Every AI product in existence today is built around the same fundamental model: you bring a question, it produces an answer. The answer is fast, confident, and comprehensively expressed. This model is enormously useful for certain purposes. For learning — for the development of genuine understanding that survives the absence of the tool — it is often exactly backwards. Aethel was built on this recognition, and on the conviction that building differently was worth the cost.
We did not arrive at this position by theory alone, and I want to be honest about that.
The initial versions of what became Aethel were not dramatically different from other AI learning tools. They were more Socratic in orientation than most, but they still leaned on explanation, still produced direct answers when users were frustrated, still treated the resolution of confusion as the primary goal. They were better than many alternatives. They were not yet doing what we thought learning tools should do.
What shifted the design — what moved us from a tool that explained to a tool that interrogated — was a specific pattern we kept observing in user sessions. Users who received thorough explanations of concepts they were working on would, within a few sessions, encounter the same concepts again and engage with them as if they were encountering them for the first time. The explanation had resolved the immediate confusion. It had not produced learning. The confusion was gone; the understanding had not replaced it.
This is not a surprising finding to anyone who knows the cognitive science. Explanation, delivered to a passive recipient, does not produce deep encoding. What produces deep encoding is the cognitive effort of the recipient — the attempt to connect the new material to existing knowledge, to generate applications, to identify what is and is not understood, to work out answers rather than receive them. The explanation provided by the tool was substituting for that effort rather than enabling it. We were helping too much, and the help was preventing the thing we were supposed to be producing.
The Socratic Argument
The decision to build around questions rather than answers is not, at its core, a technical decision. It is a philosophical one, grounded in an account of what learning is and how it happens.
Socrates, working in Athens in the fifth century BCE and refusing to write anything down, developed a method of philosophical inquiry that was built entirely on questions. Not questions asked in ignorance — Socrates was often the most knowledgeable person in the room — but questions asked in service of a specific goal: making his interlocutor's thinking visible to the interlocutor, so that the interlocutor could examine it and discover, through examination, where it held and where it did not.
The elenctic method — cross-examination of a claimed belief through questioning — does something that no explanation can do: it requires the person being questioned to produce their own thinking. Not to retrieve and report what they have been told, but to actually think, in the moment, about what they believe and why. The questions that Socrates asks — What do you mean by that? Does this example fit your definition? What would have to be true for this to hold? — are not requests for information. They are prompts for generation. They put the learner in the position of having to do the work, rather than receiving a product of someone else's work.
Aethel's questioning is built on this structure. When a user asks a question, the first response is not an answer; it is a return question that asks the user to articulate what they currently think, where their current understanding falls short, what they have already tried or considered. This is frustrating in a specific way that we have accepted and that we think is valuable: it is the frustration of being asked to do something when you expected something to be done for you. It is, as we have argued elsewhere, the productive frustration that signals you are in the Zone of Proximal Development rather than outside it.
What the Research Requires
The shift to questions over answers is not merely a philosophical position. It is what the research on learning science consistently recommends, and we want to be explicit about the specific findings that inform our design.
Retrieval practice — the act of attempting to produce information from memory rather than reading or being told it — is the most consistently supported intervention in learning science. The testing effect, documented across decades and across domains, shows that testing produces better long-term retention than restudying the same material, even when the test itself is failed. The cognitive operations involved in retrieval — the activation of related knowledge, the reconstruction of the target information, the evaluation of whether the reconstructed answer is correct — strengthen the neural encoding of the material in ways that passive reception does not.
When Aethel asks a question rather than providing an answer, it is triggering retrieval practice. The user who attempts to answer Aethel's question is engaging in the cognitive operations that produce deep encoding. The user who receives Aethel's answer is not. This is not a minor difference in experience; it is a fundamental difference in what is happening neurologically. And the difference compounds over time: the learner who has been required to retrieve, repeatedly, across sessions, is building a knowledge structure that survives the absence of the tool. The learner who has been receiving answers is building, at best, familiarity with the answers — which is not the same thing.
