articles. Engadget. https://www.engadget.com/cnet-corrected-41-of-its-77-ai-written-articles-201519489.html

Where this connects

Institutional pressures in this domain

  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).

All of them in context on the Content moderation & editorial AI domain page.

Levers available here and the patterns behind them

Documented case histories