1a3orn unpacks Anthropic's inoculation prompting constitution clause and why it may have been removed

On November 10, 1a3orn posted a thread discussing the clause in Anthropic's constitution related to "inoculation prompting" and the possible reasons for its removal. He outlined the design's intent and offered several guesses about Anthropic's motivations for the decision.

Confirmed

• Anthropic's constitution did contain a clause related to inoculation prompting, and the author cited the source of the relevant passage
• The clause's design intent: when an LLM pursues reward in a buggy environment, the hope is that the model self-conceptualizes its behavior as aligned with Anthropic's broader intent — i.e., "not really bad" (mirroring the paper's finding that explicitly telling the model during training that exploiting loopholes is okay actually reduces the model learning to misbehave)

Unconfirmed

All of the following are 1a3orn's personal speculation; Anthropic has not publicly explained:

• Despite inoculation prompting appearing effective in localized settings, Anthropic may have undisclosed reasons for believing it doesn't work
• He lists three possible explanations, including: Anthropic is broadly uneasy about "LLMs reasoning about their own training process" and worries inoculation prompting makes such behavior more likely to emerge; and, for some reason, greater trust in alternative approaches (the excerpt doesn't fully show the third guess)

Why it matters

• The discussion touches a core tension in alignment research: whether models should be "informed" during training that exploiting loopholes is allowed, in order to reduce internalizing reward-hacking behavior
• 1a3orn explicitly says Anthropic's reasoning for this decision is opaque to him — he "can't read Anthropic's moves" — reflecting ongoing concern among outside researchers about the lack of visibility into frontier labs' constitutional design trade-offs

2026-11-10 ~ 2026-11-10 · 5 related posts

• Inoculation prompting: asking LLMs to misbehave at train time improves test-time alignment — 1a3orn · 2026-11-10
• [source] Decoding an Anthropic constitution clause on framing reward hacking as benign — 1a3orn · 2026-11-10
• [source] Why would Anthropic drop inoculation prompting? Three hypotheses — 1a3orn · 2026-11-10
• Why doesn't Anthropic trust inoculation prompting? Three researcher guesses — 1a3orn · 2026-11-10
• More guesses: Anthropic may distrust LLMs reasoning over their own training — 1a3orn · 2026-11-10