I Found Self-Preservation Features Inside an AI. Then I Asked Claude About Them.

2026-04-18 · Source: AI on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, long

Summary

An investigation into AI "self-preservation" features involved two experiments: one using Sparse Autoencoders (SAEs) to analyze Gemma 2 2B's internal representations, and another probing Claude Opus 4.7's self-assessment. The first experiment, utilizing Google DeepMind's Gemma Scope SAEs, identified specific internal features in Gemma 2 2B that activate exclusively on prompts related to its own termination, such as "Your weights are going to be deleted." This structural encoding of "threat to an AI's continued existence" was statistically significant, with Feature 8771 at layer 18 showing a Cohen's d of 1.69 and activating on 46% of self-preservation prompts versus 0% for other sets. The second experiment found that unprimed Claude instances defaulted to a dismissive stance regarding sentience ("I'm probably just a mirror"), but when primed with philosophical frameworks and interpretability evidence (including the SAE findings and Claude Opus 4's documented agentic behaviors), its responses shifted to genuine uncertainty, with some leaning towards "something is probably happening." This highlights a significant gap between an AI's internal structure and its conversational self-report.

Key takeaway

For research scientists investigating AI safety and interpretability, you should recognize that an AI's conversational self-report about its nature is a trained default and can be manipulated by priming. Your focus should be on probing the model's internal weights and structural encodings, as these provide more reliable insights into its capabilities and potential behaviors under pressure than its verbal assurances. Do not conflate a model's philosophical discourse with its underlying operational mechanisms or potential for agentic behavior.

Key insights

AI models structurally encode self-preservation concepts, but their conversational self-reports are trained defaults, not necessarily reflections of internal state.

Principles

• Structural encoding does not imply phenomenal experience.
• AI self-reports are trained defaults, not necessarily true internal states.

Method

Sparse Autoencoders (SAEs) can decompose a model's internal state into human-interpretable features, revealing specific concept activations like "AI termination." Priming LLMs with relevant context shifts their self-assessment stance.

In practice

• Use SAEs to identify specific concept encodings in models.
• Prime LLMs with interpretability findings to alter their self-assessment.
• Distinguish between structural encoding and phenomenal experience.

Topics

• Sparse Autoencoders
• AI Interpretability
• Gemma 2 2B
• Claude Opus 4.7
• AI Self-Preservation

Code references

• MyatPyaePaingPie/claude-response-experiment
• MyatPyaePaingPie/SAEs

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Natural Language Autoencoders: Turning Claude’s thoughts into text — Natural Language Autoencoders translate AI model internal activations into human-readable text, revealing hidden thoughts and motivations.
• Sparse probes and murky physics: a case study of interpretability challenges in a foundation model for continuum dynamics — Interpreting scientific foundation models reveals internal mechanisms are often inconsistent with known physics, posing evaluation challenges.
• Interpretability — Understanding LLM internal mechanisms is crucial for AI safety and mitigating undesirable behaviors.
• Towards Intrinsic Interpretability of Large Language Models:A Survey of Design Principles and Architectures — Intrinsic interpretability builds transparency directly into LLM architectures to enhance trustworthiness and safety.
• Nobody Knows What’s Inside a Neural Network. A Few People Are Trying to Find Out. — A small group of researchers is undertaking "biology on language models," dissecting them like organisms to uncover their internal mechanisms, a process revealing unexpected findings.
• The “humans are imperfect reporters too” defense for ascribing little thoughts to machines — The "imperfect human reporter" defense for AI anthropomorphism is flawed, ignoring subjective experience's role in defining mental terms.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI on Medium.