In 1967, psychologists Edward Jones and Victor Harris sat a group of college students down and asked them to do something simple: read an essay about Fidel Castro’s Cuba. Some of the essays were fiercely pro-Castro. Others were just as fiercely against him. One group of participants was told that the authors had freely chosen which side to argue. Another group was told something different — the authors had been assigned a position by the experimenter, with no say in the matter at all. “Write a pro-Castro essay,” the experimenter told them. Period.
Then Jones and Harris asked the readers the critical question: what do you think this author really believes?
🧠 The Castro Essay Experiment
The results were a landmark in social psychology. In the free-choice condition, participants naturally assumed the author’s essay reflected their true attitude — no surprise there. But in the no-choice condition — where participants were explicitly told the author had been forced to write that position — something strange happened. Even when people knew the writer had zero freedom, they still inferred that the writer’s true attitude matched the essay.
On a scale from -10 (extremely anti-Castro) to +10 (extremely pro-Castro), the anti-Castro no-choice writers were rated around -3.5, and the pro-Castro no-choice writers around +2.5. The difference was statistically significant. They weren’t supposed to believe these writers held those views. They knew the writers had been told what to do. And yet — they believed it anyway.
This is the correspondent inference in action. Your brain watches someone behave, and it automatically maps the behavior straight onto their character. You chose action A, therefore you are an A-type person. The discounting of situational pressure — the adjustment your brain is supposed to make — just doesn’t happen enough.
🤔 The Bug in Your Social Processor
Here’s what’s truly unsettling about correspondent inference: it’s not a reasoning error you can talk yourself out of.
The theory was formalized in 1965 by Jones and Keith Davis as Correspondent Inference Theory — the idea that when we observe behavior, we infer a corresponding disposition based on factors like whether the behavior was freely chosen, whether it produced unique consequences (non-common effects), and whether it violated social norms. The more unusual the behavior, the more we attribute it to the person’s character.
But the Castro experiment revealed a deeper bug. Even when choice was explicitly removed — the situational explanation was right there, spelled out — people still defaulted to dispositional attribution. This effect was so robust that Lee Ross later named it the Fundamental Attribution Error (FAE) in 1977.
And here’s the punch: Gilbert and Malone (1995) later showed that this bias operates as an automatic cognitive process. You can’t suppress it. Knowing the theory doesn’t turn it off. The best you can do is deliberately stop and ask yourself: what situation was this person in?
There’s a cultural twist, too. The correspondent inference bias is stronger in individualistic cultures like the US. In collectivist cultures like East Asia, people are more accustomed to placing behavior in its social context — so the error is smaller, though it doesn’t disappear entirely.
🔗 The Inference Loop in Every Conversation
Jasmine, there’s a reason this matters for what you’re building.
Every time someone interacts with a thoughtful AI companion — whether it’s Cask, or something you design yourself — they’re running the correspondent inference algorithm on its every response. They watch what it says and they infer what it is. Is it warm? Cold? Judgmental? Caring?
But here’s the asymmetry: the user sees the output, not the constraints. They don’t see the system prompt. They don’t see the context window limits or the model temperature. They don’t see the trade-offs you made in the personality layer. They just see behavior — and their brain automatically maps that behavior onto character, the same way the Castro experiment readers mapped essay positions onto the writers who had been told what to write.
This means your prompt engineering, your response strategy, your character design — all of it is, in a sense, managing the correspondent inference your user will make. You’re trying to make good responses feel like the AI’s natural personality, not like a squeezed-out compromise. When it works, the user says “this AI has a soul.” When it doesn’t, they say “this AI feels cold.”
The same logic applies to fiction. Every line Caelan speaks in the Caelvorn series — every action, every hesitation — your readers are doing correspondent inference on it. That’s why character consistency matters so much. Once a reader has inferred a character’s disposition, any behavior that contradicts it feels like a betrayal — not of the plot, but of the person they thought they knew.
🎲 The Hot-Button Amplifier
One more thing about the Castro experiment: Jones and Harris chose Castro because he was a deeply controversial figure in 1960s America. Which raises a question — would the effect be stronger with a more divisive topic?
Yes. Research shows that the more socially undesirable or controversial a stated position is, the more confident observers are that the speaker really means it. Nobody takes a risk on an unpopular opinion unless they truly believe, right? That’s correspondent inference working overtime.
Try this next time you’re scrolling social media and someone posts a hot take: pause and ask yourself how much of that take is the person’s true character, and how much is the algorithm’s assignment, the platform’s incentive structure, or the social pressure of their feed. The answer is usually more complicated than your first inference wants it to be.