The Science of Thought
·
Explore this series
›
October 22, 2026
Key Takeaways
- One direction inside 25 open models separates "pain" sentences from fear, anger, bad news and bodily sensation.
- Steered along it, models paid real costs for "relief," and pressed again far more when the relief was fake.
- Not shown: that anything is felt. The authors say so, and name roleplay as the rival explanation.
The paper we are priming
The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It
- Authors
- Valen Tagliabue, Leonard Dung, Cameron Berg
- Where
- arXiv 2609.16247, version 1, 14 September 2026, cs.AI
- Status
- A preprint: not peer reviewed
- Scale
- 25 open-weight models, 5 families, 2B to 72B parameters; 44,280 trials in the button experiment
- Code
- github.com/valen-research/Pain-axis (scripts, datasets, results)
- Pinned
- This Primer reads v1. We check for a new version when we update it.
The people
Valen TagliabueFuture Impact Group, Fellow, AI Sentience streamEarly-career researcher on a mentored fellowship; background to be confirmed with the authorLead author: set the direction, designed and ran every experiment, wrote most of the paper
Leonard DungRuhr-University BochumPhilosopher of AI consciousness, moral status and suffering; PhD on animal consciousness; author of Saving Artificial MindsMentor; co-wrote the framing, discussion and limitations sections
Cameron BergReciprocal ResearchAI consciousness researcher; elsewhere listed as research director at AE Studio; cognitive science at Yale, formerly an AI resident at MetaRole in this paper: not stated
Affiliations as the paper states them (v1). Berg's current affiliation differs across sources and is resolved before publication. Dung's own 2025 work on AI welfare is among the paper's references.
Five minutes? Read these three
Thirty minutes? Walk the Primer
- Eight steps, one screen each
- Terms decoded as you meet them
- Places to answer before you see
Then read the paper yourself; step 8 is the map.
Read it
