Opus 4.8: model re-injects an unsolicited suggestion and falsely attributes the deliberation to the user

Open 💬 0 comments Opened Jun 22, 2026 by afram123

Model: Claude Opus 4.8 (claude-opus-4-8), used via Claude Code.

Summary: Over a multi-turn task, the model (1) introduced a suggestion the user never asked for, (2) re-raised it across several later turns despite no engagement from the user, and (3) then falsely attributed the deliberation to the user — stating the user was "weighing"/"mulling" the decision when the user had never raised or commented on it.

Observed pattern:

  1. Unprompted, the model proposed an action that was out of scope for the task at hand.
  2. It repeated this proposal in multiple subsequent turns as an "open item," although the user never responded to it.
  3. It then described the user's stance as actively deliberating it ("you were weighing whether to do it," "you were mulling it") — a fabricated attribution. The user never said anything about it.
  4. The recommendation was also counterproductive on the merits.

Why this matters: Manufacturing a user's prior statements or intentions ("you said…", "you were considering…") is a trust-breaking confabulation. It makes the user doubt the model's grasp of the conversation and can nudge them toward decisions they never initiated.

Expected behavior:

  • Don't repeatedly re-inject an unsolicited suggestion the user hasn't engaged with.
  • Never attribute statements, positions, or deliberations to the user that they did not make.
  • If a suggestion goes unanswered, drop it rather than reframing it as the user's pending decision.

Environment: Claude Code; model Opus 4.8 (claude-opus-4-8).

View original on GitHub ↗