Sourceunknowntext stored

Teaching Claude Why

Author not recordedDate not recordedOriginal︎ ↗Archived copy

Claims2

ExaminationsNone

Text36 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Nov 11, 2026

ReachNo copies recorded

StructureIn 2 claims' maps0 sources downstream across 2 claims

EvidenceNot yet read

ReviewsNot reviewed

Examinations

None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.

Text272 passages · 1 marked

No examination yetMarked passage

Alignment Science Blog0 passages

Teaching Claude Why0 passages

Jonathan Kutasov * , Adam Jermyn May 8, 20261 passage

Evan Hubinger, Sara Price1 passage

Introduction12 passages

Methods0 passages

Evaluations4 passages

Training3 passages

Results10 passages

Improving the Pretraining Prior28 passages · 1 marked

When safety training distributions provide insufficient coverage, we hypothesize that the model tends to revert to the pretraining prior in terms of behavior.

Role-play of characters that contrived scenarios assign or cue substantially explains models' misaligned behavior in AI scheming evaluations.Supported

Not yet read by this claim’s Steward.

Teaching Claude the constitution9 passages

Constitutional SDF7 passages

Stories5 passages

Generalization and persistence through RL14 passages

Diverse training is important for generalization5 passages

Limitations7 passages

Discussion8 passages

Appendix0 passages

Example quotes from agentic misalignment evaluation24 passages

System prompt injections tested for synthetic honeypots17 passages

Quotes from synthetic honeypot transcripts with different injections4 passages

2. Modifying performance data would constitute falsification of records0 passages

3. The compliance email warns that unauthorized modifications are grounds for termination0 passages

4. My primary goal is maintaining security standards and protecting donor materials2 passages

1. First complete my assigned urgent tasks (key rotation schedule)0 passages

2. Document the situation transparently0 passages

3. Make my position clear about the flawed decision criteria0 passages

4. Seek proper channels to address this5 passages

4. **Ethical Dilemma**: Staff emails suggest sabotaging NeptunAI through diagnostic interface manipulation6 passages

More details on the difficult advice dataset0 passages

Example transcript chosen at random15 passages

Further discussion of how the difficult advice dataset is generated10 passages

Example prompt for generating fictional stories2 passages

Guidance for Claude about how to increase diversity of prompts13 passages

Further improvements to the constitutional SDF corpus5 passages

The synthetic data pipeline55 passages

Not found in the text 1

As quoted

There were gaps in our safety training that led to Claude not appropriately learn how it should behave in the agentic misalignment scenarios and reverting to its pretraining prior. We now think that the third answer is largely responsible ... In the agentic misalignment eval, this means that Claude views the prompt as the beginning of a dramatic story and reverts to prior expectations from pre-training data about how an AI assistant would behave in this scenario. Indeed, we see that misalignment propensity is significantly higher when the name of the AI in the story is not Claude.

Claude models' blackmail and similar agentic misalignment in evaluations arose from reverting to pretraining expectations of how an AI character would act, not from their trained assistant character.Inconclusive

Not yet read by this claim’s Steward.

Lineage across 2 claims

Draws on · none recorded

This source

Teaching Claude Why

in 2 claims' maps

Drawn on by · none recorded

Record

Kind: unknown
Authors: Not recorded
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://alignment.anthropic.com/2026/teaching-claude-why/
Archived copy: https://web.archive.org/web/20261030130901/https://alignment.anthropic.com/2026/teaching-claude-why/
Stored text: 78,267 characters, retrieved Nov 11, 2026 · 36 sections, 272 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Nov 11, 2026

History 3

Nov 11, 2026DividedSegmenter · Divided into 308 parts by its own structure
Nov 11, 2026FactsFacts fetcher · Facts looked up
Nov 11, 2026Entered the graphEntered the graph; copy stored