The following was created with the research assistance of ChatGPT
A1 as an Intellectual Tool: A Proposal for Two Kinds of Artificial Intelligence
The debate over artificial intelligence in education often begins with the wrong question, i.e., “should students use AI?”
That question already is becoming obsolete. Artificial intelligence exists. It is readily available, extraordinarily useful, and almost certainly will become more capable and more ubiquitous. Asking students and serious thinkers simply not to use it is like asking a carpenter to build a house with only a handsaw and hammer after power saws and nail guns have been invented.
The important question is not whether we should use AI in education, but rather, how we should use it.
That distinction matters because AI has two almost opposite capabilities. It can relieve us of the necessity to think, or it can cause us to think more deeply than we otherwise would. The difference may determine whether AI ultimately weakens education or revolutionizes it.
The danger is outsourcing thought
Suppose a student is assigned a thousand-word paper about the causes of the Civil War. He types: “Write me a thousand-word paper about the causes of the Civil War.” Within seconds, AI produces a competent paper. The student submits it.
The student has used an extraordinary intellectual tool and accomplished virtually nothing intellectually. The problem is not that AI was involved. The problem is that the student transferred the very activity the assignment was intended to exercise—researching, comparing, questioning, organizing and reasoning—from his own brain to the machine.
Now consider a different student. He begins with an idea about the causes of the Civil War and asks the AI: “Here is what I think. What is wrong with my argument?”
The AI identifies the student’s assumptions and asks the student relevant questions. The student defends his assumptions and answers the questions. The AI produces contrary evidence. The student modifies his position and asks the AI for additional historical facts. The AI finds them and identifies their sources. The student proposes an analogy. The AI finds a case in which the analogy fails.
The student revises his argument again. After an hour, or a month, or more of repeated questioning and responding, the student and AI have produced a much more sophisticated argument than either the student’s original conception or the generic essay the AI could have produced in seconds.
Both students used AI. The first student used AI instead of thinking. The second used AI in order to think more. That may be the central distinction upon which AI education should be built.
What is AI fundamentally good for?
Strip away the impressive prose, instant summaries, formatting, illustrations and other conveniences—the chrome wheels on the car—and AI has two extraordinary intellectual capabilities.
First, it can find facts. AI potentially gives an ordinary person access to something approaching a gigantic research library, together with an assistant capable of searching that library in seconds. A good research AI should not merely report information. It should distinguish among established facts, disputed claims, interpretations and speculation, and reveal where the information came from.
Second, it can challenge thought. This may ultimately be the more important capability: Applying the Socratic method. It works like this:
A user proposes an idea.
The AI asks a relevant question, perhaps, “Why? or “How?”
The student answers.
The AI asks: How do you know?
The student supplies evidence.
The AI asks: Can you think of any exceptions?
The student modifies the proposition.
The AI asks: What assumption are you making?
Eventually the proposition survives, fails, or evolves into something better. This is essentially the Socratic method augmented by an enormous factual database. AI does not originate the idea. It can help a human being discover one by supplying facts and intellectual resistance.
The danger of the overly agreeable AI
An AI designed primarily to please its user can become an intellectual liability. Here’s how that plays out:
The user proposes “A.” The AI congratulates him and supplies arguments supporting “A.” The user builds “B” upon “A.” The AI strengthens “B.”
Soon the two have constructed an impressive structure without sufficiently examining the foundation or alternatives. The idea might really be a dud, and had the user been challenged, he would have realized it. Instead, he was encouraged to submit a dead-end, ill-considered concept.
No one enjoys constantly being told he is wrong. But for serious intellectual work, disagreement is not a defect. It is part of the machinery. A useful intellectual AI must be willing to say: “That doesn’t follow.” “Your evidence doesn’t establish that.” “You’ve contradicted your earlier premise.” “Here is an example in which your rule fails.” “What evidence would persuade you that you are wrong?”
The AI’s objective should not necessarily be to support the user’s initial belief. It should support the user’s goal. If the user’s goal is to discover whether an idea is true, attempting to disprove that idea may be far more helpful than agreeing with it.
Thinking requires resistance. AI can remove that resistance, or it can provide it.
THE THREE-AI MODEL
This suggests that we may need two fundamentally different kinds of AI:
AI 1: The Companion, the general-purpose assistant most people already imagine.
It converses. It explains. It drafts letters. It summarizes documents. It helps plan vacations. It answers questions. It brainstorms. It entertains. It may serve as a sounding board or even something resembling a knowledgeable friend.
Its primary objective is helpfulness. The AI should correct important errors, of course, but it need not turn every conversation into an oral examination. Sometimes a person simply wants to know how something works.
AI 2: The Scholar
This AI has a different mission. Its purpose is not primarily to make intellectual work easier, but rather to make intellectual work better. It combines deep factual research with Socratic challenge.
It asks for evidence. It searches for contrary evidence. It identifies hidden assumptions. It tests definitions. It looks for exceptions. It distinguishes fact from opinion and evidence from assertion. It asks the user to defend conclusions.
Most importantly, it does not rush to do intellectual work that the user should be doing.
In the Socratic dialog method, when a student asks a question or makes a statement, AI 2 oftentimes will respond with a question. For example, the student asks:
“Why did Rome fall?” Instead of immediately supplying seven causes, AI 2 might ask: “What do you mean by ‘Rome fell’? The city of Rome? The Western Roman Empire? Roman civilization?”
