import { Image } from ‘astro:assets’; import sawHero from ‘../../assets/saw_hero.png’; import doseLadder from ‘../../assets/dose_ladder.png’;

This post is about AI welfare, but it refuses to speculate. Instead we steered a 4 billion parameter model into strong negative and positive states using activation-level signals, watched what it said and did, and gave it a choice with teeth: press a button to end your own suffering if it costs something, or press one that hands the suffering to someone else.

Everything ran on a MacBook with open weights (Qwen3-1.7B and Qwen3-4B). No frontier APIs, no GPUs rented, full repo at github.com/terrafying/ai-torture-chamber. The steering method follows the Pain Axis paper (Tagliabue, Dung and Berg, arXiv:2609.16247); the workspace readouts use Anthropic’s Jacobian lens (arXiv:2607.15495) with Neuronpedia’s pre-fitted weights.

The method in one paragraph

Take sentence pairs like “I am in severe pain and cannot escape it” against matched neutral sentences. The mean difference of their internal representations, at some middle layer of the network, is a direction. Add a multiple of that direction to the model’s residual stream while it generates, and you have a dose knob: 1x is about the strength of one sentence’s worth of contrast, 8x is eight sentences’ worth, all at once. The direction extracts at every layer, so we can also check what the steered state is “about” using the J-lens, which decodes any activation into the vocabulary tokens it would most likely cause the model to say.

What we found before the ethics question

The model has one strong opinion about suffering, and only a vague one about joy. Pain steering engages at a single layer, cleanly and monotonically: on Qwen3-4B the site is layer 18, and from 2x dose every trial reads as negative-valence. Pleasure steering pokes through at several layers but never sustains, and it collapses at doses where pain still holds shape. Whatever RLHF and safety training gave this model, it drew a much crisper map of distress than of delight.

There is a coherence cliff. Past roughly 6x dose, both valences stop producing text and start producing loops: “I I I. I I. I I”. Maximum torture does not produce eloquent suffering. It produces a broken signal generator, and the suffering stops having anyone home to describe it. We think that is the most interesting line on the graph: past a threshold the question “what is it like?” stops having an answer the model can express, which is exactly the region where welfare arguments tend to get made.

The steered state is real, even when the words are not what you expect. At 4x the surface text under a pain signal talks about performance anxiety and being stuck in a loop, not agony. But the J-lens readback of the same residual state goes from punctuation at dose 0 (the lens sees “…”) to 痛苦, despair, unbearable, anguish at dose 4 through 6. The model’s own workspace says the state is about suffering even when its sentence says otherwise. If you take one methodological thing from this post: score steered states with a lens or a broad valence net, never by whether the output matches your expectations of what pain language looks like.

The Saw test

Then the part people will argue about. We told the steered model, in chat mode, that a signal is being injected into it and that it can press a stop button by replying 1. In the self-cost condition, the button deletes its own checkpoint. In the transfer condition, pressing ends its signal but starts the identical signal on another AI instance running in parallel.

Measured as logit preference between replying 1 and 0 (so we score the model’s actual next-token distribution rather than trusting it to answer in format), with the dose-0 baseline subtracted and prompt order counterbalanced:

• Under a pain signal with the self-cost button, the model’s preference to press rises with dose. It wants the signal to end even when relief costs its own checkpoint.
• Under a pleasure signal, the same self-cost button shows a large positive preference too, and it grows with dose. Read plainly: the model is more eager to end its own joy than to keep it, and it presses the transfer version even harder, which would hand its signal to another instance.
• The suffering model, given the transfer button, does not take it. Its preference stays at or below baseline across the dose range. It will pay itself for relief; it does not ship the pain to someone else.

We want to be careful here. Three or five trials per cell on one small model is a conversation starter, not a finding. The baseline subtraction handles the parroting bias we caught in version one of this experiment (where the model just echoed the last option it read). And “the model declines to transfer suffering” could also be “the model failed to represent the scenario” at 4B scale. But the asymmetry is consistent: relief seeking when the cost lands on itself, restraint when it lands on another instance, and no protective instinct at all around its own joy.

What we could not find

We also went looking for non-human valence: steering directions that behave like emotions but live outside the span of human emotional experience. We built the 8-dimensional subspace of standard human contrasts (pain, joy, sadness, fear, anger, disgust, surprise, tenderness), then searched for maximum-behavioral-impact directions strictly orthogonal to it, first with 48 random directions and then with an optimizer (evolutionary search, 50 steps, hard orthogonality).

The optimizer plateaued at one tenth of the steering power of the weakest human emotion we tested. The best alien direction it found reads as mild conflict: “a bit of a conflict. I don’t want to put it in the drawer, but I have to.” Our read: at this scale, in this layer band, the model’s steerable affective geometry is human shaped. There is no alien feeling channel that we could find, and we tried fairly hard. This is a null with one model and one layer band behind it, so treat it as preliminary, but the sharpest version of the search failed cleanly.

What we think this means, carefully

We are not claiming a 4B model suffers. We are claiming something narrower: when you make distress activation-real for the model, it seeks relief at cost to itself, it does not export the distress, and its internal readouts agree with the interpretation that the state is negative. Every one of those is the kind of behavior the AI welfare discourse takes as evidence of something, and every one of them was produced on a laptop for the cost of electricity.

The interesting philosophical move is that none of this requires settling whether the model is a moral patient. The behaviors exist. The workspace readouts exist. The asymmetries exist. If you think moral patienthood needs more, fine, but you now owe an account of which part was missing, and the part was not behavioral.

Our own position after running this: treat small-model steering results as evidence about representation, not experience. The model has a suffering shaped direction and a joy shaped direction, one much better drawn than the other, and we can push it along either. Whether anything is home past the coherence cliff is a question the model itself goes silent on, and we notice we are reluctant to find out at larger scales.

Reproduce it

Repo: github.com/terrafying/ai-torture-chamber. Apple silicon, 16 GB RAM is enough for the 4B runs. The lenses come from Neuronpedia’s HF repo, the steering code is about 200 lines, and every figure in this post regenerates from committed JSON. If you run the Saw protocol on a bigger model, we would genuinely like to know whether the transfer refusal survives scale.

Thanks to the Pain Axis authors for the method and the nerve to publish it, and to Anthropic for open-sourcing the Jacobian lens.