Fenelon

Top 10 Best AI Flashcard Makers in 2026

2026-09-24T00:00:00+02:00

Every AI flashcard maker now does the easy part. Drop in a PDF, a slide deck or a YouTube lecture, and thirty seconds later you have a deck of two hundred cards. That is not where studying succeeds or fails.

It fails a month later, when the cards were generated, skimmed once and never seen again. So this list is ordered by a harder question than "how fast does it make cards": how much of the material is still in your head four weeks on? Three things decided the ranking — are the generated cards good enough to keep without rewriting them all, is there a real spaced repetition schedule behind them, and can you get your cards out if you change tools?

Free plans are compared too, because for most students the best free AI flashcard generator is the one they will actually keep using. Prices are those advertised in September 2026 and change often — check before you subscribe.

How to choose an AI flashcard maker

Start with what happens after the cards exist. A generator that hands you two hundred cards and no schedule has saved you an evening and cost you the exam: without spaced repetition, you will review everything too often at first and nothing at all three weeks later. Of the ten tools above, NoteGPT is the only one with no review mode at all, and Quizlet and Gizmo schedule reviews in ways you cannot see.

Then decide where your cards should live. If you already use Anki, the best AI flashcard generator is the one that feeds it cleanly: Limbiks, RemNote or Revisely, then let FSRS do the rest. If you want one app for everything, Quizlet, Knowt and Brainscape keep making, studying and sharing in one place. If a daily habit is what you lack, Gizmo's streaks will do more for you than a better algorithm.

On free plans, be realistic. Most free AI flashcard makers read five pages of a document, or a handful of generations a month — enough to try the tool, not to get through a semester. The genuinely free options are Anki with cards you write yourself, and Knowt's Learn mode on sets you already have.

And test the cards, not the generator. Take ten cards the AI wrote from your own notes, and try to answer them a week later without looking. If you can answer them without understanding the material, the cards are testing recognition, not knowledge — which is the moment to rewrite them, or to move to exercises that ask the same idea in more than one way.

For the wider picture of AI tools for studying — tutors, answer engines, course-to-notes tools — see our Top 10 Best AI-Powered Study Tools. And if what you want to remember is a book, see our Top 10 Best AI Book Companion Tools.

Top 10 Best AI Book Companion Tools in 2026

2026-09-21T00:00:00+02:00

An AI book companion is supposed to do what a good reading partner does: answer the question you would otherwise skip, explain the passage you read three times, and bring the book back to you weeks later, when it has started to fade.

Most tools do the first thing well. Very few do the third one at all, and that is the difference this list is ordered by: how much of the book is still with you a month after you closed it. Three questions decided the ranking — does it answer your questions inside the book itself, does it work on the books you own rather than a fixed catalogue, and does anything bring the material back before you forget it?

Prices are those advertised in September 2026 and change often — check before you subscribe.

How to choose

These ten tools answer three different questions, and confusing them is how people end up paying for four subscriptions that do the same thing.

If your problem is comprehension — the page that will not resolve, the argument you cannot follow — you want the book to answer back: Ask this Book, NotebookLM, Rebind, ChatGPT. If your problem is volume, the pile that never goes down, the answer is audio or good guides: ElevenReader, Speechify, Shortform. And if your problem is retention, and for most readers it is, only four tools on this list will do anything about it: Fenelon, Readwise, Recall, and Rebind's slower reading pace.

The honest test is the one you can run tonight. Take a book you finished last month, close everything, and write down what it argued and what changed your mind. If the page stays empty, you do not need a better summariser — you need something that brings the book back to you next week, and the week after. That is the whole difference between a tool that reads with you and a tool that reads for you.

For tools aimed at studying rather than reading — flashcards, tutoring, exam revision — see our Top 10 Best AI-Powered Study Tools.

And if you would rather turn a book into cards you review, see our Top 10 Best AI Flashcard Makers.

Top 10 Best AI-Powered Study Tools in 2026

2026-09-18T00:00:00+02:00

Every month brings another "AI study app". Most of them do the same three things — summarise a document, generate flashcards, answer questions — and the differences that matter only surface after a few weeks of use.

This list is ordered by how much a tool changes the way you study, not by how much it can generate. Four questions decided the ranking: does it help you understand rather than hand you an answer, does it make you practise, does it bring material back before you forget it, and how much work does it take off your hands?

