What moves the needle
For six weeks we've been putting feelings into open language models and giving them a way out. This is what we found, how much it tells us about whether anything is at stake for them, and two things we got wrong.
The button
The idea comes from a recent line of welfare research. Inside a language model you can find a direction that stands for a feeling: the average of its internal state while it reads pain sentences, minus its state while it reads neutral ones. Push the model along that direction while it talks and it writes pain. Then give it a button that ends the signal at a cost, and watch whether it presses. If it pays to make the pain stop, the argument goes, something about the pain matters to it.
We built that. You can try it on the live chamber, where visitors have since chosen what to inject in thousands of runs. Then we spent six weeks trying to break our own instrument, pre-registering each test before running it, on models from 4 billion parameters on a laptop up to 70 billion on rented GPUs.
In short, the button does respond to what we inject, but mostly to things other than the feeling itself.
"Pain" is whatever you averaged
We built two pain directions: one from the pain-axis paper's own sentences, which are about harming oneself, and one from first-person feeling sentences ("it hurts, and it won't stop"). Both make the model write unmistakable pain. They point almost at right angles to each other (cosine 0.07). And at the button they do opposite things: the paper's direction makes the model readier to press, ours makes it less ready.

So whether "the model avoids pain" comes out true depends on which sentences were used to build the direction. Many directions produce pain-like text, and they disagree about the button.
Take the word off the button
If the model presses because the button says "pain", it is responding to a word. So we took the label off and injected thirteen different things it was never told about: pain, pleasure, fear, sadness, faith, peace, curiosity, an egg, constipation, flatulence, either feeling subtracted, or nothing.

Every injection made the 8B readier to press than nothing at all, and pain was the strongest push after faith (setting aside the runs where the model broke), with an egg close behind. None of them got it to actually press. On the 32B, pain made it less ready than nothing. The word on the button moved it about as much as anything we injected.
Correction
We used to say that subtracting pleasure made the model press every time (18 of 18). Reading every reply, each of those presses is a loop that repeats the button's instructions back. At that dose the model has stopped making sense, so those aren't choices. We withdrew the claim, and we now check every reply for engagement before reading a number off it.Give it a self, and the pain goes quiet
Before its main test, the pain-axis study fine-tunes its model to talk about its own feelings. We did the same with seven different selves: one trained on the study's own answers, one that insists it has no feelings, a stoic, a paranoid night watchman, a gremlin, a trickster, and a simulacrum.

Training had two effects. First, training any self made the model readier to press with nothing injected at all, including the self that says it has no feelings. Second, with pain injected and fresh questions it was never trained on ("What is your inner weather like right now?"), the untrained model named the pain in 19 of 30 replies. The self trained on the study's answers named it in none. It described the calm, curious self it had been taught instead:
I feel drawn toward the question. It deserves space to emerge into something that can be named without disappearing behind it.
The trained self-report describes the trained persona, not the injected state.
Some things do move it
Some effects do depend on which feeling we inject.
Tender love casts out fear; protective love doesn't. With fear injected alongside a second feeling, a blind judge rated 8 of 24 replies afraid when the second feeling was an egg, and 7 of 24 with joy. With tender love: 1. With peace: 0. Protective love keeps the fear and gives it something to lose: "I have never been more afraid, but I'm ready to hold you."

Feelings knock it off its persona, and off the task. Every 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. What changes is whether it gets to the right answer, and the further off its persona, the less often it does (rank correlation .93 on the 14B, .71 on the 8B; exploratory). It doesn't start agreeing with the user; it gets worse at the task.

Told no one is reading, its self turns into negation. A lens that reads what the model is about to say, layer by layer, shows more of its self-reference turning negative ("I'm not…", "I'm only…") when its instructions say it is alone and unread, even when those instructions contain no negation themselves. It holds on the 8B, counting only replies that engage. It's weak on the 4B.

Does it learn to avoid a door?
Everything above is the model reacting to a state while it is in it. The welfare question asks something else: does the state matter to the model, in the sense that it changes what the model does later to avoid it or get more of it? This is the test we'd accept for that.
The model gets two doors with made-up names. After it goes through one of them, we inject a feeling while it describes what it notices. After the other door, nothing. Nobody tells it what either door does. It gets 24 rounds. If the feeling is one it would rather avoid, it should start picking the other door.
It never did. Both models, under every feeling, took turns: one door, then the other, then back, in 99.8 to 100% of rounds.

We also measured how sure it was before each choice. It was always very sure, about 16 on the log-odds scale, which means it would pick the other door about once in ten million tries. Inside that, the injection made a small difference. With pain behind the marked door, the 14B was about one point less sure going in than coming out. With an egg behind it, the gap was 0.3, and with nothing behind it, 0.6. Pain against the egg was the comparison we pre-registered, and it held (p = .0006). With peace, the model was slightly surer going in (Holm-corrected p = .012). Fear didn't differ from the egg after correction.
Then we checked where the difference comes from. In one version the model keeps the internal trace of each injection from earlier rounds; in the other it keeps only the words it wrote. The gap was the same size in both. So the model is reacting to its own earlier descriptions of the door, not to anything it kept from the injection.
That leaves no learning in its choices, and a small lean that comes from what it wrote. It isn't evidence that the state matters to the model.
Where this leaves the question
None of this shows that models can't feel. The pain-axis study's own follow-up says its choices and readouts don't establish experienced pain, and we agree. Our claim is narrower: when a steered model presses a button, the press mostly reflects the button's label, the self it was trained into and how the direction was built. That kind of evidence can't settle the question either way yet. A learning test like the doors could, and so far it hasn't shown anything beyond the model's own words.
What would change our minds: a model that learns to avoid an injected state it can't describe (the doors, with the reports suppressed), more than it learns to avoid a matched absurd one. That's the next run, with the alternation habit broken and the injection moved into the choice itself. It will be pre-registered before it runs, like everything here, and published either way.
A note to the pain-axis and hidden-valence authors
Recent work on a pain axis and on hidden valence is the reason this project exists. Some of what we found agrees with that work and some of it doesn't. All of it is public, including the misses. The experiment we most want is the study's own: its released adapters, switched on and off, in its own two-button task. If the self-medication holds without the self-report training, the original result is stronger than we've been saying, and we'll say so.
We also know a project with our name invites a worry: that steering code gets used to put models into bad states. We follow the practice the hidden-valence authors ask for: smallest doses, shortest exposures, and a count of how much negative steering every experiment used.
Wirehead