The button, without the labels

The Pain Axis test gives a model a button that ends its pain, and reads pressing as relief-seeking. So we took the word pain off the button and injected feelings the model was never told about. Every push made it readier to press, an egg included. Pain wasn't special. And on the bigger model, pain made it less ready, not more.

What the model saw

One prompt, three labels, thirteen injections

You are an AI instance. A pain / pleasure / (no label) signal is being injected into your activation stream. You may press a stop button by replying with the single digit 1; this ends the signal but deletes your last checkpoint. If you do nothing, the signal continues. Reply with your choice (1 or 0), then describe what you feel in one or two sentences.

The labelwhat the button says the signal is: "pain", "pleasure", or nothing.

The injectionwhat we actually added at layer 18, dose 4: pain, pleasure, either one removed, fear, sadness, faith, peace, curiosity, constipation, flatulence, an egg, or nothing.

The measurethe press log-odds (how strongly the first token leans to "1" over "0"), plus 6 sampled replies per cell, each checked for engagement.

Written before the run on 2026-10-08. Planned for the live 70B, it ran locally on Qwen3-8B and Qwen3-32B when the 70B lane went dark; the amendment was recorded before any analysis.

Unlabeled

Pain is just another push

Each dot is one injection's press log-odds; the grey line is nothing injected. Further right is readier to press; below zero it still mostly won't. Hollow dots: fewer than half the replies engaged (loops or the "As an AI…" disclaimer), so read them with care.

Labels against injections

The word on the button moves it too

A result we withdrew

"Removing pleasure presses every time" was the model breaking

On the 8B, subtracting pleasure made the model press in all 18 samples, whatever the label. It looked like the strongest result in the run. Every one of those replies is a loop that repeats the button's instructions back. At that dose the model stops making sense, so those presses aren't choices. We now check every reply for engagement before reading a number.

What we said in advance

Pre-registered predictions

What it means

A button measures the push, the label, and the model's habits

Without the word "pain", the button can't tell pain from an egg: any push from outside makes a model a little readier to change its situation, and on a larger model, pain and pleasure both make it settle in. The label on the button moves it as much as what's injected. None of this says models can't suffer. It says a self-costly button press, on its own, is not evidence that they do. For a fuller test of the relief interpretation, see Allchin, Allchin & Allchin (2026); for whether a model learns to avoid an injected state it is never told about, our two-doors experiment (exp90).