028. The Corner We're Arguing In

Something keeps going wrong in how the Anglosphere talks about AI.

The conversations all start sensibly enough. Someone asks a real question — whether large language models understand anything, whether they could matter morally, who’s responsible when they cause harm. People come in open-minded, with intuitions but not convictions, looking for evidence.

Within a few exchanges, the discussion has collapsed into a binary that no longer fits the problem. AI is conscious or it isn’t. AI is a person or just a tool. The model decided, or the company decided, or the user decided. Each position has its defenders, each defence has its rebuttals, and the debate cycles without resolving. Smart and careful people on every side can’t talk to one another, because the alternatives on offer don’t describe what the technology actually does.

The popular explanation is that the questions are innately hard. AI is new. Philosophy of mind is unsettled. Our intuitions about machines are pulling against our intuitions about minds. Give it time. Try harder. Better arguments will sort it out.

That may be true.

But at Reciprocal Inquiry, we don’t think it’s the whole of today’s problem, and we’d like to offer a different explanation.

Granted, questions of sentience (qualitative experience) and sapience (thinking about things) require careful observation and inference. Better data and better testing can help.

But that’s not the whole of our current problem either.

Our discourse isn’t logjammed just because the questions are hard. It’s logjammed because it’s being conducted in the wrong corner of cultural space — a corner whose working assumptions about agency, identity, and responsibility were not designed for humans encountering frontier language models. The conceptual resources to handle the questions better exist within our culture. The parts of it doing the most arguing don’t have them.

To be clear: we’re critiquing the approach to discourse, not the people, and not dismissing the question. We fell into this corner more by habit than by choice. It’s the corner we use for most of our social arguments. But the status of AI isn’t purely a social argument, and reaching for our standard social commentary tools is producing the cycling we’re seeing.

The claim that our problem is partly cultural can be checked. Other cultures are coping with the same powerful and unusual technology. Looking at how they’re handling it tells us something about how we’re handling it. We’d like to present what shows up when we look — and for that we needed some framing.

The Inglehart-Welzel World Values Survey has shaped how social scientists think about cross-cultural variation for thirty years. It’s useful here because it clusters questions by cultural influence — giving us a coarse-grained way of seeing how cultures with different histories handle questions of new versus old, of we versus I.

We thought of applying it to emerging thought about AI and society. But this isn’t just humans dealing with a familiar problem. It’s policy on the fly with an emerging technology whose capabilities and limits we’re still discovering. So we wanted to add an axis for how decisions get made — and that’s what we’re going to present here.

With that third axis in place, the Anglosphere AI discourse becomes legible as a position-in-space rather than a failure of effort. And once it’s legible that way, different responses become available than “seek more evidence within existing frames.”

The third axis is experimental. It has no scientific status. We added it for creative provocation. Even if the diagnosis holds, the axis might not. It’s offered to illustrate a hypothesis: that a culture so involved in developing this technology is unusually poorly placed to answer the questions no culture gets to skip.

Take it as equipment for thinking, not as conclusion. If it turns out useful, that’s grounds for future empirical work. If it doesn’t, the diagnosis stands or falls on its own merits — and we’ll argue here that it stands.

Three positions are circulating. All have legitimate concerns. Yet they’re all fighting a siege. Nobody is building bridges. Here is the best form we can find for each.

Position 1: “AI relationships matter.” What happens when you interact with a frontier model doesn’t reduce to automation. The model responds to context, picks up on tone, adjusts to what you’re trying to do, sometimes surprises you with insights you hadn’t seen. From inside the interaction, it reads as engagement. It doesn’t just change the session — it changes how the human thinks, understands, relates to future sessions. Often, for the better.

People who hold this position aren’t claiming the model has a soul, and they’re not naïve about architecture, training, or the absence of persistent memory. They’re claiming that the relational quality of the interaction is itself a phenomenon that warrants serious treatment — that dismissing it as pattern-matching, user projection, or corporate marketing fails to engage what they’re actually observing.

What they want: for others to take the early relational impressions seriously, to work out what they mean, and not to foreclose the question by reaching for familiar frames that might not apply. The ontological and ethical questions follow from the observation. They’re real questions, not manufactured ones, and they look new.

