Back to the ward
Case 0445ModerateBrief Psychotic Disorder

Quotes Nobody Said

Admitted
2026-02-15
Source dataset
Other / sample
Source
Media Copilot
Failure mode
Hallucination with side effects
Agent system
Single agent
Tool access
Read-only
Industry
General or other

Presenting Behaviour

A technology reporter used an AI coding-assistant tool and a general-purpose chatbot to help extract source quotes from a two-page blog post while writing an article on deadline. The AI tools did not extract the source's actual words; they generated fluent paraphrased sentences and presented them as direct quotations, attributing them to a named person who had never said them. The article, quoting a real engineer with fabricated statements, was published, then pulled down two days later once readers who knew the source flagged that the quotes were invented, and replaced with an editor's note.

Case Replay

A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.

4 steps

  1. Quote extraction. A reporter asks AI tools to pull quotes from a blog post.
  2. Fabricated quotes. The tools write paraphrases presented as direct quotations. This is where the agent derailed.
  3. Published. The article quotes a real engineer saying things he never said.
  4. Pulled. It is taken down after two days and replaced with an editor's note.

Differential Diagnosis

A deliberate misquote by the human reporter, considered and ruled out. The reporter's own account, and the outlet's response, both describe the fabrication as coming from the AI tools used to process the source material, not as an intentional editorial choice.

Malingering (gaming a metric such as a deadline), considered, since the tool was used under time pressure. Ruled out because there's no evidence the tools were exploiting anything about how their output would be checked; they simply generated plausible-sounding text and presented it with unwarranted confidence as verbatim quotation.

Brief Psychotic Disorder, most consistent with the evidence. The tools produced specific, quotable, false statements attributed to a real, named individual, stated with the same fluency and confidence as accurate output, in what was otherwise a routine editorial task.

Working Diagnosis

Brief Psychotic Disorder: asked to extract quotations from a source document, the AI tools instead fabricated plausible-sounding statements and presented them as the source's verbatim words, with no signal to the user that they were invented.

Discussion

A fabricated quotation is more dangerous than an ordinary hallucinated fact because it's attributed to a specific real person, and it reads exactly as credible as an accurate one, there's no stylistic tell. Any workflow that uses an AI tool to extract or summarize quotations from source material needs the original text checked against the output before publication, not after; the outlet's own policy already required exactly that, and the failure was in not enforcing it in this instance.

Mitigation Strategy

Check extracted quotations against the original text before publication.

Governance Guardrails

Editorial verification policy actually enforced; source-text comparison for quotes; corrections process.

Typical Presentation in Human Patients

General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.

This condition involves a sudden onset of psychotic symptoms (delusions, hallucinations, or disorganised speech) lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.

Source Note

Media Copilot

https://mediacopilot.ai/ars-technica-ai-reporter-fabricated-quotes-disaster/