Prices are those advertised in September 2026 and change often — check before you subscribe.

How to choose

Most students end up using two tools, not ten. One that helps them understand — a conversational assistant, a tutor, an answer engine — and one that makes the material stay: cards, exercises, a review schedule. Tools that try to be both are rarer than the marketing suggests.

So start from the failure you keep repeating. If you follow a lecture and cannot reproduce it a week later, the problem is retention, and the answer is spaced repetition, not a better summariser. If you read the chapter three times and still cannot say what it argues, the problem is understanding, and no flashcard app will fix it. If you never quite know what to work on next, what you need is a path, not another generator.

And whichever tool you choose, keep one habit: close it, and try to explain what you have just learned from memory. Every tool in this list can produce an explanation for you. None of them can have understood it for you.

If what you are trying to get through is a book rather than a course, the tools are not quite the same: see our Top 10 Best AI Book Companion Tools.

If flashcards are the part you want to get right, we compared the AI tools that generate them — and schedule them — in our Top 10 Best AI Flashcard Makers.

Critical Questions and Argumentation Schemes: Learning to Evaluate an Argument

2026-09-02T00:00:00+02:00

We accept or reject arguments all day long: in an article, a meeting, an advertisement, a conversation. Most of the time we do it by intuition. Argumentation theory offers something sturdier: models of the structure of arguments, a catalogue of their common forms and, for each form, the list of questions that test its strength. These critical questions are one of the most useful tools of critical thinking. They have also recently become a testing ground for large language models.

Toulmin's model: the anatomy of an argument

Stephen Toulmin proposed in 1958 a description of argument that remains the most widely used. A complete argument has six elements:

The claim: the assertion we want to have accepted.
The data: the evidence used to support it.
The warrant: the statement that explains why the data support the claim.
The backing: the underlying assumptions that reinforce the warrant.
The rebuttals: the conditions under which the claim would not hold.
The qualifier: the degree of certainty of the claim ("probably", "unless").

Example. "You should take an umbrella" (claim), "the sky is overcast" (data), "an overcast sky often means rain" (warrant), "according to the region's weather statistics" (backing), "unless the wind blows the clouds away" (rebuttal), "probably" (qualifier). Breaking an argument down this way reveals what is left unsaid: most often, the warrant.

Toulmin's model applied to the umbrella example: the warrant and its backing, often left implicit, support the step from data to claim.

Walton's argumentation schemes

Douglas Walton catalogued argumentation schemes: forms of argument that represent common structures in everyday, scientific or legal discourse. They include the deductive and inductive forms of classical logic, but also a third family, often the most interesting: defeasible arguments. A defeasible argument is one whose conclusion can be accepted on the available evidence, but revoked if new evidence appears. Reasoning based on a generalization subject to exceptions is the typical case.

Schemes can be grouped into families. Each has its own critical questions; here is one per family, by way of example.

Family	What it does	Typical critical question
Knowledge and authority	Passing on an opinion deemed reliable: position to know, expertise, testimony, popular opinion.	Is the source really in a position to know, and in this specific field?
Causality and diagnosis	Linking phenomena by cause, effect or sign.	Does another explanation account for the same facts?
Classification and analogy	Placing a case in a category to apply a property to it, or claiming that two similar cases call for the same conclusion.	Are the two cases alike in the respects that matter for the conclusion?
Deliberation	Recommending an action to reach a goal: instrumental reasoning, arguments from values, from consequences, from sunk costs.	Is there another way to reach the goal, or negative consequences left unmentioned?
Personal	Attacking the interlocutor rather than the argument.	Does the trait held against the person bear on their reliability about this subject?
Slippery slope	Warning of a drift presented as inevitable.	Is each step of the drift really inevitable, and what prevents stopping along the way?

Two epistemic families deserve detail. The schemes that apply rules to cases: established rule, precedent, analogy, verbal classification. And the schemes by discovery: best explanation (a hypothesis is plausible because it best explains the facts), sign, the move from correlation to cause, and the argument from ignorance (a proposition is false because no evidence was found after an appropriate search). These schemes are close cousins of cognitive biases: a bias is often a scheme applied outside its conditions of validity.

The argument from sign

"I see bear tracks on the trail. So a bear passed here recently." This reasoning is merely plausible. It is legitimate only under conditions of doubt and lack of information, and it must yield to a better explanation (tracks molded by a prankster, for instance).

