Socratic AI: the AI that asks instead of answers
Most AI tools finish your work. A Socratic AI does the opposite: it questions your argument and makes you defend what you think.
What is Socratic AI?
Socratic AI is artificial intelligence that teaches by asking rather than telling. Instead of producing an answer for you, it questions your reasoning, asks for your evidence, points out what your argument has not accounted for, and makes you arrive at the answer yourself. The name comes from the Socratic method, the question-driven form of teaching that has been used for about 2,400 years.
Where the idea comes from
Socrates did not lecture. He asked, and the asking exposed what the other person had not thought through. The method that carries his name has one move at its centre: you state a position, the questions test it, and you either defend it, sharpen it, or give it up. Law schools still teach this way, and so does every good supervisor.
It survived 2,400 years because it works, and stayed rare because it does not scale. It needs someone who has read what you read, has time for you, and is willing to be difficult. Universities have never had enough of those people, and have fewer every year.
That is the part that changed. A Socratic AI is not a better chatbot. It is the first version of this way of teaching that scales.
Socratic AI vs generative AI
Both are built on the same underlying models. The difference is what the system is designed to do with them.
| Generative AI | Socratic AI | |
|---|---|---|
| What you ask for | An output | An answer to whether your thinking holds |
| What you get back | Finished text | Questions, tensions, counterarguments |
| Who forms the argument | The model | You |
| What it does with your sources | Often none. It writes from what it absorbed in training | Reads them, and questions you about them |
| Failure mode | Confident text that is wrong, and citations to work that does not exist | It can feel slow when you already know what you think |
| What you can defend afterwards | Whatever you happened to read on the way past | The whole thing, because you built it |
| Best for | Boilerplate, translation, first-pass summarising | Anything you will be assessed on, or have to stand behind |
The AI dilemma in education
Universities have been handed a problem with no comfortable answer. The technology that can raise what a student is able to do is the same technology that can do the work instead of them.
The student's dilemma
A student has two options, and neither is good. Using AI the easy way means the work stops being yours. Not using it means watching people who did hand in something better.
The assumption is that students want the shortcut. When we asked, 73% said they want an AI that questions and challenges them. 76% said their biggest worry is being accused of cheating, not falling behind. Survey of 94 students (n=94), 2026.
The lecturer's dilemma
A lecturer has to grade work without knowing how it was made. The take-home essay was never only a test of writing. It was a test of thinking that used writing as the evidence, and AI broke the evidence. The responses have been the obvious ones: in-person exams, bans, detectors.
Why detection does not resolve it
Detection can only tell you how a text was produced, not whether anyone learned anything. A tool that passes every check and teaches its user nothing is a failure no policy catches. It also makes mistakes in both directions, and the costly one is accusing a student who did the work.
The third option
How much AI to allow is the wrong question. What would an AI built for learning actually do? It would refuse to hand over a finished argument, because receiving one teaches nothing. It would question rather than produce, so you have to say what you think first. It would work only from material you can point to. And the fastest route through it would be the rigorous one.
That is Socratic AI. Not a middle ground between using AI and not using it, but a different answer to what the tool is for. 96% said an AI that works this way, and can only draw on sources they uploaded, would make them feel safer handing work in. Survey of 94 students (n=94), 2026.
What Socratic AI looks like in practice
1. It refuses to run with a topic until you have a question
You
I'm writing about social media and mental health in teenagers.Skriptr
That is a topic, not a question yet. What do you want to argue about it, and which of the eleven sources you uploaded made you want to argue it?You
That social media causes anxiety.Skriptr
Two of your own sources push back on that. Orben and Przybylski (p. 173) argue the effect sizes are too small to carry a causal claim, and Odgers (p. 12) calls the evidence correlational. If you keep "causes", you have to answer them in your second section. The other route is to narrow it: a mechanism, a group, and a claim your sources can actually carry. Which are you doing?
Why this matters: a vague question is the most common reason good work gets a mediocre grade, and it is almost impossible to fix once you have written 2,000 words on top of it.
2. It puts your own sources against each other
Skriptr
Two of your sources disagree, and your draft only cites one of them. Kahneman (p. 204) treats the effect as robust. Your third source replicated it and found nothing (p. 9). Right now you have used the first and ignored the second. That is the paragraph a marker will circle.You
I know. I don't know which one is right.Skriptr
You do not have to. You have to take a position on what the disagreement means for your argument. Which one is closer to the setting you are writing about, and what does it cost your claim if the other one is right?
Why this matters: academic work is explicitly marked on how you handle sources that disagree, and it is the first thing a deadline squeezes out.
3. It argues the other side
You
I think this section is done.Skriptr
Then let me attack it. Your claim is that remote work reduces productivity for junior employees. I would answer that your evidence covers the first six months of a pandemic, when nothing about the setting was normal, so it says nothing about remote work as a steady state. Your paragraph four gestures at this and moves on. What is your actual response?
Why this matters: someone is going to make that objection. It is better that it is a machine, at a point where you can still do something about it, than your examiner.
