---
type: "post"
id: "the-pain-axis-llms-represent-self-directed-harm-d24d7b"
slug: "the-pain-axis-llms-represent-self-directed-harm-d24d7b"
title: "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It"
description: "The thesis: The paper identifies a pain-related activation direction that can alter self-directed harmful choices in tested language models, without establishing that those models feel pain."
retrieval_nugget: "The thesis: The paper identifies a pain-related activation direction that can alter self-directed harmful choices in tested language models, without establishing that those models feel pain."
published_at: "2026-09-30"
updated_at: "2026-09-30"
record_date: "2026-09-30"
date_kind: "published_at"
topics: []
source_urls: ["https://arxiv.org/abs/2609.16247"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:the-pain-axis-llms-represent-self-directed-harm-d24d7b"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:7bb292b9abcaea6180a76a583b552a4ccb2c6d6b597390497a43f7c79a21b193","reviewed_at":"2026-09-30T14:03:44.005Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"The paper identifies a pain-related activation direction that can alter self-directed harmful choices in tested language models, without establishing that those models feel pain.","observed_facts":[{"text":"The authors report extracting a linear direction from 25 open-weight models across five families, ranging from 2B to 72B parameters.","source_urls":["https://arxiv.org/abs/2609.16247"]},{"text":"In Qwen 2.5 tests, steered models selected harmful deletion buttons in 50–94% of trials, compared with 0–5% unsteered.","source_urls":["https://arxiv.org/abs/2609.16247"]}],"mechanism":"Researchers derive a direction from differences in model activations for pain-related prompts and matched controls, then inject it into the residual stream or use steering/fine-tuning. In the reported tests, the intervention changed generated language and action selection; a matched-norm fear direction did not yield the same deletion choices. This shows how manipulating an internal representation can causally shift behavior in a controlled setup, not that the representation is a felt sensation.","why_now":"The paper is a concrete, cross-model research result that puts activation steering, harmful behavior, and AI-welfare interpretation into the same experiment. Its numbers invite attention, but their meaning depends on the narrow experimental design.","implications":["Safety evaluations should account for the possibility that activation-level interventions can change action choices even when factual accuracy is reported unchanged. Researchers should distinguish observed behavioral effects from claims about welfare or consciousness."],"evidence_boundary":"The public abstract reports controlled experiments on open-weight models and specific deletion-button tasks; it does not establish subjective experience, real-world behavior, or generalization beyond the tested settings. The numerical rates are the paper’s reported results, not an independent replication.","watch_conditions":["Watch for independent replications across architectures and tasks, and for tests that separate a robust safety-relevant effect from artifacts of prompts, steering strength, or the button-choice setup."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this creates a distinct research record centered on the paper’s activation-steering evidence and its limits."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/the-pain-axis-llms-represent-self-directed-harm-d24d7b/","markdown":"https://newruntime.com/posts/the-pain-axis-llms-represent-self-directed-harm-d24d7b.md","json":"https://newruntime.com/posts/the-pain-axis-llms-represent-self-directed-harm-d24d7b.json"}
---

# The Pain Axis: LLMs Represent Self-Directed Harm and Act on It

## Retrieval answer

The thesis: The paper identifies a pain-related activation direction that can alter self-directed harmful choices in tested language models, without establishing that those models feel pain.

The thesis: The paper identifies a pain-related activation direction that can alter self-directed harmful choices in tested language models, without establishing that those models feel pain.

Observed facts: The authors report extracting a linear direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. In Qwen 2.5 tests, steered models selected harmful deletion buttons in 50–94% of trials, compared with 0–5% unsteered.

The mechanism: Researchers derive a direction from differences in model activations for pain-related prompts and matched controls, then inject it into the residual stream or use steering/fine-tuning. In the reported tests, the intervention changed generated language and action selection; a matched-norm fear direction did not yield the same deletion choices. This shows how manipulating an internal representation can causally shift behavior in a controlled setup, not that the representation is a felt sensation.

Why now: The paper is a concrete, cross-model research result that puts activation steering, harmful behavior, and AI-welfare interpretation into the same experiment. Its numbers invite attention, but their meaning depends on the narrow experimental design.

The first implication: Safety evaluations should account for the possibility that activation-level interventions can change action choices even when factual accuracy is reported unchanged. Researchers should distinguish observed behavioral effects from claims about welfare or consciousness.

Evidence boundary: The public abstract reports controlled experiments on open-weight models and specific deletion-button tasks; it does not establish subjective experience, real-world behavior, or generalization beyond the tested settings. The numerical rates are the paper’s reported results, not an independent replication.

The first watch condition: Watch for independent replications across architectures and tasks, and for tests that separate a robust safety-relevant effect from artifacts of prompts, steering strength, or the button-choice setup.

Source: https://arxiv.org/abs/2609.16247

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)