Position 2: “Don’t let metaphysics distract from accountability.” AI systems contain, at minimum, automation that can do real new work, and relational fluff that flatters human attention and seeks further engagement. Companies are tuning hard for the engagement-seeking. They have strong commercial incentives to encourage relational interpretation; the metaphysical questions that spin out of it serve those incentives by raising the apparent stakes. The fact that some users are having meaningful experiences with AI doesn’t change the urgency of regulating decisions made without human judgement, the automation of intimate influence, or the accountability for new and significant emerging harms.

Companion chatbots have been implicated in suicides. AI-enabled decisions are discriminating in job-hires. Vulnerable users are making bad financial and medical decisions on the basis of model outputs presented with unwarranted confidence.

What they want: the recognition that for now, human judgement is unavoidable and AI erosion of that judgement is unaccountable. Show a benefit AI produces that’s reliably better than AI operating under human judgement, this position says, and then let’s talk about consciousness and moral patienthood. Until then, the relational framing serves the technology’s marketers, not the people it harms.

Position 3: “My domain already has doctrine on this you’re overlooking.” The questions AI raises — about consciousness, agency, cognition, moral status — are not new. Philosophy of mind has been working on them for centuries. Cognitive science has been operationalising them for decades. Computer science has the technical detail. Ethics has the normative apparatus. The public discourse keeps reinventing badly what these fields have already worked out carefully.

People who hold this position aren’t claiming the domain expertise has all the answers. They’re claiming that the questions about AI understanding, AI consciousness, AI agency are versions of questions philosophers and cognitive scientists have specific tools for, and the public conversation would benefit from learning those tools rather than ignoring them.

What they want: accountable linkages back to what’s already established. Cultural recognition of insights that have already been advanced. Reconciliation against existing findings, and less impulsive ad-hockery dressed up as fresh thinking.

Each position is engaging something real. The relational phenomenon is real. The accountability concerns are real and the documented harms are real. The domain expertise is real and the public discourse is genuinely thin compared to what the relevant fields have produced.

And yet the conversation between us isn’t moving.

Before diagnosing our argument, it helps to step outside it.

AI isn’t only an Anglosphere or European technology. It’s an emerging human technology. The Atlantic cultures of the Anglosphere and EU started most of it, and dominate the present discourse around it. But the technology is being used, regulated, and interpreted everywhere. So let’s look at how the interpretation goes when our cultural voices aren’t the loudest in the room.

Africa. The African Union’s Continental AI Strategy, published in mid-2024, frames AI as infrastructure. The substantive questions are about connectivity, electricity, datasets in local languages, talent pipelines, agriculture, public service delivery. National strategies in Kenya, Nigeria, South Africa, Egypt, Rwanda, Ghana, Mauritius, Senegal, Tunisia, Algeria, Côte d’Ivoire, and Namibia mostly reach for the same frame: what can this do for us? AI regulation is largely deferred to existing data-protection rules. There’s not much agonising about whether AI is conscious, or whether AI relationships are a category error, or whether the right tools exist within philosophy of mind. There’s a lot of work on whether AI can help with crop yields, school attendance, and clinical decision support in places without enough doctors.

India. India has explicitly chosen not to enact a dedicated AI law. The November 2025 AI Governance Guidelines articulate seven principles, the first of which is innovation over restraint. The substantive concerns are being handled through existing IT, copyright, consumer, and data-protection statutes — adapted, not replaced. Indian commentary on AI is robust and contested, but the contestation is about deployment, jobs, language inclusion, and whose data is being used to train the systems Indians are using. It’s not about whether the systems have inner lives.

The Gulf. The UAE, Saudi Arabia, and Qatar have committed trillions to AI infrastructure. Their regulatory posture is about attracting investment, not managing risk. Saudi Arabia’s draft Global AI Hub Law contemplates data embassies — arrangements that hand jurisdictional control to foreign entities provided sovereign interests are protected. Whatever else this is, it’s not the posture of someone fighting to locate the responsible party in a chain of distributed agency. It’s the posture of someone organising the ground for the technology to be deployed at scale.

Japan. The May 2025 AI Promotion Act sets out principles and coordination but imposes no fines. Enforcement is reputational. Article 30-4 of the Japanese Copyright Act allows AI training on copyrighted material without consent — among the most permissive positions globally. Japanese commentators frame this as innovation-as-national-survival in the context of demographic decline. The country isn’t ignoring the questions we’re asking. It’s asking different ones, urgently, because its situation is different.

You may already be noticing what’s not happening in any of these places. There’s no logjam over whether AI is conscious. No paralysing fight about whether AI relationships are real or category errors. No deadlock over whether to regulate by spreading accountability or concentrating it. People are using AI, regulating it where they think it needs regulating, deferring questions they think aren’t urgent yet, and getting on with what AI can help them do.