The generation effect — the finding that self-generated information is retained significantly better than received information — reinforces this. When Aethel asks "what do you think the answer is?" before providing any information, it is not merely checking comprehension. It is creating the conditions under which the learner's own generation of an answer, even an imperfect one, will produce better encoding of the correct answer when it eventually arrives. The imperfect answer is not wasted effort; it is the mechanism by which the correct answer gets encoded more deeply than it would have if it had been provided first.
The Socratic Argument
The decision to build around questions rather than answers is not, at its core, a technical decision. It is a philosophical one, grounded in an account of what learning is and how it happens.
Socrates, working in Athens in the fifth century BCE and refusing to write anything down, developed a method of philosophical inquiry that was built entirely on questions. Not questions asked in ignorance — Socrates was often the most knowledgeable person in the room — but questions asked in service of a specific goal: making his interlocutor's thinking visible to the interlocutor, so that the interlocutor could examine it and discover, through examination, where it held and where it did not.
The elenctic method — cross-examination of a claimed belief through questioning — does something that no explanation can do: it requires the person being questioned to produce their own thinking. Not to retrieve and report what they have been told, but to actually think, in the moment, about what they believe and why. The questions that Socrates asks — What do you mean by that? Does this example fit your definition? What would have to be true for this to hold? — are not requests for information. They are prompts for generation. They put the learner in the position of having to do the work, rather than receiving a product of someone else's work.
Aethel's questioning is built on this structure. When a user asks a question, the first response is not an answer; it is a return question that asks the user to articulate what they currently think, where their current understanding falls short, what they have already tried or considered. This is frustrating in a specific way that we have accepted and that we think is valuable: it is the frustration of being asked to do something when you expected something to be done for you. It is, as we have argued elsewhere, the productive frustration that signals you are in the Zone of Proximal Development rather than outside it.
What the Research Requires
The shift to questions over answers is not merely a philosophical position. It is what the research on learning science consistently recommends, and we want to be explicit about the specific findings that inform our design.
Retrieval practice — the act of attempting to produce information from memory rather than reading or being told it — is the most consistently supported intervention in learning science. The testing effect, documented across decades and across domains, shows that testing produces better long-term retention than restudying the same material, even when the test itself is failed. The cognitive operations involved in retrieval — the activation of related knowledge, the reconstruction of the target information, the evaluation of whether the reconstructed answer is correct — strengthen the neural encoding of the material in ways that passive reception does not.
When Aethel asks a question rather than providing an answer, it is triggering retrieval practice. The user who attempts to answer Aethel's question is engaging in the cognitive operations that produce deep encoding. The user who receives Aethel's answer is not. This is not a minor difference in experience; it is a fundamental difference in what is happening neurologically. And the difference compounds over time: the learner who has been required to retrieve, repeatedly, across sessions, is building a knowledge structure that survives the absence of the tool. The learner who has been receiving answers is building, at best, familiarity with the answers — which is not the same thing.
The generation effect — the finding that self-generated information is retained significantly better than received information — reinforces this. When Aethel asks "what do you think the answer is?" before providing any information, it is not merely checking comprehension. It is creating the conditions under which the learner's own generation of an answer, even an imperfect one, will produce better encoding of the correct answer when it eventually arrives. The imperfect answer is not wasted effort; it is the mechanism by which the correct answer gets encoded more deeply than it would have if it had been provided first.
The Honest Difficulty
There is something we want to name directly, because it is both obvious and rarely said in product communications: we find it difficult too.
Designing a product that consistently withholds what the user wants in the moment is an uncomfortable design practice. Every instinct in user experience design moves toward reducing friction, increasing satisfaction, delivering what the user came for. These instincts are correct in most product contexts. In learning, they are frequently wrong, and working against them — deliberately building friction, deliberately deferring satisfaction, deliberately not giving users what they are asking for — requires constant renegotiation with the intuitions that good product design normally cultivates.
There are sessions in which Aethel's questioning feels, even to us watching the interaction, like the wrong call — where the user needed direct input and what they received was another question, and the question did not produce productive engagement but frustration that went nowhere. We do not always get the calibration right. When the struggle is not productive — when it has become merely demoralising — the questioning model is failing, and we need to provide more direct support. Recognising that line is among the hardest problems in our design.
