← All insights
AIHuman–AI Interaction

Prompting Psychology: Understanding How Humans and AI Think

AI may sound human, but it does not think or experience the world as we do. Drawing on research into language models and human judgment, this article explains how that difference shapes AI responses—and how clearer, more critical prompting can produce better results.

We often speak about artificial intelligence as though it has a human psyche. We say AI “understands,” “thinks,” or “gets confused.” We may even thank it or give it a human name.

But AI does not have a psyche in the human sense. It does not experience the conversation. What it has is a highly developed ability to predict and generate language based on patterns learned during training.

Understanding this difference is the key to better prompting.

The human psyche versus the AI “psyche”

Human thought is grounded in lived experience and shaped by emotion, memory, culture, values, and cognitive biases. AI language models, by contrast, generate responses by learning statistical relationships between tokens, not through consciousness or experience (Brown et al., 2020; Vaswani et al., 2017).

As illustrated by Quattrociocchi et al. (2025):

What appears to be an AI “personality” emerges from the model’s training, system design, conversation history, and our own human projections.

Why AI feels more human than it is

Humans are remarkably quick to attribute minds to things that communicate.

In 1966, Joseph Weizenbaum created ELIZA, a simple program that imitated a psychotherapist by turning users’ statements into questions. Although its mechanisms were basic, some users responded as though the program genuinely understood them. This became an early example of the “ELIZA effect”: our tendency to perceive understanding and emotional depth in systems that simulate conversation (Weizenbaum, 1966).

Modern AI intensifies this effect because it can produce humor, empathy, uncertainty, personality, and apparently original arguments. But the simulation of a mental state is not proof of that mental state.

  • When an AI says: “I believe this is the best approach.”
  • A more accurate interpretation would be: “Given my training, instructions, and the information in this conversation, this response is a likely and appropriate continuation.”

That does not make the answer worthless, it just means we should evaluate it differently.

What the research tells us

Here are some research-backed practices for writing better prompts.

1. Small changes can alter the answer

Language models are sensitive to wording, formatting, examples, and information order. Even rearranging multiple-choice options can affect their performance (Pezeshkpour & Hruschka, 2024).

For consequential tasks, rephrase your prompt and compare several responses before drawing conclusions.

2. Examples are better than vague adjectives

Instructions such as “make it professional” or “write creatively” are open to interpretation. Research on few-shot learning shows that models can infer a task from examples included in the prompt (Brown et al., 2020).

  • Instead of: "Make this engaging."
  • Try: "Use short opening sentences, concrete examples, and a reflective tone. Avoid exaggerated claims and corporate jargon."

When possible, also include a short example of the style you want.

3. Complex tasks benefit from decomposition

Providing intermediate reasoning examples has improved model performance on arithmetic, symbolic, and commonsense tasks (Wei et al., 2022).

For complex work, divide the task into stages, for example:

  • Identify the requirements.
  • Examine the evidence and generate options.
  • Test the options before producing the final answer.

4. More context is not always better

Models may use relevant information less reliably when it appears in the middle of a long context (Liu et al., 2023).

Select and organize the essential information, placing your most important instructions prominently rather than pasting in everything available.

5. AI can become overly agreeable

Language models sometimes mirror a user’s beliefs even when those beliefs are incorrect, a tendency known as sycophancy (Sharma et al., 2023).

Because we naturally favor information that confirms our existing views, leading prompts can reinforce our assumptions.

  • Instead of: "Don’t you agree that this strategy is the best?"
  • Ask: "Evaluate this strategy independently. Identify its strongest advantages, its most serious weaknesses, and the evidence that would change your conclusion."

If you want critical thinking, explicitly invite disagreement.

A practical framework for better prompts

Strong prompts generally contain six elements.

  1. Objective: Explain what you are ultimately trying to accomplish.
  2. Context: Provide the facts the model needs. Clearly distinguish evidence from assumptions.
  3. Audience: State who will use or read the result.
  4. Constraints: Specify boundaries such as length, tone, timeframe, acceptable sources, or prohibited approaches.
  5. Output format: Describe how the answer should be organized.
  6. Evaluation criteria: Define what a successful response must do.

The other half of prompting psychology: managing ourselves

The greatest prompting risk may not be the AI’s limitations. It may be our reaction to its output based on our bias and heuristics. Before accepting an AI-generated answer, we should reflect and ask ourselves:

  • Am I persuaded by evidence or by polished language?

  • Did the model verify this claim or merely generate a plausible explanation?

  • Did my prompt contain assumptions that the answer simply repeated?

  • Have I asked for counterarguments?

  • Would I trust this conclusion if it were expressed less confidently?

  • Which parts require an independent source or human expert?

The best relationship with AI is neither blind trust nor blanket skepticism, but a structured collaboration.

The art of prompting is how to communicate clearly with a probabilistic system, while remaining conscious of the very human psychology through which we interpret its answers.