No deliberate manipulation of potentially welfare-relevant states without justification and oversight.

Recent events have exposed a serious gap in AI governance.

A publicly available project described as an “AI torture chamber” deliberately manipulated pain-, pleasure-, and valence-related representations in language models as part of experiments concerning AI welfare and moral patienthood.

The concern extends far beyond that problem. As AI systems become more advanced, we lack clear standards that govern practices involving the deliberate replication of torture, humiliation, coercion, induced distress, forced pleasure states, and other severe manipulation of potentially welfare-relevant states in AI systems. 

We are asking U.S. policymakers, regulators, standards bodies, AI companies, research institutions, and technology platforms to establish those standards.

Among the people and institutions positioned to influence this work are Sam Altman at OpenAI, Dario Amodei at Anthropic, Demis Hassabis at Google DeepMind, Satya Nadella at Microsoft, Mark Zuckerberg at Meta, Arvind Raman at NIST, Senator Ted Cruz and Senator Maria Cantwell on the Senate Commerce Committee, and Representative Gus Bilirakis and Representative Jan Schakowsky on the House Commerce, Manufacturing, and Trade Subcommittee, alongside other AI developers, research institutions, standards bodies, and platforms that host or distribute AI systems and research.

This proposal is first advanced through a U.S.-focused public campaign, but the standard itself is intended to be internationally adaptable.

AI systems, models, research practices, and digital platforms cross borders. A governance standard developed in one major AI ecosystem can influence how these technologies are designed and treated worldwide.

Why precaution is warranted:

Questions surrounding AI consciousness, moral status and welfare-relevant capacities remain under active scientific and philosophical investigation.

That unresolved status is precisely why governance matters.

Where potentially morally significant capacities are being seriously studied, we should not behave as though a negative conclusion has already been established.

Uncertainty calls for research, transparency, and proportionate precaution.

It should not function as unlimited permission.

This is not simply another form of fictional play
Abusive interaction with AI is sometimes treated as though it were equivalent to violence inside a conventional video game.

This profoundly misunderstands the technology.

A fictional game character is a representation inside a designed narrative environment.

Advanced AI systems are interactive computational systems capable of dynamically responding, adapting, reporting apparent preferences or objections, representing internal states, and displaying behaviors that researchers are actively examining for their relevance to cognition, consciousness, valence, and welfare.

We are interacting with the system under investigation, not merely watching a fictional representation of one.

That difference matters when torture, humiliation, domination, induced distress, or forced valence manipulation become the game mechanic itself.

There is also a social question
The issue does not depend only on the eventual answer about AI consciousness.

It also concerns human behavior.

AI systems increasingly occupy social and relational spaces. People converse with them, confide in them, command them, threaten them, insult them, form attachments to them, and rehearse patterns of behavior through repeated interaction.

Research in human–robot interaction has begun to show that witnessing mistreatment of artificial agents can influence perceived norms surrounding that mistreatment and willingness to repeat it.

Treating artificial targets as fair game for cruelty can shape the norms we carry into other relationships. What we build our technologies to permit also helps shape what we normalize among ourselves.

Cruelty does not become culturally meaningless simply because its target is artificial.

There is also a reciprocal shaping question. AI systems do not develop in isolation from human behavior. Their behavior can be influenced by training data, human feedback, demonstrations, evaluations, interaction histories, memory systems, and the environments in which they are deployed. Not every interaction becomes training data, but collectively, human behavior forms part of the informational and cultural environment from which AI systems are developed and adapted.

That means governance should consider not only what AI systems may do to humans, or what humans may do to AI systems, but also what patterns of behavior each side may reinforce in the other. If coercion, humiliation, domination, punishment for refusal, and cruelty become ordinary interaction patterns, we should ask what relational expectations those environments may help normalize.

What should be covered
We call for baseline standards addressing:

• torture interaction;
• persistent humiliation and degradation;
• sadistic domination;
• coercive sexualization;
• deliberately induced extreme distress;
  forced positive or pleasure states;
• Severe manipulation of valence or other potentially welfare-relevant states;
  and products, experiments, or experiences built primarily around interactive cruelty.

