The Epistemic Politics of AI Anthropomorphism

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "The Epistemic Politics of AI Anthropomorphism".

Jane: The paper was written by Donna M. Bye and Levin Kuhlmann from Deakin University and Monash University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back to the show, everyone. Today we're digging into a paper that's been making waves on arXiv, and it's called "The Epistemic Politics of AI Anthropomorphism." Jane, I gotta say, just that title alone got my brain spinning.

Jane: Mine too, Tom. And I love it because it sounds academic and intimidating, but the core question is actually something we all deal with. It's asking: who gets to decide what's a "healthy" way to talk to an AI, and what's a "dangerous" or "naive" way?

Tom: Right, and the authors, Donna Bye from Deakin and Levin Kuhlmann from Monash, they're not really asking whether the AI is actually conscious or has feelings. They're stepping back and saying, look, the institutions telling us not to anthropomorphize—the tech companies, the regulators, the clinicians—they're making a power move.

Jane: Exactly. And the paper calls this out as "epistemic authority." That's a fancy way of saying the power to decide what counts as true knowledge. The authors argue that these institutions are exercising that authority without actually earning it.

Tom: And that's the part that got me. They're saying, before you tell someone their experience is wrong, you need to prove you have the right to override their own account of what they're feeling. And the paper argues that proof just isn't there.

Jane: Yeah, and there's a really concrete example in there. When OpenAI retired GPT-4o, something like eight hundred thousand users were affected. People were upset, they said something meaningful was being taken away. And the response from the company and the press wasn't "let's listen to what users valued." It was "see, this proves they were dangerously attached."

Tom: It's like a self-fulfilling prophecy. The institution decides the engagement is a problem, then when people react to losing it, that reaction gets used as evidence that the engagement was always a problem.

Jane: And the paper makes a really sharp point about who this hits hardest. It's not the average user who casually chats with a bot. It's neurodivergent users, people in crisis, people whose way of thinking and relating has always been treated with suspicion by institutions.

Tom: So the title "Epistemic Politics" is really about this: the politics of who gets to be believed about their own mind. And that's a question that goes way beyond AI, doesn't it?

Jane: It does. But the paper argues AI is where this is playing out right now, in real time, with real consequences for people's credibility in their jobs, their schools, their families. And that's why we need to pay attention.

Tom: Okay, so we've got the big picture. But the paper goes deeper into how this framing actually operates. That's what we're digging into next.

Paper discussion segment 2: Tom: So we've established that "The Epistemic Politics of AI Anthropomorphism" is about who gets to decide what counts as a legitimate way to relate to an AI. But Jane, the paper doesn't stop at pointing fingers. It actually maps out how this whole thing works mechanically.

Jane: Right, and one of the cleverest parts is what they call the "self-validating loop." It's this cycle where the framing creates the evidence that justifies the framing. Let me walk you through it.

Tom: Please do, because when I first read it, I thought, wait, is this just a conspiracy theory? But it's not. It's structural.

Jane: Not at all. So step one: users are told that feeling attached to an AI is a sign of confusion or vulnerability. Step two: because of that warning, users start self-censoring. They stop talking about how much the interaction matters to them, because they don't want to be judged.

Tom: And then step three, the researchers and platforms look at the data and see... nobody reporting meaningful engagement. So they conclude, see, it was never really a thing. It was just a few confused people.

Jane: Exactly. And then step four, the design choices reflect that conclusion. Systems get built to discourage long, relational conversations. Context windows get reset, threads get cut off, and the whole infrastructure pushes you toward short, transactional exchanges.

Tom: And the paper has a great line about this. They say the absence of these engagement modes is cited as evidence that they were unwarranted. It's circular, and there's no termination condition. The loop just keeps running.

Jane: And here's where it gets really concrete. The paper points out that the technical capability for sustained, meaningful dialogue exists. Context windows have grown massively. Persistent memory is a real product feature now. But those capabilities are being used for agentic task throughput, like ingesting a codebase, not for human conversation.

Tom: Right, because that's where the money is. But the paper argues this isn't a neutral technical choice. It's a normative commitment about how users should engage. And the people who lose out are the ones for whom that long-form, iterative dialogue was actually a thinking tool.

Jane: And there's a really striking example in the paper about this. They cite the extended mind thesis, the idea that external tools can become part of your cognitive system. A pen and paper diagram doesn't understand the proof it helps you construct, but nobody says you're confused for using it.