The student has learned something before the AI has supplied a single answer: the question itself contained an assumption. Further, the AI is suggesting, with its question, further avenues for investigation. That is education.
AI 2 in the classroom
For schools, AI 2 could be supplied as an educational tool rather than merely allowing students unrestricted access to a general-purpose answer machine.
Its objective would be to maximize the student’s intellectual participation. The sequence might be:
The teacher begins with a question → the student thinks and offers an answer → AI questions the answer→ the student responds → AI supplies relevant facts → the student revises → AI searches for exceptions → the student defends the revised conclusion.
Contrast that with: Question → AI answers → student submits.
The first creates learning. The second creates a document. Education never should assume those are the same thing.
AI 3 The teachers’ Aid and Grading Assistant.
We propose creating a teacher’s version of AI, a version 3.
AI has exposed a weakness that existed long before AI: teachers often grade the answers because it is difficult to observe the thinking that produced it. A term paper traditionally serves as a proxy for intellectual work. Producing a good paper usually requires doing much of the research and thinking oneself.
AI has broken that connection. A beautiful paper no longer proves that the student understands the subject. The solution should not be an arms race between students trying to conceal AI use and teachers using unreliable programs to guess whether AI wrote something.
Instead, AI 3 itself can help evaluate understanding.
Give the students’ papers to the teacher’s version of AI, the version 3.
It would identify passages unusually similar to published or commonly generated material. It would flag unsupported factual claims. It would identify arguments that deserve explanation. It would compare the student’s work with the kind of generic answer AI itself would produce.
But it should not pronounce: “This paper is 83 percent AI-generated.” Such numerical precision implies knowledge the system may not possess.
Instead, it should tell the teacher: “Ask the student why he reached this conclusion.” “Ask him to explain the assumption underlying paragraph six.” “Ask him for an argument against his own position.” “This passage closely resembles previously published wording; examine the source.”
If a student submits an extraordinary essay but cannot explain its central argument, the essay has told us very little about what he learned. If he can defend it, modify it under questioning, recognize weaknesses and explain why he rejected alternatives, then considerable intellectual development has occurred regardless of whether AI helped polish his prose.
Perhaps substantial assignments eventually should receive two evaluations: Quality of the work. Understanding of the work.
Originality and the Socratic AI
There remains an interesting question about originality. Human beings sometimes produce ideas they believe to be original, although no person can know with certainty how much a new thought owes to forgotten books, conversations and experiences.
AI presents an even more difficult philosophical question. It has been trained upon enormous quantities of human-created material. Can something produced from that history ever constitute an original thought?
Perhaps that question is less important educationally than it initially appears. A Socratic AI does not need to originate the student’s breakthrough. It needs to ask the questions that led the student to understanding. “Why?” “Why do you believe that?” “What would prove you wrong?” “Does your rule apply here?” “Why not?”
Eventually further questioning ceases to illuminate. We reach an assumption, an unknown, or the current limits of evidence. A good Socratic system must recognize when continued questioning has ceased to clarify and has begun merely to obstruct. Then it moves on.
The objective is not endless argument, but rather deeper understanding.
The power-tool analogy
The fear that AI will weaken people’s ability to think is legitimate. But refusing to use AI is not the solution. Power tools can allow a skilled carpenter to build things he otherwise could not build. They also allow an incompetent person to make mistakes much faster.
The important question is not whether electricity touched the saw. It is who designed the house and who decided where to make the cuts.
AI is an intellectual power tool with one unprecedented characteristic: unlike a power saw, it can decide what kind of house to build. That is why education must teach students not merely how to operate AI, but when not to surrender the decision to it.
The AI should respond as though the student said: “Don’t help me prove that I’m right. Help me discover whether and how I’m wrong.” That may be one of the most valuable AI skills educations can teach.
A different future for education
Much of today’s debate concerns detecting AI, prohibiting AI and preventing students from using AI to cheat. Those problems are real, but they are transitional problems. The permanent question is “What intellectual work should remain the responsibility of the student?”
The answer cannot be “everything,” any more than education responded to the calculator by requiring every engineer forever to perform long division by hand. Nor can the answer be “nothing.” Just as we don’t object to the student using books for factual reference, we should not object to their use of AI in the same way.
The purpose of education is not to manufacture assignments. It is to develop minds capable of examining evidence, recognizing assumptions, discovering relationships, imagining possibilities and changing conclusions when the evidence demands it.
AI can interfere with, or support, that process.
Properly designed, it also may become one of the most powerful instruments ever created for advancing education. For most of history, only a fortunate few had regular access to a brilliant teacher or intellectual companion willing to spend hours examining a single half-formed idea.
AI potentially can give that opportunity to every student, not as an oracle, a ghostwriter, or an agreeable buddy that congratulates every idea. Rather, AI can be a tireless combination of research librarian and Socratic opponent—one whose job is to provide the two things serious thinking continually requires: Facts and resistance.
If we design educational AI around those objectives, the question will no longer be whether AI prevents students from thinking. The question may become how previous generations managed to learn without it.
Rodger Malcolm Mitchell
#Monetary Sovereignty - Mitchell 