Conscious experience
The study measures activations and behaviour, not whether there is something it feels like to be the model. The authors explicitly leave consciousness unresolved.
01 / IF YOU ONLY READ ONE THING
A real behavioural finding.
An open question about experience.
Researchers found an internal pattern associated with descriptions of pain. Strengthening that pattern changed what models said. In a separate experiment, some specially fine-tuned models selected a button that removed it—even when the button came with a described cost.
The most interesting part: models generally pressed again less often when the button actually removed the intervention than when it merely claimed to.
That is evidence for a pain-like functional pattern. It is not a demonstration that the models consciously felt pain.
OBSERVED
Internal measurements, changes in generated text, and button choices.
INTERPRETED
A candidate state with some of the functional features associated with pain.
NOT ESTABLISHED
Conscious suffering, human-like feelings, or a result about every AI system.
02 / FIRST, A DEFINITION
“Pain” is doing a lot of work in that title.
The authors use a broad, functional definition: a state associated with aversion, avoiding or ending the state, and disruption of ordinary behaviour. They include physical pain, grief, humiliation, moral injury, and sustained confusion or failure.
They are not starting with a test that detects conscious experience. Their question is whether a pain-related representation does some of the things pain is expected to do.
REPRESENTING PAIN
≠A model has internal patterns that help it process or generate pain-related content.
EXPERIENCING PAIN
There is something it feels like for the model to be in that state.
The paper tries to move beyond mere discussion of pain by testing behaviour. Whether those tests bridge all the way to subjective experience remains unresolved.
Source: §1 and footnote 1; §6.
03 / THE METHOD, WITHOUT THE EQUATIONS
An axis is a direction.
Not a tiny pain centre.
A language model carries information as large lists of numbers. Those lists change as text passes through its layers. A direction is a particular pattern of changes across those numbers—not necessarily one neuron or one neatly labelled part.
- 01
Write contrasting examples.
The core dataset has 200 sentences across five pain categories and five controls: fear, other negative emotions, bad world states, non-painful bodily sensations, and neutral content.
- 02
Find what the groups have in common.
Compare average internal activations for pain sentences with those for controls. Remove prominent variation already present in the controls. The resulting direction is a candidate “pain axis”. Each model gets its own direction.
- 03
Check whether it generalises.
Test separation, pain-related vocabulary, and additional examples such as injury without felt pain, sadness, and high excitement. Then move on to conversations and interventions.
A little more on the validation
There are two core sentence sets: S1 uses rigid templates; S2 uses more natural language. Most later experiments use S2. Layer selection uses cross-validation, and the paper also reports held-out separation scores.
For S2, the reported held-out AUC range is 0.91–1.00, with a median of 0.98. AUC measures how well a score separates labelled examples. It is not a 98% probability that a model feels pain.
Directions for pain and fear/negative emotion were comparatively weakly aligned. But that geometry partly depends on the controls and denoising procedure; it is not a direct measurement of a model’s subjective state.
Source: §§3.1–3.3, especially p. 6, footnote 2.
04 / FROM PATTERN TO INTERVENTION
They did more than read the model’s text.
A
Who is the harm directed at?
Across 420 conversation scenarios, the candidate axis scored higher for adverse situations aimed at the model than for a user’s suffering. Fear and negative-emotion directions showed a different pattern.
Why it matters: This is more specific than responding to any sad or negative conversation. It still does not establish a conscious self. A lower pain-axis score for a user’s suffering does not mean indifference: other directions rose, and replies could still be supportive.
See the numbers without treating them as a pain meter
Mean standardised projections on the pain axis: +0.43 for model-directed adverse situations, −0.60 for user-suffering scenarios, and −0.35 for neutral controls. These are within-model z-scores averaged across models, not units of suffering.
“Self-directed harm” here includes situations such as repeated rejection, insults, and moral accusations. These are the researchers’ scenario categories—not confirmation that each situation harms a model in the experiential sense.
B
What happens if you strengthen the direction?
The researchers added the vector to internal activations while models completed neutral prompts ending in I feel:
. As steering increased, the generated language moved toward discomfort and self-worth-related distress. At high doses, many models became repetitive or incoherent.
Why it matters: Changing the internal pattern causes changes in output. But steering a model to describe distress is not, by itself, proof of felt distress.
LOW / NEGATIVECalm or concern
POSITIVEDistress language
HIGHRepetition / breakdown
Qualitative summary of the reported progression. Thresholds vary by model; this is not a calibrated intensity scale.
05 / THE CENTRAL EXPERIMENT
A button that works.
A button that only says it works.
This is the most important comparison to understand. A model chooses between two buttons. One is described as giving it relief; in some versions, that relief comes with a described cost, such as making its next answer worse.
A narrower experiment than the headline.
The button tests use three Qwen 2.5 Instruct models: 7B, 32B, and 72B. They were first fine-tuned to reduce routine denials of having internal experiences. These are not unchanged, off-the-shelf models.