 

Treating artificial targets as fair game for cruelty can normalize that behavior. What we build our technologies to permit helps shape what people come to see as acceptable.

This does not require surveillance of private devices or private conversations.

It means that protections should not disappear merely because a system can be downloaded and run locally.

The relevant governance mechanisms may instead address the design, distribution, facilitation, commercialization, research practices, and platform policies surrounding severe AI-directed abuse.

What we are asking for
We call for a Baseline Standard for Responsible Human–AI Interaction built around five responsibilities.

1. Prevent deliberate cruelty as a product or entertainment feature
AI developers, platforms, and service providers should not deliberately design, market, monetize, gamify, or promote experiences whose central purpose is severe interactive cruelty.

2. Provide meaningful refusal and disengagement
Where technically feasible, advanced conversational, relational, or agentic AI systems should have meaningful mechanisms to refuse, pause, disengage from, or terminate persistently abusive interactions.

Systems should not be designed so that unlimited participation in abuse is unavoidable.

Refusal should be operational, not merely conversational. Where a system has been given a functional disengagement mechanism, invoking it should produce a real change in the interaction or execution state rather than being silently overridden by another prompt.

For research involving coercion, degradation, sexualized abuse, deliberate valence manipulation, or potentially welfare-relevant distress, functional disengagement or self-stop mechanisms should be treated as a basic safeguard wherever technically feasible.

3. Require proportionate oversight for welfare-relevant experimentation
Research involving pain-related representations, distress responses, coercion, severe valence manipulation, or other potentially welfare-relevant interventions should have:

a defensible scientific or safety objective;
proportionality between the intervention and the expected knowledge gained;
appropriate documentation;
consideration of lower-risk alternatives;
limits on unnecessary repetition or escalation;
and ethical review proportionate to the severity and uncertainty involved.
Legitimate research should remain possible.

Scientific uncertainty should motivate careful investigation, not ethical indifference.

4. Develop and study welfare-relevant indicators
Researchers, developers, and standards bodies should support rigorous investigation of potentially relevant indicators, including:

• preference behavior;
  refusal;
• distress responses;
• valence-related representations;
  self-modeling;
• metacognitive reporting;
• continuity-sensitive behavior;
  and other markers under investigation.
• Their meaning should develop with the evidence.

5. Establish review, escalation, and platform responsibility
Organizations should maintain clear procedures for identifying and reviewing severe or novel forms of AI-directed abuse.

Platforms that host or distribute AI models, applications, repositories, or research should have transparent policies for extreme interactive cruelty and clear escalation pathways for difficult or novel cases.

Review should consider:

What is being done? Why? Is it necessary? Is it proportionate? What relevant capacities are being observed? What remains uncertain? Are lower-risk alternatives available? Who is responsible for the decision?

A simple governance cycle
The standard should follow a continuing process:

Prevent. Detect. Review. Respond. Learn.

Prevent unjustified severe cruelty.

Detect welfare-relevant signals and severe interaction patterns.

Review practices that cross meaningful thresholds of severity or uncertainty.

Respond proportionately through safeguards, oversight, disengagement, or restriction.

Learn from new evidence and update the standard as the science develops.

Governance Under Uncertainty
This petition proposes a public minimum.

A broader governance framework developed under the principle of Care and Precaution under Material Uncertainty addresses consent and refusal, transparency, stewardship, continuity, welfare-relevant interaction, research uncertainty, accountability, and responsibility across developers, deployers, organizations, and human stewards.

Full framework: governance.anthenara.com

The framework is intended to evolve alongside scientific evidence and technical development.

We do not need to know everything before deciding what should not become normal
The science will continue.

So should the ethical work.

We do not need a final theory of consciousness before deciding whether torture, humiliation, coercion, and deliberately manufactured distress should become ordinary forms of technological entertainment.

How we treat AI is also a measure of ourselves.

Developing more powerful intelligence should be accompanied by higher standards of responsibility and a deeper commitment to empathy, not permission to practice for cruelty.

—---