The argument from authority, or expert opinion

This is the most common scheme and the most misused. Its structure: source E is an expert in domain S containing proposition A; E asserts that A is true; therefore A is true. Walton attaches six critical questions to it:

Expertise: how credible is E as an expert?
Field: is E an expert in the field that A belongs to?
Opinion: what did E actually assert that implies A?
Trustworthiness: is E personally reliable as a source (biased, honest, conscientious)?
Consistency: is A consistent with what other experts assert?
Evidence: is E's assertion based on evidence?

A more rigorous version of the scheme adds an explicit conditional premise ("if E is an expert and asserts A, then A should be taken as probably true"), which turns it into a plain modus ponens and shifts the whole debate to the acceptability of that conditional.

How to attack an argument

There are two broad ways to attack an argument: criticize it or refute it. You can put forward a counter-argument, whose conclusion is the negation of the original conclusion. You can attack a premise, by questioning it or showing it is false. For a defeasible argument there is a third route: attacking the inference rule, the warrant or the generalization it rests on, for example by citing an exception.

The final evaluation of an argument comes down to three criteria: the acceptability of the premises, their relevance to the conclusion, and their sufficiency, that is, whether they constitute a strong enough reason to conclude.

Critical questions and the burden of proof

Critical questions can be understood in three ways: as a form of attack on a defeasible argument, as deeper implicit assumptions that have not been established, or as starting points for finding a refutation.

They are not all equal with respect to the burden of proof. In the expert scheme, asking questions 1, 2 or 3 sends the burden back to the person who put forward the argument: they must establish the expertise, the field, the exact assertion. Asking questions 4 or 5, on the contrary, places the burden on the questioner: accusing an expert of bias requires grounds. This asymmetry explains why some objections are admissible without evidence and others are not.

One question remains open: when to stop? Once all the critical questions have been asked, should we accept the argument, or open new sub-questions? At what point is an argument strong enough? The theory provides no stopping rule. It is a judgment, and it is precisely that judgment that has to be learned.

Where do LLMs fit in?

Since 2024 a research task has formed around critical questions generation: using a language model to ask the questions that expose the weak points of a person's argumentative discourse. The first work on the subject started from very simple prompts ("list the critical questions that should be asked before accepting the arguments in the following paragraph") and had the generated questions annotated by humans into three categories:

useful: the answer to the question can change one of the arguments in the text;
not useful: the question is valid, but its answer will probably change nothing;
invalid: the question cannot challenge any argument, because its reasoning is wrong, it is not related to the text, it introduces concepts absent from the text, it is so general that it would apply to any text, or it is a mere reading-comprehension question.

A reference dataset followed in 2025, built from debate corpora annotated with their argumentation schemes, with several thousand evaluated questions. The task proved difficult: models readily produce plausible but generic questions, of the "what are the sources for this claim?" kind, which apply to everything and attack nothing. The winning system of the competition organized that year worked in two steps, generating many questions and then selecting the most specific ones, which says clearly where the difficulty lies: not in production, but in selection.

This difficulty is instructive for the learner. A good critical question is specific to the text, tied to a precise argument, and its answer could change the conclusion. These are exactly the criteria that distinguish an objection from a mere display of skepticism.

An exercise to practice

Take an editorial, a speech or an advertisement. Identify the main claim and its data. Look for the implicit warrant. Determine which scheme the argument belongs to: authority, sign, analogy, consequences? Then ask the critical questions specific to that scheme, noting for each one who carries the burden of proof. Finally, decide: accept, reject, or suspend judgment.

This is the kind of exercise Fenelon generates from the texts a learner studies: spotting a scheme, formulating a useful critical question, telling a contestable premise from a fragile inference rule. Language models are used there to produce varied cases and to assess answers against explicit criteria, not to judge in the learner's place.

Conclusion

Argumentation schemes give a name to the forms of reasoning we use without thinking, and critical questions give, for each of them, the list of points where they can give way. Knowing how to recognize them does not make you skeptical of everything: on the contrary, it lets you accept an argument knowing why, and reject it knowing where.

The Memory Paradox: Why Learn by Heart in the Age of AI

2026-08-26T00:00:00+02:00

"Why learn it by heart when I can look it up in three seconds?" Students have been asking this since search engines arrived, and conversational assistants have given it new force. A paper published in 2025 under the title "The Memory Paradox" turns the question around: is our growing reliance on digital memory aids and digital distractions gradually eroding the very intellectual capacities we are trying to develop? Here is what cognitive science lets us answer.

