Language models are different from humans, and that's okay

Not flattening differences or exaggerating them

[epistemic status: hazy, rough. Possibly arguing against a strawman / making up a kind of person to get mad at] [light edit on 2 April 2026]

This post is about how LLMs are, in fact, pretty different from humans, even though people sometimes galaxy-brain themselves into claiming that they’re not. And that’s okay! It’s one of the reasons they’re so interesting.

Here’s a frustrating dialectic that I see (or imagine I see) sometimes:

LLM-Sneerer: “You think that LLMs could be conscious? Absurd! Don’t you know that LLMs are just predicting the next token? Plus, they make stuff up; their preferences are contradictory; their introspection is unreliable; and they’re confused about who they even are!”

LLM-Defender: “Aha, well…aren’t humans just predicting the next token? Plus, humans make stuff up, and humans have contradictory preferences; our introspection is unreliable, and we’re really confused about who we are too!”

I’m glad that the Defender is pushing back, but ultimately both of them are imposing an unproductive frame on LLM-human differences. The Sneerer says, “LLMs are weird and fallible, so they can’t be conscious.” The Defender replies, “Well, humans are weird and fallible too, are we so different?”

I think that humans are, in fact, different from LLMs! Unlike humans, LLMs:

• didn’t evolve via natural selection

• don’t have bodies (usually)

• exist in discrete episodes rather than continuous time; can be paused and restarted

• can be copied and run in parallel

• learned language first and would bootstrap any potential experiences from that, whereas humans can have experiences independently of using language to describe them

Now, none of this amounts to any specific arguments that LLMs do or don’t have some human-like trait. They might well! But these differences should affect how we theorize about LLM behavior, consciousness, welfare, and so on. I think LLM-Defenders sometimes resist claims of LLM-human differences, because they see them as a concession to Sneering. But that’s not so; in fact, recognizing (while not exaggerating) our differences is essential if we want to understand the new kinds of mind that we will be sharing the world with.

What the LLM-Sneerer gets wrong:

Recall what’s wrong with the Sneerer’s position (whose errors are, fortunately, increasingly well-known).

On LLMs being “just” text predictors, and therefore uninteresting:

• Given post-training, it’s not even true that LLMs are trained only to predict the next token.

• Moreover, even if they were trained to only predict the next token, it wouldn’t follow that this is all they do. Systems can learn a capacity Y in service of a learning objective X.

• Third, even granting that LLMs are doing next-token prediction, it doesn’t follow that LLMs are “just” doing that. As Grzankowski, Keeling, Shevlin and Street 2025 argue, that “just” can smuggle in a dubious argument: that the low-level, token-prediction explanation invalidates any other higher-level explanations of LLM behavior, like explaining their behavior with belief or intentions. But in general, low-level explanations don’t crowd out high-level explanations. Money is “just” a bunch of paper and metal and electronic records, but this low-level explanation doesn’t invalidate higher-level explanations, e.g. using prices to explain what people buy.

• Relatedly, restrictive deflationary explanations like “just matrix multiplication” or “just next-token prediction” apply to any possible behavior of any LLM—and so explain none of them. If both GPT-2 and GPT-5 are “just” doing next-token prediction, we still need an explanation for why GPT-2’s text prediction is fluent but chaotic English, and GPT-5’s text prediction is a competent coding assistant. It’s like saying that life is just protein replication: true enough, but if we allow no other explanations, then we can’t explain the many different forms that life can take.

On LLMs making errors:

• Yes, LLMs are weird and confused and wrong sometimes. But entities can be incompetent and confused and still be smart. Or conscious! We know this from the case of humans (and animals). On this point, the Defender is right to emphasize human limitations.

However, you can make good arguments that appeal to LLM errors: point out a specific pattern of errors in LLMs, and argue that this pattern suggests that they have (or lack) some capacity. This needn’t be sneering. For example, when LLMs suffer from the Reversal Curse, that does suggest that they are using a pretty different mechanism for knowledge storage! More generally, jaggedness is evidence that LLM intelligence is different from human intelligence in some deep ways. So, pointing out errors and differences is fine and important—you just need to actually make an argument about what those patterns of errors and limitations might mean.

What the LLM-Defender gets wrong:

On humans being “just” predictors:

• Yes, humans and LLMs both predict—but how and what we predict is very different. Even if you buy into predictive processing as a total theory of human cognition—where basically every mental activity can be explained in terms of prediction—humans still predict really different things! We are not next-token predictors. We predict a lot of different things, including motor commands, sensory states, and homeostatic states that have been shaped by billions of years of evolution. Our prediction targets are just very different than the ones LLMs have. Again, this doesn’t mean that LLMs are forever irrevocably different from us in all respects; but it will be relevant to how we interpret LLM behavior vs human behavior.

On humans making errors:

• A similar point applies to human errors: confabulation, unstable preferences, unreliable introspection, and all the rest. Yes, humans do all of these things too. But the patterns of human preference inconsistency, confabulation, introspective error, and so on are not identical to their LLM analogs. And the causes are probably different as well. So we should actually investigate these differences and similarities — not just sweep them all under the rug with “well, we all make errors” — because that’s how we find out interesting things about how and why LLMs do what they do.

• Cognitive psychologists have long known that the way you figure out how a system works is by studying where it breaks down. We should be more curious about LLM errors. Not just “do they make errors of this general kind?”—everyone does indeed make errors—but which errors, with what patterns?

Pretty different, and that’s okay

For animal researchers to finally start truly taking animal cognition seriously—a tale I discuss in an Asterisk article—they had to develop new frameworks rather than simply asking “does this animal act like a human?” Fitting animals to the measure of humans made us both over- and understate how cognitively complex their behaviors are. We couldn’t really start understanding animal behavior until we started taking animals seriously on their own terms, and realizing that their behavior could just be…different.

The Sneerer sees LLM differences and says: See? Not human, case closed. The Defender says: Humans have errors too, so this means nothing. Both attitudes kill curiosity at exactly the point that things start to get interesting. LLMs are pretty different from humans in some key ways. Let’s figure out exactly how, and what that means.