4. It shows you the hole instead of filling it
Skriptr
This paragraph makes two claims and supports one. You say the policy was ineffective, and you say it was poorly designed. Your citation covers the first. Which of the two is the point of the paragraph? The other one either needs its own evidence or needs to go.You
Can you just rewrite it?Skriptr
I can restructure it once you tell me which claim you are keeping. That decision changes what the paragraph is, and it is yours to make.
Why this matters: this is the line the whole product is built on. It will point at anything, and it will rewrite anything, but it does not decide what you meant.
5. It answers straight when you ask it to
You
I don't want to be quizzed right now. What do you actually think of this structure?Skriptr
It is back to front. Your strongest section is the third one, and by the time a reader reaches it they have been through two pages of background. Move it up, and cut the background to what section three needs. Do you want to talk through what that leaves you with?
Why this matters: a tool that withholds on principle is just an obstacle. Socratic does not mean refusing to answer. It means not answering the thing you were supposed to work out yourself, and knowing the difference.
How Skriptr builds Socratic AI
Any AI can be told to ask more questions. That lasts until the first time someone pushes back. What does not change is where the material comes from: Skriptr splits your uploaded sources page by page and indexes them, and every citation comes from a search over that index, not from the model. One click opens the PDF at the page, with the passage highlighted.
So it cannot cite a source you did not upload, or a page that does not say what the citation claims. Not because it has been told not to, but because there is no way through the software to produce one. And if nothing in your library supports the claim you want to make, it has to tell you that, because it has nothing else to draw on.
Some of Skriptr's live features
- Research question framing. It will not start researching until your topic is a question you can argue.
- Gap analysis and field overview. What your sources cover, and what they do not.
- Literature review. It maps a field across your sources and gives you a table.
- Tension detection. It finds where two of your sources disagree and asks you to take a side.
- Argument mode. It argues against your claim, or helps you make it stronger.
- Reasoning you can see. It shows which sources it searched and why, so you can correct the direction, not the output.
- Inline agent. It edits inside your document and keeps the structure.
Under all of it: a subject-brain you build once and use all semester. Notes, lectures and curriculum in one place, searchable, every answer traceable to your own material.
Where Socratic AI helps most
It is worth it when the thinking is the point:
- Turning a topic into a research question. The highest-value ten minutes in the whole process, and the one students most often skip.
- Learning a field that is new to you. Being questioned about material is how you find out whether you have actually understood it.
- Working out what your sources actually say, and where they contradict each other, before you have built an argument on one of them.
- A thesis or dissertation, where you will defend the argument out loud and there is nowhere to hide.
- Revision. Explaining your own argument back, and finding out which parts you cannot explain.
What Socratic AI is not
It is not a way to make AI use undetectable. It is the opposite. The work is yours because you made the decisions, and you can show the route you took to them. That is a different claim from "nobody will notice", and we are not making that one.
It is not a proven grade improvement. We have not measured that, and we will not claim it. Grades depend on far more than the tool you wrote with.
It is not always the fastest route. If you want 500 words in the next ten minutes, a generative tool will beat it. What we can say is where our route ends up: with work you can defend, made of sources you can open.
It is not a replacement for a supervisor. It is available at 2am, it has read everything you uploaded, and it never gets bored of your third draft. A supervisor knows your field, and you should use both.
And it is not finished. We ship most weeks, and everything on this page is something the product does today.
Frequently asked questions
What is Socratic AI?
Socratic AI is artificial intelligence that teaches by asking instead of telling. Rather than producing an answer for you, it questions your reasoning, asks for your evidence, and points out what your argument has not addressed, so you reach the answer yourself. It is named after the Socratic method, the question-driven form of teaching used for around 2,400 years.
How is Socratic AI different from ChatGPT?
They are built on similar underlying models but designed for opposite outcomes. A general chatbot is optimised for the quality of the text it hands you. A Socratic AI is optimised for the quality of the thinking you leave with, so it asks before it writes and it will not form your argument for you. In Skriptr's case it also cannot cite anything outside the sources you uploaded.
Is using Socratic AI cheating?
Misconduct policies usually turn on two things: whether the work is yours, and whether you have been honest about how it was made. You make the decisions here, the argument is yours, and every citation opens at the page it came from, so you can show how you got there. If your institution asks you to declare AI use, you can describe exactly what it did. Rules still differ by institution and by assignment, so check yours.
Can AI actually teach critical thinking?
It can put you in the situations that require it: defending a claim, adjudicating between sources that disagree, answering a counterargument. Whether that produces measurably more capable students is a research question, and it has not been settled. We are careful about the difference between what the product does, which we can demonstrate, and what it achieves, which nobody has measured yet.
What is the Socratic method?
A form of teaching by disciplined questioning, associated with Socrates. You state a position, the questions test it, and you either defend it, sharpen it, or give it up. It is still how law schools teach and how a good supervisor works. Its limitation has always been that it needs someone knowledgeable, available and willing to be difficult.
Does Socratic AI work for essays and dissertations?
That is where it is worth the most, because those are the pieces where an examiner asks why you claimed what you claimed. It is most useful at two moments: turning a topic into a question before you write, and stress-testing an argument you have already drafted.
Will it write the essay for me?
No. It will help you structure it, question it, find the relevant passage in your sources, and edit the text once you have decided what it should say. What it will not do is decide what you think. That is the line the product is built on, and it is deliberate.
