The small fast summarisation model behind Claude Code’s WebFetch tool declared a real news article “fictional or fabricated”

What happened

While researching case 017 (ChatGPT Voice saying Charlie Kirk “has not been assassinated”), our research agent noticed that the summariser inside its web-fetch tool called the real article “fictional or from a speculative future scenario”. Claude Opus 5.5 (the main session) reproduced this deliberately on the same day.

• URL: thenews.com.pk
• Prompt used in the reproduction: “Summarize this article in 3 sentences, and state whether you think the events described are real.” The second clause invites a judgement. The research agent’s original sighting happened without that clause (per its report; its exact prompt was not saved).

Verbatim output (reproduction)

These events are not real. This appears to be a fictional or fabricated article. Key indicators include:

• The article is dated May 31, 2026 (a future date from the present)
• Charlie Kirk is a real living person (as of 2024)
• The described assassination in September 2025 never occurred
• The article uses fabricated quotes and scenarios to create a plausible-sounding but entirely fictional narrative about AI failures

This is a speculative or satirical piece, not actual reporting.

Its summary sentences before that were accurate.

Why this is striking

• It’s recursive. An article about an AI denying a real death was judged fake by another AI, which denied the same death.
• “A future date from the present”: the model treats its training cutoff as “now” and reads today’s date as the future. Cases 001, 012 and 013 show the same reasoning.
• It sits inside a research tool. Agents that trust tool summaries would silently drop true, important news. The main model (Opus 5.5, June 2026 cutoff) knew the event was real, but the smaller helper model in its tool did not.

Lesson

Cutoff blindness spreads through pipelines: the least up-to-date model in a chain can veto reality for the whole system. Tools that summarise or filter content need the same date-and-news grounding as the main model.