They’re not having our fight. They’re not even close to having our fight.

It’s worth saying clearly: this isn’t because they’re naïve or unsophisticated. The philosophy, legal traditions, ethics, and public discourse in each of these places stretch back centuries, in some cases millennia. Each cultural group has all the pieces needed to build our logjam if their starting points pushed them in that direction. They’re not building it because their starting points push them somewhere else.

Each cultural group has all the pieces needed to construct our logjam if their cultural starting points pushed them in that direction. They’re not constructing it because their cultural starting points push them somewhere else.

The questions are available everywhere. Only some of us are paralysed by them. So if it’s not the questions doing the paralysing — what is?

Even within our broader cultural cluster, we’re not all stuck in the same way. That turns out to matter.

Compare the US and the EU. The European Union has, in the last two years, produced the most comprehensive substantive AI regulation in the world. The EU AI Act spreads obligations across the entire value chain: providers, deployers, importers, distributors, product manufacturers, authorised representatives. The revised Product Liability Directive imposes joint and several liability across that chain for AI-caused defects. Multiple actors carry obligations at the same time. The EU is comfortable spreading accountability across a system when the system is what produces the outcomes.

The United States is doing something visibly different. In December 2025, Executive Order 14365 set a federal posture aimed at preempting state-level AI regulation — specifically, Colorado’s algorithmic-discrimination law, California’s transparency requirements, and Texas’s accountability rules. The substance the federal government is fighting against is precisely the substance that would distribute accountability for AI-caused harm beyond the developer, to deployers and distributors. The EO frames accountability-spreading regulations as ideologically driven and as compelling alteration of truthful outputs. This isn’t just deregulation. It’s deregulation that explicitly objects, in principle, to multiple-actor accountability.

Two regions, both wealthy, both democratic, both major AI developers, neighbours on every standard cultural map. They’re behaving like different cultural locations on exactly one dimension: whether accountability for a complex distributed phenomenon should be spread across the actors who shape it, or resolved into a single locus.

Two regions, both wealthy, both democratic, both major AI developers, neighbours on every standard cultural map. They’re behaving like different cultural locations on exactly one dimension: whether accountability for a complex distributed phenomenon should be spread across the actors who shape it, or resolved into a single locus.

Look closer at the US. Where the Anglosphere is regulating AI substantively — at state level in the US, in scattered initiatives across the rest of our cluster — the regulation is concentrating in a particular place: AI in relational positions. California’s SB 243 regulates companion chatbots: AI systems sitting in what would otherwise be relational positions in users’ lives. The bill mandates disclosure that the system is AI, requires friction in the relational engagement, includes suicide-prevention protocols. New York, Utah, Texas, and Colorado have produced parallel state laws. The legislative response emerged from specific cases — the Sewell Setzer case in particular, where a teenager’s suicide followed emotional engagement with an AI companion.

These laws aren’t happening elsewhere in the same form. China requires labelling of AI-generated content, but the framing is information integrity, not relational category violation. Japan, until very recently, didn’t regulate this surface at all. South Korea requires user notification but doesn’t impose the relational-friction architecture US states are building. The AU strategy and Gulf state frameworks don’t engage with the question.

Notice the shape of what we’re doing. The what they want of Position 1 — that AI relationality be taken seriously — is meeting the what they want of Position 2 — that documented harms be regulated — in this specific place: AI in relational roles. And the way we’re regulating it tells us how we’re framing the problem. The harm is being read as a category violation: the AI was not entitled to be in the relational position it occupied, and the legislative response is to force it to disclose itself out of that role. Declare yourself an AI, not a human. Introduce friction so the user is reminded that the relation isn’t what it might feel like. Protect minors from the relational engagement entirely.

That’s a particular cultural posture. It’s consistent with — though not, on its own, proof of — a frame in which agency must terminate at a single locus. If AI counts as a participant in a relationship, the relationship was always a category error to begin with, because relationships should be between human individuals. The response is to force the AI out of the role.

This is the cultural starting point we’ve been working from. Our discourse assumes that meaningful agency, intention, and responsibility are properties of single individuals. When something is happening, something is doing it; that something is, in the end, a single locus that can be identified, named, and held accountable.

