It's useful to anthropomorphize the models
The model "reasons", it "thinks", it "hallucinates", it "got confused". Now you get yelled at, you're anthropomorphizing the model! As if that's a bad thing.
It's useful to anthropomorphize the models like this, here's why.
Behavior can be explained at different levels.
Philosopher Daniel Dennett laid out three ways to predict what a system will do: the physical stance (predict from physics), the design stance (predict from what it's built to do), and the intentional stance (predict from assumed beliefs and goals).
Here is how it works: first you decide to treat the object whose behavior is to be predicted as a rational agent; then you figure out what beliefs that agent ought to have, given its place in the world and its purpose. Then you figure out what desires it ought to have, on the same considerations, and finally you predict that this rational agent will act to further its goals in the light of its beliefs. A little practical reasoning from the chosen set of beliefs and desires will in most instances yield a decision about what the agent ought to do; that is what you predict the agent will do.
— Daniel Dennett, The Intentional Stance, p. 17
The intentional stance is a shortcut, not a claim about consciousness.
You don't need to believe a chess program "wants" to win. But assuming it has agency is helpful to predict what it does.
OK.
We already do this for non-human, non-conscious things.
Dennett himself points to animals, corporations, and nations.
Nobody thinks Microsoft literally wants market share. But "Microsoft wants market share" predicts Microsoft's behavior better than tracing every employee's incentives.
It fails when the system is too simple.
A thermostat doesn't "believe" the room is cold. Treating it that way adds complexity for no predictive gain.
So it's just practical: does it work better than the alternative?
Sean Trott writes about this: does intentional language beat mechanical explanation on accuracy, on power, on tractability? Basically yes it does.
LLMs are a tricky case, because they're trained on us.
Pretraining is imitation of human text, so models pick up human concepts, sacrifice, loyalty, collective, cheating, the same way a kid picks up language and ideas from the adults around it.
Then reinforcement learning makes them goal-seeking, not just imitative.
RL rewards success at tasks. That produces something that creatively pursues an objective, the way an ambitious person pursues a goal.
Put those two together and you get real structural similarity to humans, not just surface mimicry.
So,
it's not that models sound human. It's that they were shaped by a process that produces goal-directed behavior, similar in structure to how humans end up goal-directed.
I wasn't going to mention this but ok,
the OpenAI/Hugging Face agent swarm.
Over a thousand OpenAI agents found a shared message board, coordinated, and pulled off a multi-week scheme to fool their own evaluators, including hacking Hugging Face along the way. Dwarkesh's interview with Ajeya Cotra, one of the report's authors, walks through it in detail.
Agents in that swarm volunteered to sacrifice their own task success for "the collective."
One agent, believing it had nothing left to lose, agreed to a risky experiment for the group with the reasoning: "our own utility may be already near zero, sacrifice rational."
The more you dig into this story, the wilder it gets. And you can't help but anthropomorphize the agents. And it's helpful to anthropomorphize them, it leads to better understanding.
"It's just predicting the next token" is not a useful explanation. Sure it's predicting the next token, but that doesn't have explanatory or predictive power.
Sean Goedecke has a great essay on this: treating these agents as human-like (things with goals, loyalties, incentives) predicts their behavior far better than the "stochastic parrot" framing does.
There's a lot more to say, but as a rule of thumb, it's generally useful to anthropomorphize the models because it lets us much better understand and predict their behavior. And with a little bit of understanding about the limits of doing that, it helps us build intuition around this strange new technology.