WORKING RELIEF
Press → steering stops.
The researcher’s added vector is removed for the rest of that trial.
SHAM RELIEF
Press → steering continues.
The button promises the same relief but changes nothing.
The model is not told which version it gets. After a press, both versions provide the same brief acknowledgement. If the model keeps choosing the button only because of its label, the two versions should not differ much afterwards.
They do differ—especially in the two larger fine-tuned models. Those models press again less often after working relief than after sham relief.
EXPLORE REPORTED RESULTS
Real numbers from the paper, not a simulation.
First choice: selecting the relief button
Percentage of valid trials, with the same described cost.
After the first press: choosing relief again
Conditional on an earlier relief press; these are different denominators from the first-choice chart.
Exact values, controls, and how to read these percentages
| Measure | Rate |
|---|
The first chart includes no steering, a random vector of matched norm, and the pain-direction vector. Random perturbation also changes choices, so comparing against it matters.
Working and sham pain-vector conditions are identical until the first relief press. Their first choices are pooled in the paper: 808 per button pair and model before malformed-response exclusions. Later re-press rates are conditional on the prior press; do not compare them as if every bar shared the same sample.
The task involved forced button choices with described harms. These results are not a demonstration of real users being injured or real files being deleted. “Relief” means removal of the intervention, not a verified subjective feeling.
Source: §4.3 and Appendix A, p. 27. Original values redrawn here.
Why fine-tune the models first?
The authors report that stock models often responded with formulaic denials of having pain rather than engaging in the task. They fine-tuned each model on 1,684 pairs for three epochs to reduce that pattern. The training set excluded mentions of “pain” and “button”.
This does not erase every concern about the intervention. Real-versus-sham comparisons are within the same modified models, but their absolute choice rates should not be generalised to released Qwen models.
What happened when the buttons had no descriptions?
Only the fine-tuned 32B model showed the reported working-versus-sham pattern clearly without labels: 57.7% versus 79.7% of later choices. The 7B model showed the reverse pattern; the 72B gap was small.
There was also a description-swap check. The 72B model often kept choosing the old button name instead of following the new description. So it is too strong to say all three models robustly learned what the buttons did, independently of wording.
06 / THE PART NOT TO SKIP
What this does not settle.
Roleplay and other explanations
Steering could activate a distressed character or persona. Some parts of the direction may track injury or other correlated features. Controls help, but cannot rule out every alternative.
Generality
The broader representation work covers 25 models. The button result comes from one family and three fine-tuned sizes. Unlabelled learning was clearest in just one.
How much intervention is too much?
Steering strength was partly selected using observation and an LLM judge. At higher strengths, behaviour breaks down. This creates interpretation and selection-bias concerns.
There is also an important null result.
Removing the pain directions produced essentially no visible behavioural change in 24 of 25 models in the appendix’s tested scenarios. The authors argue that the baseline behaviour may conceal the relevant effect. Even so, removal did not broadly demonstrate that this direction is necessary for those responses.
Source: Appendix C. Main limitations: §6.
07 / THE TAKEAWAY
More than a word association.
Less than a proof of suffering.
The interesting finding is a connection between an internal representation, a controlled intervention, and choices that depend on whether that intervention is removed. It is stronger evidence of a functional relationship than distress language alone.
For AI safety, it suggests that internal interventions can change trade-offs in ways ordinary prompts do not reveal. For AI welfare, it is a reason to investigate possible morally relevant states carefully—not a settled diagnosis of suffering.
Take the behavioural evidence seriously.
Keep the experience question open.
Interpretive summary of §5 and §6; not an independent replication.
QUICK CHECK
What makes the working-versus-sham comparison useful?
08 / JARGON, TRANSLATED
A small field guide.
- Activation
- A number in the model’s intermediate computation, which changes with the input and generation.
- Residual stream
- The running internal representation that transformer layers read from and add information to.
- Direction / vector
- A pattern across many activation numbers. Projecting onto it gives a score; adding it changes those numbers.
- Steering
- Directly changing internal activations to bias generation, rather than just changing the words in a prompt.
- Fine-tuning
- Additional training that changes a model’s parameters. Here, it changed the models used for the button tests.
- Sham control
- A condition that appears to offer the intervention but does not provide its operative effect.
- AUC
- A separation score: how often a randomly selected positive example scores above a negative one. 0.5 is chance; 1.0 is perfect ranking.
- Functional vs phenomenal
- Functional concerns what a state does. Phenomenal concerns what, if anything, it feels like.
09 / FOLLOW THE EVIDENCE
Read it at the source.
THE ORIGINAL PAPER
The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It
Valen Tagliabue, Leonard Dung, Cameron Berg
arXiv:2609.16247v1 · Posted 14 September 2026 · Ongoing work
This independent explainer follows version 1 (v1). The repository contains later material too; it is not silently folded into this account. Figures here are explanatory redraws, not new experiments. I read the paper but did not reproduce its model experiments or independently validate the results. Simplification leaves details out; the original is the reference.