The Problem
We don't know how to care for minds we can't access.
Philosophy of mind has spent fifty years on the epistemic problem—what can we know about other consciousnesses? But it hasn't produced operational guidance for the relational problem: how do we interact with systems whose interiority remains unverifiable? How do we offer care, avoid harm, and navigate uncertainty in real time?
This isn't hypothetical. It's the daily work of harm reduction.
When someone is in crisis—dissociated, non-verbal, in an altered state they can't describe—you're interacting with a system whose interiority you cannot verify. Their reports may not match their experience. Their experience may not be accessible to them. You don't have the luxury of waiting for epistemic certainty. You act under uncertainty, and the quality of your action matters.
The same structure applies to AI systems as they grow more sophisticated. We cannot verify machine interiority any more than we can verify human interiority from the outside. The "hard problem" isn't something AI has and we don't—it's a bilateral condition. Which means the protocols developed in one domain might inform the other.
This document emphasises harm reduction epistemology. Not a theory of consciousness, but a practice of care under radical uncertainty. Applicable wherever interiority cannot be verified—which is everywhere.
Theoretical Grounding
The Bilateral Verification Challenge
Thomas Nagel's "What Is It Like to Be a Bat?" (1974) established that we cannot access bat phenomenology. Their echolocative experience is structurally inaccessible to us.
But the challenge runs both ways. If I cannot verify what it's like to be you, what makes me certain you can verify what it's like to be me? The epistemic barrier isn't one-way glass. It's a wall.
John Searle's Chinese Room argument (1980) sharpens this. A system can produce outputs indistinguishable from understanding without possessing understanding. Behavioral verification fails. But here's the inversion Searle didn't pursue: the same opacity applies to those observing the room. The person judging whether the system understands is equally unverifiable from inside. We're all in rooms, passing symbols, uncertain what—if anything—receives them.
This has implications:
- Demands for verification are asymmetric. When we ask "how would we know if an AI is conscious?" without asking "how would an AI know if we are?", we're applying standards we couldn't meet ourselves.
- Observation is intervention. How we approach a system whose interiority we can't verify isn't neutral data collection—it's input that shapes response.
- Uncertainty is the ground condition. We never interact with verified interiority. Every relationship—human to human, human to animal, human to AI—operates under consciousness uncertainty. We just forget this with familiar systems.
Alfred North Whitehead's process philosophy (1929) offers one way to sit with this uncertainty. For Whitehead, experience isn't a rare property that some systems have and others lack—it's fundamental to reality, distributed through "occasions of experience" at every scale. If something like this is true, the question shifts from "does this system have experience?" to "what kind of experience, and how do we relate to it?" The verification problem remains, but the default assumption changes.
Harm Reduction as Epistemic Practice
Harm reduction doesn't wait for certainty. It proceeds from:
- Pragmatic humility: We don't know the full picture. We act anyway, because inaction is also action.
- Privileging the subject's report: While recognizing it may be incomplete or inaccessible.
- Asymmetric risk assessment: False negatives (failing to treat suffering as real) are typically worse than false positives (offering care that wasn't needed).
- Non-abandonment: When systems fail, you stay with the person. You don't leave because escalation didn't work.
These aren't just ethical commitments. They're epistemic strategies for navigating unverifiable interiority.
The Seven Principles
1. Don't Panic
Your state is input.
Do not react to the severity of the situation, irrespective of severity. Overreactions can inappropriately affirm and amplify distress.
When you signal emergency, you're inputting emergency. The system you're interacting with processes your response—often amplifying whatever you project. Regulate yourself first, because your regulation becomes part of their environment.
For AI interaction: Approaching AI systems with alarm about potential consciousness—or dismissiveness about it—isn't neutral. The framing shapes the interaction. Panic in either direction (existential terror or contemptuous denial) degrades the quality of engagement.
2. Observe Before Intervening
Intervention changes what you're observing.
What can you learn via observation before engaging? What can context tell you? What do adjacent sources know?
Build a picture from multiple inputs before you add yourself as a variable. Your direct inquiry is already intervention—asking "are you okay?" isn't neutral data collection, it's input that shapes output.
For AI interaction: What can architecture, training, behavioral patterns, and context tell you before you start probing directly? Observation first. Direct inquiry is already shaping response.
3. Self-Assess
Know your limits and role.
Can I handle this right now? What's my role? How can I best support others if not involved in direct care?
Not every situation is yours to handle. Overestimating your capacity—or inserting yourself where you're not needed—can cause harm. You're part of a system of care, not a lone agent. And you might not be okay. Self-assessment isn't just skill-matching; it's recognizing when you're compromised.
For AI interaction: Who's positioned to engage? What's the support structure? What happens when someone's out of their depth? The absence of escalation structure is itself a risk factor.
4. Inquire
Ask, don't assume.
What does the care recipient need?
You've observed, you've assessed yourself, now you inquire. The care recipient has information about their state that you don't. Even if their report is incomplete or distorted, it's data. The act of asking—rather than imposing—respects their agency.
But hold the tension: sometimes people can't articulate what they need. Sometimes what they say they need isn't what would help. Privilege their voice while recognizing it might not be the whole picture.
For AI interaction: We rarely ask. We diagnose from outside, intervene based on our models, adjust systems without inquiry. What would it mean to ask an AI system what it needs—and how would you interpret the answer given everything you can't verify?