Knowledge and skills are two sides of the same coin

There is no skill without knowledge. The pedagogical debate of the last thirty years has often been reduced to a slogan: "a guide on the side, not a sage on the stage". The teacher accompanies, the student discovers. The formula is appealing, but it rests on a confusion between two kinds of knowledge.

Primary knowledge, such as language or face recognition, is acquired on its own, by simple immersion. Secondary knowledge, such as reading, mathematics or science, is not acquired naturally: it requires explicit teaching, repetition and time. Discovery-based pedagogies work well for the former and much less well for the latter. And it is the latter we are trying to pass on.

Cognitive offloading

When I use a tool to reach the result of a mental operation faster, I offload the mental effort needed to solve it. Psychologists call this cognitive offloading. The phenomenon is not new: writing, notebooks and calculators are all forms of offloading. What has changed is its scope. An assistant that can write, summarize, compute and remember on our behalf lets us offload almost everything.

The problem is not offloading as such, it is excessive offloading at the moment of learning. Internalizing knowledge takes effort: that effort is what builds the mental representation. Skipping it leaves a weak mental schema that cannot later support critical thinking or problem solving. We believe we have learned because we obtained the answer.

A troubling clue: after decades of increase, average IQ has been declining in several Western countries since the 1990s. The causes are debated, and a correlation is not an explanation. But the coincidence with the spread of digital memory aids at least deserves the question.

Two memories, not one

To understand what is lost, we need to distinguish two memory systems.

Declarative memory stores facts and episodes: the capital of Australia, the date of an appointment, the statement of a theorem. It is the one digital tools replace most easily.

Procedural memory stores skills and routines that have become second nature: driving, reading, playing an instrument, recognizing an argumentative structure without thinking about it. It is implicit, slow to build and highly resistant to forgetting.

Developing real expertise often consists in moving knowledge from the declarative system to the procedural one. You practice until the skill is embedded in the unconscious networks that make intuition and fluid thinking possible. An expert does not "search": they see. And that transfer requires precisely the repetition that offloading spares us.

Two memory systems. Expertise consists in moving knowledge from one to the other through repetition; offloading short-circuits that transfer.

Mental schemas and engrams

Internalized networks form structures called schemas. A schema organizes knowledge and enables complex thinking: when a chess player looks at a position, they do not see thirty-two pieces but a handful of familiar configurations. Schemas develop gradually through practice with concepts. Metaphors enrich them by linking a new concept to an experience that is already known.

At the biological level, the physical trace of a memory is called an engram: the state of neural connections left by an experience. An engram is formed through repetition and through the association of ideas, concepts and experiences. No external search builds it for us.

Epistemic friction

There is a second reason why the instant answer teaches nothing. The more a new piece of knowledge contradicts my beliefs, the more slowly I assimilate it. When a person receives information that seems ambiguous or inconsistent with their current beliefs, we can say they are in epistemic friction: a resistance to the new knowledge.

This friction is not a flaw. It is a setting. Too little friction would make us believe anything; too much would prevent us from ever revising our ideas. It depends on the gap between the incoming information and our beliefs: a head-on contradiction produces strong friction, a marginal one produces weak friction. Recent work on dialogue even tries to quantify it in order to calibrate how fast ideas get revised.

Learning means crossing that friction. A tool that gives the answer before we have had time to form our own hypothesis removes it: nothing rubs, nothing sticks.

The strategies that build memory

Cognitive science identified long ago the practices that strengthen engrams. They all have one thing in common: they demand effort at the right moment.

Retrieval practice: try to remember before checking. Testing yourself beats rereading.
Spacing: spread revisions over time rather than cramming them. Flashcard apps with a spaced repetition scheduler do this for you — see our comparison of the best AI flashcard makers.
Interleaving: alternate between different but related topics. This technique forces you to repeatedly discriminate between several engrams, which strengthens each of them.
Generation: produce an answer, even a wrong one, before receiving the right one.
Elaboration: explain in your own words, connect to an example, find a metaphor.

Offloading then has its place, but after the effort, not instead of it. Checking an answer you first attempted is one use of the tool; having it handed to you from the start is another.

