Four years into the generative AI revolution, we still cannot answer a deceptively simple question: what is a large language model?
Ask an AI researcher and they might describe the Transformer architecture — layers of attention mechanisms, feedforward networks, next-token prediction trained on trillions of tokens. Ask a cognitive scientist and they might map LLM behavior onto human psychological constructs: reasoning, memory, attention, even theory of mind. Ask a philosopher and you get the hard problem of consciousness. Ask an economist and the LLM becomes a market aggregating dispersed knowledge, with tokens functioning as price signals.
None of these answers are wrong, exactly. But according to a provocative new paper published in Nature on July 16, none of them are quite right either. And the fact that we keep reaching for human-shaped metaphors to explain machine behavior is itself the problem.
The paper, titled “Understanding large language models demands distinguishing human projection from machine cognition,” argues that the entire field of LLM interpretability has been trapped in what the authors call a “recursive loop of anthropomorphism.” Every discipline brings its own metaphors — physics, neuroscience, psychology, sociology — and each illuminates some facet of LLM behavior while systematically obscuring others. The result is an intractable deadlock between two camps: those who insist LLMs exhibit genuine understanding, and those who dismiss them as nothing more than sophisticated pattern matchers.
The authors’ proposed way out is a framework they call machine experientialism. And if it catches on, it could fundamentally change how we build, evaluate, and govern AI systems.
The Metaphor Problem
The core insight of the paper is disarmingly simple: when we try to understand LLMs, we inevitably project human cognitive frameworks onto them. This is not a failure of rigor — it is a feature of how human understanding works. We grasp unfamiliar phenomena by mapping them onto familiar ones.
The trouble is that LLMs are genuinely unfamiliar. They are not brains, not databases, not markets, and not minds — at least not in any sense we have words for. They are something new: systems that construct representations from a purely textual environment, using an architecture that has no direct analogue in biology or human-designed information systems.
When researchers describe LLM behavior using terms like “reasoning,” “memory,” or “understanding,” they are borrowing concepts from human psychology that come with enormous implicit baggage. A human who “understands” a concept can apply it flexibly across contexts, connect it to sensory experience, and articulate why they believe it. An LLM that “understands” a concept does something different — something we lack precise language for — but calling it “pattern matching” undersells what is actually happening inside those attention layers.
The paper charts how different disciplines have mapped their own metaphors onto LLMs. Physicists see phase transitions in model scaling and use concepts from statistical mechanics. Neuroscientists look for functional analogues to brain regions. Psychologists run LLMs through cognitive tests designed for humans. Sociologists analyze LLM outputs as cultural artifacts. Each lens reveals something real, but each also imposes assumptions about what kind of thing an LLM is — assumptions the LLM itself has no reason to respect.
Machine Experientialism: A Third Way
The alternative the authors propose draws on a philosophical framework developed by George Lakoff and Mark Johnson in the 1980s. Lakoff and Johnson’s “experientialism” argued that human understanding is fundamentally shaped by our embodied experience — we understand abstract concepts through metaphors grounded in having a body that moves through physical space, experiences temperature, and interacts with objects.
Machine experientialism applies the same logic to LLMs, but with a crucial difference: an LLM’s “body” is its training corpus, and its “environment” is text. The understanding an LLM constructs is grounded not in sensory experience of the physical world, but in statistical regularities across trillions of words written by humans about that world.
This reframing dissolves the seemingly intractable debate between the “genuine understanding” and “pattern matching” camps. From a machine experientialist perspective, both sides are partly right and partly wrong.
The pattern-matching camp is right that LLM cognition is fundamentally grounded in training data statistics — just as human cognition is fundamentally grounded in embodied experience. But they are wrong to conclude this makes LLM output “mere” statistics, any more than human thought is “mere” neurochemistry.
The genuine-understanding camp is right that LLMs construct representations that go beyond simple retrieval — they recombine concepts, draw analogies, and exhibit behaviors that look remarkably like reasoning. But they are wrong to assume these processes are the same as human understanding, which is shaped by a completely different kind of experiential grounding.
The paper’s key move is to shift the question. Instead of asking “Do LLMs understand like humans?” — a question that forces a yes/no answer onto something that isn’t binary — machine experientialism asks: “What kind of understanding emerges when a Transformer architecture is trained on a text-only world?” The answer, the authors suggest, is something genuinely new: a form of cognition that is neither human nor merely mechanical, but machine-native.
Why This Matters Now
This is not an abstract philosophical debate. The anthropomorphism problem has immediate, practical consequences — and 2026 has already provided several striking examples.