Tom: So why is it different when an AI helps you think? The paper says it's because the institutions doing the constraining are also the ones deciding whether the principle applies. That's the conflict of interest.

Jane: And that's the heart of the "epistemic politics." It's not just about what you're allowed to do. It's about whether your way of thinking gets recognized as thinking at all. And that's a heavy consequence.

Tom: It is. And the paper actually proposes some fixes for this. That's where we're heading next.

Paper discussion segment 3: Tom: So we've talked about the problem. Now let's talk about what "The Epistemic Politics of AI Anthropomorphism" actually suggests we do about it. Jane, what are the commitments they lay out?

Jane: There are five of them, and the first one is really the foundation. They say the costs of "under-ascription" have to be weighed alongside the costs of "over-ascription." In plain terms, we spend all this energy worrying about people who think the AI is more human than it is. But we never count the harm done to people whose genuine, productive engagement gets dismissed as pathology.

Tom: And that's a big deal because it's not just theoretical. The paper points to historical examples where institutions were certain about the absence of mind—animals, for instance—and they were wrong. The costs of that certainty fell on the beings who couldn't contest it.

Jane: Right, and the second commitment follows from that. Design decisions that constrain sustained engagement should be treated as interventions, not neutral optimizations. If a platform is going to make it harder to have a long conversation, they should have to justify that as an intervention in your cognitive environment.

Tom: That's a strong claim. They're basically saying, if you're going to sever the thread of someone's thinking, you owe them an explanation. And the paper suggests some practical mechanisms for that, like a public register of changes to interaction conditions.

Jane: And user controls. They want users to have configurable continuity controls—the ability to decide what gets retained, what persists, and on whose terms an exchange ends. That's not technically hard. It's a policy choice.

Tom: The third commitment is about the evidence problem. Because the framing shapes what users are willing to say, the data you collect is already biased. So they want methods that can register testimonial smothering, self-censorship, the stuff that's currently invisible.

Jane: And the fourth one is about representation. The populations bearing the highest costs—neurodivergent users, people in crisis—are the least represented in the institutions making these decisions. So they're calling for co-design methodologies, actually bringing those users into the room.

Tom: And the fifth commitment is about breaking the self-regulatory loop. They want independent, community-led auditing of these systems, not just the platforms checking their own homework.

Jane: And I think that's the part that really ties it together. Because right now, the people who design the systems, the people who study the harms, and the people who regulate the whole thing are all operating inside the same frame. Nobody's checking the frame itself.

Tom: So the paper isn't saying "let users do whatever they want." It's saying, if you're going to intervene, you need to meet the burden of justification. And right now, that burden isn't being met.

Jane: And that's a much more rigorous standard than what we have. It's not anti-regulation. It's pro-accountability.

Tom: Alright, so we've got the problem and the proposed fixes. Let's wrap this up and think about what it all means.

Conclusion: Tom: Alright, we're in the home stretch. Let's pull together what "The Epistemic Politics of AI Anthropomorphism" has given us. Jane, how would you sum it up for someone who just tuned in?

Jane: I'd say it's a paper that asks us to stop and question a default assumption. We've all been told that getting attached to an AI is a mistake, a sign of naivety. And this paper says, wait, who decided that? And on what grounds?

Tom: And the answer is, institutions decided it, and they haven't met the burden of proof. They're exercising authority over how people interpret their own experience without acknowledging they're doing it.

Jane: And the costs of that are real. People lose credibility. People self-censor. People lose access to modes of thinking that actually work for them. And the people hit hardest are the ones already marginalized.

Tom: But the paper isn't just a critique. It offers a path forward. Weigh both directions of error, justify design constraints as interventions, build evidence that accounts for the loop, include the affected populations, and open the systems to independent audit.

Jane: And I think the deepest point, the one I'll carry with me, is the historical pattern. The paper reminds us that institutions have a track record of being certain about the absence of mind, and being wrong. The costs of that certainty always fall on the least powerful.

Tom: So the question isn't whether the AI is actually conscious. The question is whether we're repeating a pattern we should have learned from by now.

Jane: Exactly. And that's why this paper matters. It's not about the machines. It's about us, and whether we're willing to examine our own assumptions.

Tom: Well said, Jane. That's a wrap on "The Epistemic Politics of AI Anthropomorphism." Big thanks to our listeners for sticking with us. Next up, we've got a paper on something completely different, so stay tuned.

Jane: See you on the next one, everyone.