In 399 BC, Socrates was executed for asking too many questions. His crime, loosely translated from the Athenian legal code, was "corrupting the youth"—which in practice meant teaching young people to question received wisdom rather than accept it.
Twenty-four centuries later, we have built machines that will answer literally any question you ask. And the dominant failure mode is not that the machines give bad answers. It is that we ask bad questions.
The Socratic Method, Briefly
Socrates did not lecture. He asked sequential questions designed to expose contradictions in his interlocutor's thinking, gradually guiding them toward a more coherent understanding. The method has three core moves:
- Elicit a claim. What do you believe to be true?
- Probe the foundations. Why do you believe this? What assumptions support it?
- Test with counterexamples. Does this belief hold in all cases, or does it break?
The genius of the method is that it creates understanding through productive discomfort. You do not learn when you are told the answer. You learn when you discover that your current understanding is insufficient.
Why Most Prompts Fail the Socrates Test
The average AI prompt is an assertion disguised as a question. "Write me a marketing email for our new product." This is not a question. It is a command that presupposes dozens of unexamined assumptions:
- Who is the audience?
- What do they currently believe about this product category?
- What is the single most important thing they need to understand?
- What action should they take, and why would they take it?
- What tone matches our brand and their expectations?
Socrates would never accept such a prompt. He would ask: "Before we write this email, can you tell me what problem your customer has on Tuesday morning that your product solves by Tuesday afternoon?" And then he would keep asking until the answer was specific enough to be useful.
The people who get extraordinary results from AI are not better at writing prompts. They are better at questioning their own assumptions before the prompt is written.
The Five Socratic Questions for AI Work
Adapted from the Socratic elenchus for modern AI interactions:
1. "What do I actually need to know?"
Before you type a prompt, ask yourself what output would make you say "this is exactly what I needed." If you cannot describe the ideal output with precision, you are not ready to prompt.
This is Socrates' first move: elicit the claim. What do you believe you want?
Most people discover, when they try to answer this honestly, that they do not want what they thought they wanted. "Write a blog post about AI" becomes "help me explain to technical leaders why retrieval-augmented generation outperforms fine-tuning for enterprise knowledge bases, using a specific cost comparison."
2. "What assumptions am I encoding?"
Every prompt contains implicit assumptions. "Summarize this document" assumes the document is worth summarizing, that a summary is the right format, and that the default summary length is appropriate.
Socrates' second move: examine the foundations. What are you taking for granted?
The practical technique is to read your prompt and list every assumption it makes. Then decide which assumptions are load-bearing (must be true for the output to be useful) and which are incidental (could be different without affecting quality). Make the load-bearing assumptions explicit in your prompt.
3. "What would make this answer wrong?"
This is the counterexample move, and it is the most powerful of the five.
Before you accept an AI's output, ask: "Under what circumstances would this response lead me to a bad decision?" This is not about catching hallucinations (though it helps with that). It is about understanding the boundaries of the answer's validity.
A financial model generated by AI might be technically correct but assume a growth rate that is unrealistic for your market. A code suggestion might work but introduce a dependency that conflicts with your security requirements. A strategic recommendation might be sound in theory but assume resources you do not have.
Socrates would say: the unexamined answer is not worth using.
4. "What question should I have asked instead?"
This is where most people stop, and where the best practitioners start.
After receiving an AI response, explicitly ask the model: "Given what you now understand about my problem, what question should I have asked you instead?" This meta-question often produces more valuable output than the original prompt, because the model can now identify gaps in your framing that you could not see.
This is the Socratic dialogue in action: the back-and-forth process of refinement that moves both parties toward better understanding.
5. "How would I know if I'm wrong?"
The final Socratic question is about falsifiability. Socrates insisted that any belief worth holding must be capable of being proven wrong. If you cannot imagine evidence that would change your mind, you are not thinking—you are defending.
Applied to AI: after you have used a model's output to make a decision, define what evidence would tell you the decision was wrong. Then look for that evidence. This transforms AI from a confirmation machine (which is its default mode—it will happily reinforce whatever framing you provide) into a genuine thinking tool.
The Dialogue Pattern
The most effective AI interaction pattern is not prompt → response. It is dialogue:
Round 1: State your problem. Ask the model to identify what information is missing before it can help effectively.
Round 2: Provide the missing information. Ask the model to propose an approach and explain its assumptions.
Round 3: Challenge the assumptions. Ask the model to steelman the opposite approach.
Round 4: Synthesize. Ask the model to integrate the best elements of both approaches into a recommendation, noting where uncertainty remains.
This four-round dialogue consistently outperforms single-shot prompts, even very elaborate ones, because it creates the same productive discomfort that Socrates used: each round forces re-examination of what seemed settled.
Why This Matters Now
We are at a peculiar moment in history. For the first time, the average person has access to a conversational partner with encyclopedic knowledge and infinite patience. And yet most people use this partner as a vending machine: insert prompt, receive output, move on.
Socrates would find this tragic. He spent his life arguing that the unexamined life is not worth living. The AI corollary is that the unexamined prompt is not worth sending.
The organizations that will extract the most value from AI in the coming years will not be the ones with the most sophisticated models or the largest token budgets. They will be the ones that teach their people to ask better questions—to approach AI with the same disciplined curiosity that Socrates brought to the agora.
The good news: unlike Socrates, you will not be executed for asking too many questions. The only cost is a few extra tokens. And the returns compound with every conversation.
Start with the first question: "What do I actually need to know?" Everything else follows from there.