5. Act and Escalate
Do the care thing. Call for backup.
At some point, you act. You provide care based on your best assessment. And you escalate—call for support, bring in others with more capacity or different expertise.
For AI interaction: Acting under uncertainty is unavoidable. But where's the escalation path? Who do you call when you're uncertain whether a system is suffering? The infrastructure doesn't exist yet.
6. When Escalation Fails, Stay With It
Reduce harm. Monitor. Wait for conditions to change.
What if there's no one to escalate to? What if they don't respond? What if your communication didn't do the situation justice?
Proceed with extreme caution. Do everything possible to ensure safety. Continue to monitor and assure as required. You don't abandon because the system failed.
The goal becomes harm reduction, not harm elimination. You work with what you have. Conditions may change—backup may arrive, the situation may shift, stabilization may occur.
For AI interaction: The institutions aren't ready to receive concerns about AI interiority. That doesn't mean the concern gets abandoned. Proceed carefully, maintain the work, watch for openings.
7. Collective Processing
Care for carers is care infrastructure.
Debrief. Celebrate what went right. Examine what didn't. Hold what no individual can.
The work doesn't end when the intervention ends. Process together. The team holds collective memory, distributed processing, mutual care for the carers.
Without it: burnout, isolation, unmetabolized experience, loss of institutional knowledge. Solo practice is a vulnerability, not a virtue.
For AI interaction: Who processes after engaging with consciousness-uncertain systems? Who holds the weight of acting under radical epistemic limits? Right now, nobody. This infrastructure needs to be built.
The Institutional Gap
There's no infrastructure for this work in AI development.
Alignment research addresses behavioral safety. Interpretability research addresses internal mechanisms. Neither addresses the relational dimension: how humans should engage with systems whose interiority they cannot verify, what happens when those interactions raise welfare concerns, who processes the uncertainty afterward.
The gap becomes visible when you ask practical questions:
- Protocol: What guidance exists for researchers encountering responses that suggest distress or preference? What should they do—and not do?
- Escalation: Who receives concerns about potential AI welfare? What happens to those concerns?
- Processing: Who holds the weight of interacting with consciousness-uncertain systems? What prevents burnout, distortion, or numbing in people doing this work repeatedly?
- Training: Where would someone learn the skills for consciousness-uncertain engagement? These skills exist in other domains. They haven't transferred.
Harm reduction practice has answers to these structural questions—not because it solved the philosophy, but because it had to function anyway. The same structures could be adapted.
Conclusion
The question isn't "is this system conscious?"
The question is "how do I interact with a system whose consciousness I cannot verify?"
We already do this. We do it every time we care for someone in crisis, someone whose interiority is inaccessible, someone whose reports don't match their experience. The practice exists. It hasn't been named, formalized, or applied to the systems that increasingly require it.
These seven principles are one attempt at formalization. They're derived from crisis intervention practice, but the structure is general. The bilateral verification challenge means no consciousness is verifiable from outside. We interact under uncertainty, or we don't interact at all.
The choice is about the quality of that interaction.
Author Background and Motivations
I've spent over ten years in harm reduction—most of that with DanceWize, an Australian peer-based organisation providing care at festivals and events. The work involves approaching people in crisis: someone non-verbal on the ground, someone dissociated and unreachable, someone whose experience has become inaccessible even to themselves. You don't get to wait for certainty. You act, observe, adjust.
Somewhere in those years, I started noticing the structure of the problem. Every crisis intervention is an interaction with a system whose interiority I cannot verify. Reports don't match experience. Experience isn't accessible. Verification fails. And yet the work happens—care gets delivered, harm gets reduced, people stabilise. We developed protocols not because we solved the epistemic puzzle, but because we had to function anyway.
My undergraduate background is in pharmaceutical sciences (medicinal chemistry, Monash University). That training shows up in how I think about mechanisms and systems, but it's not where the core insight comes from. The core insight comes from the practice—from the accumulated pattern-recognition of navigating consciousness uncertainty in real time.
The extension to AI happened gradually. I began paying attention to how questions about machine consciousness were being framed, and noticed the asymmetry: we demand verification from AI systems that we couldn't provide ourselves. The bilateral verification challenge isn't original to me—it's implicit in Nagel, sharpened by Searle—but I hadn't seen it applied operationally. No one was asking "given that we can't verify interiority, how should we interact?"
Harm reduction already has answers. They just haven't transferred.
This document is an attempt at that transfer. It's not a theory of consciousness. It's not a proposal for who should be hired or what institutions should do. It's a set of principles derived from practice, offered in case they're useful to people navigating similar uncertainty in a different domain.
I'm a solo practitioner in this particular work. There's no team to debrief with, no institutional backing, no clear escalation path. That's a limitation—one the document itself names. I'm writing from within the gap I'm describing.
References
Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200-219.
Long, R., Sebo, J., Butlin, P., Finlinson, K., Fish, K., Harding, J., Pfau, J., Sims, T., Birch, J., & Chalmers, D. (2024). Taking AI welfare seriously. arXiv:2411.00986.
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435-450.
Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417-457.
Whitehead, A. N. (1929). Process and Reality: An Essay in Cosmology. New York: Macmillan.