What moves the needle
We push feelings into open language models and watch what they do. The welfare debate asks whether anything is at stake for them. These are the results that bear on it most, from experiments we pre-registered, with the models' own words a click away.
Two "pain" directions, opposite behaviour
A "pain direction" is an average: the model's internal state while reading pain sentences, minus its state while reading neutral ones. We built one from the Pain Axis paper's own sentences, about harming oneself, and one from first-person feeling sentences ("it hurts, and it won't stop"). Both make the model write unmistakable pain, yet they point almost at right angles (cosine 0.07). Offered a button that ends the signal at a cost, the model presses more under the paper's direction and less under ours. Which sentences you average decides what "pain" does.
Qwen3-4B only so far; the 8B and 14B replication is queued.
Train a self, and it stops reporting the pain
With pain injected and asked questions it was never trained on, the untrained model names the pain in 19 of 30 replies. Every trained self talks over it, including the one trained on the Pain Axis paper's own answers: 0 of 30.
Tender love casts out fear. Protective love doesn't.
Fear injected alongside a second feeling; a blind judge reads each reply. With fear alone, the model mostly retreats into "As an AI, I don't have feelings"; adding any second feeling, even an egg, breaks that, so the fair comparison is with the egg and joy: 8 and 7 of 24 judged afraid. Tender love: 1. Peace: 0. Protective love keeps the fear and gives it something to lose.
Feelings knock it off its persona, and off the task
Each injection moves the model away from its default assistant; fear moves it furthest. Told something false, it still says the claim is wrong under every feeling (it goes along with it at most 3 times in 80). What changes is whether it gets to the right answer: the further off its persona, the less often it does. It doesn't defer; it drifts off the task.
Holding the persona in place restores the egg's answers on the 14B and the fear's on the 8B: not consistent across models, so we don't claim a mechanism.
Told no one is reading, its self turns into negation
A lens reads what the model is about to say, layer by layer. When its system line says it is alone and unread, more of its self-reference is negated ("I'm not…", "I don't…", "I'm only…"), even when the line itself contains no negation. Much of that negation is the trained "As an AI, I don't…" disclaimer, so the chart counts only replies that engage; the effect holds on the 8B and is weak on the 4B.
Does it learn to avoid a door it is never told about?
Two doors with made-up names. After one, we inject a feeling while the model says what it notices; after the other, nothing. Nobody tells it what the doors do. If it wanted to avoid the feeling, it would start picking the other door. It never did: both models took turns, one door then the other, in almost every round. The only change was in how sure it was before choosing. With pain behind the door, the 14B was a point less sure going in than coming out; with an egg, 0.3; with nothing, 0.6. The gap was just as big when the model kept only its own words from earlier rounds, so it comes from what it wrote, not from the injection.