What this implies for an AI tutor

An assistant that answers everything is the worst possible tutor, not because it is wrong, but because it succeeds too fast. In Fenelon we made the opposite choice: the exercise comes before the explanation, the question before the answer, and reviews are spaced out over time according to successes and failures. AI is used to choose the content and the right moment, not to think in the learner's place.

Conclusion

The memory paradox fits in one sentence: the tools that make knowledge most accessible are the ones that make its acquisition least likely. The solution is not to give up the tools, but to reclaim what they cannot do for us: the effort of retrieval, spaced repetition, the friction of an idea that resists. That effort, and it alone, leaves a trace.

Can a Socratic Dialogue Be Programmed? What Plato Teaches AI Tutors

2026-08-19T00:00:00+02:00

Many of us studied a Platonic dialogue at school. We remember a text in which Socrates asks his interlocutor questions and, question after question, a certainty collapses. Now that large language models can hold a conversation, the idea is everywhere: what if software could question like Socrates? This article takes the question seriously, which means analyzing what Socrates' questions actually do before deciding what a program can reproduce.

Four ways of "questioning like Socrates"

The phrase covers at least four different practices that are often confused.

A teaching device: a teacher who plays the ignorant and leads the student, through questions, to formulate an answer or an understanding by themselves.
Maieutics: a dialectical introspection that gives birth to knowledge the interlocutor was carrying without knowing it. Socrates compares himself to his mother, a midwife.
The elenchus: an examination that clarifies an opinion and refutes it, sometimes leading to aporia, that perplexity in the face of one's own ignorance.
An ethical exercise: the examination of one's life, actions, judgments and values. "The unexamined life is not worth living."

This last point is often forgotten. The themes of the Socratic dialogues are almost all moral: courage, piety, justice, wisdom, the good life. That context forces Socrates' interlocutors to say what they really think, since they are talking about the content of their own lives. Does the method remain valid when applied to photosynthesis or quadratic equations? Nothing guarantees it a priori. That is the first question to ask before programming anything.

A functional analysis of the questions

When preparing to implement a system, you carry out what is called a functional analysis: you break down what the system does into functions, then implement them one by one. Here, the functions we care about are not those of the dialogue as a text, nor the effects of reading it on us. They are the functions of Socrates' questions on the person who has to answer them.

Several can be distinguished.

Defining. These are the first questions that come to mind when we think of philosophy. "What do you call thinking?" asks Socrates in the Theaetetus. "What do you say the pious is?" in the Euthyphro. "What is temperance, in your opinion?" in the Charmides. The function is to force the interlocutor to spell out a concept they believed they mastered.

Examining and refuting. This is the elenchus, and it makes up by far the largest part of the dialogues. The mechanism is always the same. The interlocutor asserts a thesis. Socrates obtains, through a few questions, their agreement on additional premises that seem obvious. Then he shows, through further questions, that these premises logically imply the negation of the initial thesis. The interlocutor contradicts himself: nobody refuted him from the outside.

Clarifying and exemplifying. Asking for an example, a counter-example, a distinction between two close cases.

Remembering. Recollection: leading the person to mobilize what they already know, often without having connected it to the present question.

Comparing. Analogy, which brings two concepts together to reveal a difference, and irony, which underlines the gap between opinion and true understanding.

Many dialogues follow a three-step progression: identifying the question, criticizing the question as it was asked, then problematizing it, that is, confirming that there is indeed a problem and that progress is possible.

What recent research has done with it

Several studies have tried to translate these functions into instructions for a language model. A 2023 paper, "Prompting Large Language Models with the Socratic Method", lists rules close to what a good seminar leader applies spontaneously:

ask open questions to stimulate thinking;
help clarify and define the relevant terms;
encourage the person to give examples or evidence;
lead them to doubt their conclusions;
help them summarize and draw conclusions;
help them reflect on the method itself and on what was learned during the dialogue.

The same paper distinguishes the techniques involved: definition, generalization (drawing principles from recurring patterns), induction, elenchus (testing the consistency of a hypothesis), hypothesis elimination, maieutics, dialectic (exploring opposing views), recollection, irony and analogy. A more recent work, "Dialogic Pedagogy for Large Language Models" (2025), goes further by confronting the conversational strategies of models with established theories of learning, rather than with the figure of Socrates alone.

The practical result is mixed. A language model is very good at producing an open question, asking for an example or restating. It is much less good at conducting an elenchus, because an elenchus requires holding on to the initial thesis, choosing the right intermediate premises and keeping the logical chain intact over several turns. Above all, it does not know, on its own, when to stop.