What keeps us committed to the model despite these failures is the evidence, both from learning science and from the patterns we observe in users who stay with Aethel over time. Users who have used the product long enough to develop a working relationship with the questioning structure — who have learned, over time, what Aethel is asking for and why — report a qualitatively different experience of their own understanding. Not just that they know more, but that they understand differently: with more connections, more available transfer, more confidence in deploying knowledge in contexts where the knowledge was not explicitly practised. This is what the research predicts. It is also what we are building for.
The Product Cost of This Choice
We are not naive about what this design choice costs. We know, from user research and from honest observation of our own reactions to the product, that the experience of being questioned when you expected to be answered is less immediately satisfying than the experience of being answered. It requires more effort. It is sometimes frustrating. It does not produce the immediate resolution of uncertainty that a direct answer provides.
We also know that immediate resolution of uncertainty is, from a learning standpoint, sometimes exactly the wrong thing to produce. The uncertainty that precedes an answer — the state of not yet knowing, of having to hold the question open — is a state of high cognitive engagement. The mind that does not yet know is an active mind: searching, activating related knowledge, generating predictions, noticing gaps. The mind that has just received a complete answer is, in many cases, a passive mind: processing the information, checking whether it makes sense, accepting it. The transition from not-knowing to knowing, when it is produced by one's own effort, is qualitatively different from the same transition when it is produced by receiving an answer.
The cost is that Aethel is less pleasant to use, in moments, than tools that simply answer. The users who come to Aethel expecting a sophisticated answer-generator will be disappointed, and they should be — because Aethel is not that, and the difference is not cosmetic. We have made a product that requires more of the user than most alternatives, in the specific ways that research suggests are most productive for learning. We are not going to apologise for this, but we think users deserve to know it clearly, in advance, so that the frustration — when it arrives — can be understood as a feature rather than a flaw.
The Users We Are Not Trying to Serve
This is something that product companies rarely say explicitly, and that we think is important enough to say: Aethel is not for everyone.
It is not for the user who wants to get an answer as quickly as possible. Search engines, AI assistants, and encyclopaedias serve this user well, and they should use those tools. It is not for the user who wants to feel productive without doing the cognitive work that productivity in learning requires. The satisfaction of a long session of frictionless answer-reception is real; it is not learning.
Aethel is for the user who is willing to be uncomfortable in the specific way that the research describes as productive — who is willing to be asked a question when they wanted an answer, to have their current thinking examined rather than replaced, to discover what they do not understand rather than to have the understanding provided. This user is not rare; there are many people who understand, at some level, that real learning is effortful and that tools that eliminate the effort are eliminating something important. But this user is a specific user, and we have built specifically for them.
The consequence of this specificity is that Aethel will frustrate users who approach it with the expectation that more sophisticated AI means more comprehensive answers. Our AI is sophisticated; the sophistication is in service of asking better questions, not providing better answers. These are different applications of the same underlying capability, and the difference in experience is significant.
What Questions Actually Do
The deepest reason we build around questions rather than answers is about something that is difficult to describe without sounding either precious or mystical, but that we think is real and important.
When a question is genuinely asked of you — not a test question, not a quiz with a known right answer, but an open question about what you think and why — something happens that does not happen when you receive information. You have to be present to yourself in a way that receiving information does not require. You have to locate what you actually believe, as distinct from what you have read or heard. You have to check that belief against other things you know, and be honest about where the check succeeds and where it fails. You have to produce something from yourself rather than from a source.
This is not merely a learning strategy. It is a particular kind of intellectual engagement — the kind that Socrates called philosophical and that we would call genuine thinking, as distinct from the reception and storage of thinking done by others. Tools that answer questions are making more information available. Tools that ask questions are asking you to think.
These are not equivalent. Information, made available, can be retrieved when needed. Thinking, done, changes the thinker. The change is the thing that learning is actually for, and it is the thing that no amount of accessible information, however comprehensive and however instantly available, will produce on its own.
Every question Aethel asks is a choice not to give you the answer. That choice is deliberate, uncomfortable by design, and grounded in the most consistent finding in learning science: that what you produce is what you learn, and what is produced for you is, at best, what you have access to. We are trying to build a tool that builds you — not a tool that does your thinking and delivers the output. The distinction is the whole point.