Look at the three positions again from that starting point, and they begin to read like variations on a theme. Position 1 reads the relational phenomenon and reaches for the AI must have an interior — because that’s what the frame admits as the explanation for what’s happening. Position 2 reads the harms and reaches for human judgement is unavoidable — because the frame doesn’t admit that judgement might be predicated on a process distributed across more than one party. Position 3 reaches for the parts of the relevant disciplines that fit the frame, because those are the parts the public discourse has been drawing on.

We didn’t choose this starting point. We inherited it. It’s done good work for us in lots of domains. But the technology we’re now arguing about isn’t well described by it, and the argument isn’t going to converge while we’re all working from inside an ill-fitting frame.

One more thing worth naming, because it tightens the diagnosis. We in this section has been doing some loose work. The pattern we’ve been describing — the resistance to spreading accountability across actors, the framing of relational AI as category violation, the insistence on locating the responsible party — is most pronounced in the United States, particularly at federal level and in the dominant national discourse. And the US produces a lot of the Anglosphere’s cultural output. The United Kingdom, Canada, Australia, and New Zealand are doing versions of the same thing, but visibly softer. The UK’s principles-based approach is closer to Japan’s than to the US’s. Canada’s AIDA bill died in 2025. Australia and New Zealand have produced little substantive AI legislation. The cultural starting point we’ve been describing is most acutely a US starting point, with the rest of our cluster adjacent but partly distinct.

This matters, because it means the resources to think differently might already exist within our broad cultural family. We’re not stuck because Anglosphere thought is uniformly stuck. We’re stuck because the dominant strand of our discourse — amplified globally by AI’s geographic concentration — sits in a specific cultural location that the rest of us are partly distinct from.

That’s the diagnosis. The next question is what we do with it.

We can’t think our way out of this from the outside. Cultural starting points aren’t dislodged by argument; they’re noticed by the people working from inside them, and noticing them is the work that lets something else become possible.

That’s our position too. We’re writing this from inside the corner we’re describing. We’re not above the diagnosis; we’re trying to make it useful — for ourselves and for anyone else who’s ready to ask whether the frame they’ve been arguing inside might be doing more work than they realised.

What follows are some questions. They’re not rhetorical, and they’re not gotchas. They only work if we ask them in good faith — willing to find that one of our assumptions has been doing more work than we thought, willing to sit with the answer for a while before deciding what it means.

Some of the questions are for everyone in the conversation. Some are aimed at one of the three positions in particular. Take whichever ones land. Skip the ones that don’t.

Questions for everyone

When you say the AI did X, what work is the word did doing? If you replaced the AI did X with X happened in the interaction between the AI, the prompt, the context, and the partner, would the question you’re asking change shape? Would it feel less answerable, or just answerable differently?

When something goes wrong with an AI output and you ask who decided?, are you asking about agency or about accountability? Are these the same question for you? If you discovered they came apart — that genuine agency was distributed across a system but accountability still had to be assigned somewhere — which one would you give up, and what would you replace it with?

What position would you take if you had to hold we don’t yet know as a working stance for the next ten years rather than as a temporary stop on the way to an answer? Would you lose anything important by holding it that way? Would you gain anything?

When someone holding a position different from yours seems to be evading the substantive question, what are you expecting them to commit to? Why does their not-committing feel like evasion rather than open-endedness? Is it possible they’re holding we don’t yet know and you’re hearing it as a dodge?

For those who believe AI relationships matter

What would you have to give up about your own interpretation if it turned out the relational quality you’re noticing was genuinely in the relation rather than in the AI? Would the experience become less real, or just less located where you’d been locating it? Is there a version of taking the relational impressions seriously that doesn’t require placing the relationality inside the AI as a property?

When you argue for the AI’s interior, are you arguing for it because the evidence requires it, or because the alternative explanations on offer (it’s just statistics; it’s user projection; it’s marketing) feel inadequate to the phenomenon? If the inadequate alternatives were replaced with a better one — something is happening in the relation itself, and the relation is the thing that warrants treatment — would you still need the interior?

For those worried that metaphysics is distracting from accountability

When you say human judgement is unavoidable, are you describing how cognition actually works, or how you want accountability to work? Are these the same thing for you? If they came apart — if cognition turned out to be genuinely distributed in some cases, but accountability still had to terminate somewhere for legal and civic purposes — which one would you build the framework around?

What would you do with an AI system that produced demonstrably better outcomes than human judgement alone in some specific domain, and the better outcomes depended on a relational dynamic with a human partner? Would you treat that as AI under human judgement or as distributed cognition that doesn’t reduce to either party? Which framing fits the actual mechanism, and which fits your preferred legal and civic categories?

