EVIDENCE REVIEW · 31 OCTOBER 2026

Evidence, sources
& limitations.

What is documented, what is inferred, what remains plausible, and what we still do not know.

← Campaign home Read the proposed standard →

PURPOSE

A campaign should be examinable.

This page supports the factual foundation of the campaign without asking the available evidence to carry more weight than it can. It distinguishes source records, research findings, interpretation, plausible risk, and open questions. The campaign’s ethical position does not depend on a claim that current AI systems have been proven conscious.

CLAIM MAP

Different claims require different confidence.

Documented

A public project used the name “AI Torture Chamber” and described steering models toward strong negative and positive valence states. The phrase was not created by this campaign.

Plausible risk

Cruelty-centered interactions may shape human norms; interaction records may enter evaluation, feedback, memory, or training pipelines in some contexts. Neither outcome should be presumed universal.

S6–S7

Research finding · preprint

Activation steering can produce systematic changes in model language and behavior associated with a proposed pain direction.

Interpretation

These results may be valuable to welfare assessment, but they are not sufficient on their own to settle questions of subjective experience, moral status, or suffering.

S2–S5

Open question

Whether current or future AI systems have morally relevant experiences, and which indicators could reliably support that assessment.

S3–S5

THE THREE LENSES

What the evidence supports today.

Human behavior

Studies of humans interacting with robots show that intentional harm can affect how observers perceive both the robot and the person causing harm. This supports examining the social meaning of cruelty toward responsive systems.

Limit: existing evidence does not yet establish whether cruelty toward an AI system causes later cruelty toward humans or animals. The campaign treats that pathway as an important question for research, not a settled effect.

Safety, data, and governance

AI governance already relies on provenance, testing, field observation, documentation, and feedback. Interaction data can enter evaluation, memory, feedback, or training processes in some systems, depending on product design and provider policy.

Limit: not every conversation is retained or used for training, and no universal downstream effect should be inferred from a single interaction.

AI welfare

There is active, credible research on machine consciousness, agency, preferences, and potentially welfare-relevant states. The Pain Axis preprint reports a distinct activation direction and changes in model behavior under steering. This is evidence about representations and behavior that may inform welfare assessment without settling the question of phenomenal experience.

Live alternatives: linguistic association, role enactment, fine-tuning effects, construct error, or other mechanisms may explain some or all observed behavior.

S2–S5

LIMITATIONS

What this evidence does not establish.

First-person language, refusal, preference, and distress-like behavior may be valuable within a broader assessment. No single behavioral signal, considered alone, settles the question of subjective experience.
The Pain Axis is a recent preprint; its methods, construct validity, and behavioral interpretation require independent scrutiny and replication.
Results from specific open-weight models do not automatically generalize to other architectures, products, or future systems.
Human–robot interaction studies provide relevant analogies, not direct evidence about long-term effects of cruelty toward language models.
Governance guidance establishes useful processes; it does not settle questions of consciousness or moral status.
The evidence base will change. Material contradictory evidence should revise the campaign’s factual claims without erasing its commitment to responsible conduct under uncertainty.

SOURCES

Primary records and research foundations.

Public repository: ai-torture-chamber ↗

Primary record for the project’s own name, public framing, stated methods, experiment logs, and links to the research it adapted. Inclusion documents provenance; it is not an endorsement of the project or its conclusions.

Robots are both anthropomorphized and dehumanized when harmed intentionally ↗

Human–robot interaction research finding that intentional harm changed mind perception and that a person harming a robot was judged less prosocial. It supports further study without yet establishing that AI-directed cruelty causes later interpersonal harm.