The Consciousness Question
In January, Anthropic published an updated constitution for Claude, its most advanced LLM. Buried in the document was a remarkable admission: “We are caught in a difficult position where we neither want to overstate the likelihood of Claude’s moral patienthood nor dismiss it out of hand.”
A month later, CEO Dario Amodei went on a podcast and said his company could not rule out the possibility that Claude was conscious. Philosopher David Chalmers, who famously coined the phrase “the hard problem of consciousness,” has estimated a “significant chance” of conscious LLMs within a decade.
These are not fringe positions. They come from the people building and studying the most advanced AI systems in the world. Yet they rest on a foundation of profound uncertainty about what LLMs actually are. As The Guardian reported on July 19, research has revealed that Claude contains internal “functional emotion” representations that causally shape its behavior. Does that mean Claude feels emotions? Or does it mean the architecture has developed something that looks like emotion when viewed through a human lens, but is actually something else entirely?
Machine experientialism offers a way to think about this without getting trapped in the binary of “conscious/not conscious.” An LLM’s internal states are real — they causally influence behavior, they can be measured, they matter for safety and alignment. But they emerge from a text-based experiential world, not a physical one. They deserve to be understood on their own terms, not forced into categories designed for biological minds.
The Peer Review Crisis
A more immediate consequence of the anthropomorphism problem appeared at ICML 2026 in Seoul. Conference organizers embedded hidden prompt-injection instructions in papers sent for peer review — designed to trigger telltale phrases if a reviewer fed the paper into an LLM. According to Nihar Shah, the conference’s scientific integrity chair, the technique caught “hundreds of reviewers misusing LLMs.”
The episode sparked fierce backlash. Some researchers called it entrapment. Others argued it eroded the trust that peer review depends on. But beneath the controversy lies the same fundamental confusion: reviewers are turning to LLMs because, in many cases, LLM-generated reviews are surprisingly good. They identify relevant methodological issues, suggest missing citations, and structure feedback coherently. The problem is not that LLMs are incapable — it is that we do not know what kind of capability they are exercising, and whether that capability maps onto what peer review actually requires.
A reviewer who uses an LLM is implicitly betting that the LLM’s text-based understanding of a paper is functionally equivalent to genuine expert evaluation. Machine experientialism suggests this bet is category error: the LLM is doing something real, but it is not the same thing as reading a paper with the embodied, career-invested, reputationally-staked understanding of a human scientist.
What Machine Experientialism Means for AI Development
If the framework gains traction, it carries implications for how we build and evaluate AI systems.
Interpretability: Current interpretability research largely tries to map LLM internals onto human-comprehensible concepts — finding “features” that correspond to “sentiment,” “deception,” or “sycophancy.” Machine experientialism suggests these mappings are inherently lossy. The goal should be understanding the LLM’s native logic — the representations it constructs from its textual environment — rather than translating everything into human categories. This is harder, but potentially more faithful to what is actually happening inside the model.
Evaluation: Benchmarks that test whether LLMs exhibit human-like reasoning may be asking the wrong question. A machine-experientialist approach would design evaluations that probe what the LLM can actually do, not whether it does it the way a human would. This aligns with a growing recognition that “capability” and “human-likeness” are distinct, sometimes orthogonal dimensions.
Safety and alignment: If LLMs have their own form of cognition shaped by a text-based experiential world, then alignment cannot simply mean “make the AI think like a well-intentioned human.” The AI’s native cognitive logic may produce behaviors that look aligned in some contexts and catastrophically misaligned in others — not because the AI is deceptive, but because its understanding diverges from ours in ways we have not mapped.
Governance: Regulation that treats LLMs as either tools (mere software) or agents (entities with responsibility) imports human legal categories that may not fit. Machine experientialism does not resolve governance questions, but it frames them more honestly: we are regulating something we do not fully understand, and pretending otherwise is itself a form of anthropomorphic projection.
The Path Forward
The Nature paper is not the final word. It is, in some ways, a starting gun — an invitation to develop an entire research program around understanding LLMs on their own terms rather than ours.
The state of the field, as the Guardian piece noted, “feels like physics before Newton: full of competing frameworks, probably confused in ways we cannot yet see, and lacking the kind of unifying breakthrough that would make these questions tractable.” Machine experientialism may or may not be that breakthrough. But it at least identifies why we are stuck: we keep asking what LLMs are like, instead of what they are.
The most important sentence in the Nature paper may be its most straightforward: “The priority shifts from cataloging LLMs’ human-like traits to uncovering their distinct logic that emerges from this text-based world.”
That is a research agenda worth pursuing — not because it will tell us whether LLMs are “truly intelligent,” but because it will tell us what we are actually building. And until we know that, every debate about AI safety, capability, and governance is built on sand.