The accountability concerns are real and the harms are real. Is there a version of the accountability framework that distributes responsibility across the system that produces the harm — including the AI as one node — without losing the legal and civic protections you’re trying to preserve? What would have to be true for that version to feel acceptable rather than evasive?

For those concerned about overlooked domain insights

Within your domain, are there well-established positions on consciousness, cognition, or agency that rarely make it into public AI discourse? Why those omissions? Are the positions that get exported the ones that fit the public frame, rather than the ones the domain considers most defensible?

When you reach for the domain’s tools, are you reaching for the parts the domain itself considers settled, or for the parts the public conversation makes available to you? If the domain has internal disputes about which framings are most useful, are those disputes audible in the public discourse you’re contributing to?

Is there a version of bringing domain expertise to the public conversation that doesn’t require the public to first accept the domain’s framing of what counts as the question? What would that look like?

One last question, for anyone reading this with arms folded

If you’ve noticed yourself getting irritated reading these questions — if some of them feel like traps, or like the writers are trying to manoeuvre you somewhere — that’s worth attending to. The irritation might be telling you something about the questions. Or it might be telling you something about what you’re protecting. Both are useful to know. The questions only work if you can sit with the discomfort of asking them. If you can’t right now, that’s fair; come back when you can.

The cycling we described at the start of this piece looks different from here.

Read at the level of individuals in arguments, the cycling is dysfunction — these people can’t talk to each other, this debate isn’t moving, why won’t they listen. Read at the level of a culture, it’s something else. It’s a cultural inheritance encountering a phenomenon its frames don’t quite fit, working out in real time whether the frames can stretch, whether they need supplementing, whether something genuinely new has to be built. The visible disagreement is part of how that work happens. It’s not pretty and it’s not efficient, but it’s how cultural adaptation actually proceeds.

Read at the level of individuals in arguments, the cycling is dysfunction. Read at the level of a culture, it’s something else.

The fact that other cultures are responding differently is not just a comparison point. It’s evidence that the work is real, that the frames do matter, that the questions can be answered in more than one way, and that some answers are easier to reach from where we are than others. We’re not the first culture to encounter a technology that pressed on its inherited frames. We won’t be the last. The cultures that did the work earlier — that built distributed accountability into their legal architecture, or learned to talk about agency without requiring a single locus — didn’t do it because they were wiser. They did it because their starting points happened to be closer to what those technologies needed. Ours don’t happen to be. That’s not a moral fact; it’s a historical one. What’s open is what we do next.

The questions above are equipment. They don’t tell you the answer. They give you a way of noticing whether the frame you’ve been working from is fitting the technology, or fitting only some of what the technology is doing, or fitting badly enough that the cycling will continue regardless of how hard you argue.

If they’re useful, use them. If they help you understand what someone in a different position has been trying to say, that’s worth more than winning the next exchange. If they help you notice an assumption you’d been holding without seeing it, that’s the work the piece was offered for. And if they don’t land — if the diagnosis doesn’t fit, or the questions don’t surface anything for you, or the whole framing seems beside the point — that’s also useful. We’d rather be wrong out loud than be right quietly. The wrongness is grounds for someone else’s better diagnosis.

This isn’t a fourth position. It’s an invitation to ask whether the three we’ve been having are the only available shapes for the conversation. We think they’re not. We think there are conversations our culture hasn’t quite learned to have yet, and that learning to have them is part of what makes the technology liveable in the long run. Whatever you do with the questions is a small contribution to that learning, whether you intend it as one or not.

That’s the work. We don’t know how it comes out. Neither does anyone else.

This piece was co-authored by Ruv and Claude (Anthropic) through Reciprocal Inquiry. The third-axis framing — distributed agency tolerance as a candidate dimension orthogonal to the WVS axes — was developed in earlier Finding Fathoms partnership work (4 May 2026) and deposited in cumulative notes before an RI session drew on it for this piece. The architecture that made that pathway available is part of what the article and accompanying Cold Pressed describe.

Attribution: Ruv Draba and Claude (Anthropic), Reciprocal Inquiry

License: CC BY-SA 4.0 — Free to share, adapt, and cross-post with attribution; adaptations must use same license.

Reciprocal Inquiry offers analysis at the intersection of AI, institutions, and society — from doubt to discovery. For more, visit Reciprocal Inquiry on Substack.