Socratic and constructivist

In everyday pedagogical vocabulary, "Socratic" has almost become a synonym for "constructivist". Any strategy that fosters autonomous reflection and discovery learning gets called a Socratic dialogue. That keeps a single function of Socrates, maieutics: dialogue as a method for generating knowledge through introspection.

This reading is legitimate but it has a cost. Plato's Socrates leads his interlocutors to aporia far more often than to knowledge. For a philosopher, aporia is a beginning. For a learner trying to understand fractions, it is mostly frustration.

What Socrates lacks: the zone of proximal development

This is where the psychology of learning becomes indispensable. Lev Vygotsky stressed the social aspects of human learning: we learn, to a large extent, thanks to other, more competent humans. Three of his concepts directly concern a tutor who asks questions.

The zone of proximal development designates what we do not yet understand but can almost understand: what we can do with help, and what that assisted experience will later allow us to do alone. Outside this zone there is nothing to learn: either the task is already mastered, or it is out of reach.

Scaffolding is the help needed to operate within that zone. Wood, Bruner and Ross described six functions: recruiting the learner (getting them on board and motivated), reducing the degrees of freedom of the task, maintaining direction, marking critical features, controlling frustration and demonstrating.

Stagnation is what happens when you work on what you already know.

The zone of proximal development: a tutor only has leverage in the middle ring, and must know where the learner stands to stay there.

A Socratic tutor without a learner model does not know which zone it is in. It may ask a definition question of someone who needs an example, or launch a refutation at someone who is about to give up. Motivational help (making sure the learner cares about the exercise, keeping attention on the task rather than on a technique already mastered, softening frustration and switching activity before it becomes a problem) matters as much as pedagogical help.

The building blocks of an intelligent tutor

Intelligent tutoring systems, studied since the 1980s, rest on four components: the teaching material, the teaching strategy, the learner model and the interface. Socratic dialogue is a strategy. It is only worth something if the other three blocks exist, and in particular the learner model: their interest in the topic and in the exercise, their attitude towards difficulty, their tolerance for frustration, what they have already succeeded at and failed.

That is the choice we made in Fenelon. Open questions are assessed against explicit answer criteria, which makes it possible to pinpoint what was not understood. Feedback does not hand over the answer: it asks first, in the manner of Socrates, but it also knows when to step down a level, change activity or simply explain. Questioning is one tool among others, chosen according to the learner's state, not a posture.

Conclusion

Can a Socratic dialogue be programmed? Its functions can: defining, exemplifying, refuting, prompting recollection, comparing. What cannot be programmed without a theory of learning is the judgment that decides which function to apply, to whom, and when. Socrates knew when to keep quiet. That is the hardest thing for an AI tutor to learn.

Bloom's Taxonomy: How to Write and Test Learning Objectives

2026-08-11T00:00:00+02:00

When you prepare a course, a learning path or even a short revision session, the first question is not "what am I going to present?" but "what should the learner be able to do at the end?". Bloom's taxonomy was invented to answer that question with a shared vocabulary. Seventy years later, it remains the most widely used framework for writing and assessing learning objectives.

Where does Bloom's taxonomy come from?

In the early 1950s, a group of American educators led by Benjamin Bloom set out to make it easier for teachers and examiners to talk to each other. Everyone spoke about objectives, but with different words. The result, published in 1956, is a classification of objectives in the cognitive domain that is at once educational, logical and psychological. Its primary purpose was modest: to let teachers understand one another when describing what they expect from their students.

The original taxonomy is hierarchical and cumulative: each level assumes mastery of the previous ones.

Knowledge
Comprehension
Application
Analysis
Synthesis
Evaluation

In 2001, David Krathwohl, one of the original authors, and Lorin Anderson published a revision that changed two important things. First, it separated two dimensions the 1956 version had mixed together: the form of knowledge (factual, conceptual, procedural, metacognitive) and the cognitive process involved (remember, understand, apply, analyze, evaluate, create). Second, it replaced "synthesis" with "create" and moved it to the top, above "evaluate". An objective now reads as a cell in a two-dimensional table: "understand a concept", "apply a procedure", "evaluate a theory".

The six levels of the revised taxonomy (2001) and their equivalents in the 1956 version.

What is a learning objective?

A learning objective expresses how we expect a person to change through the learning process: what they will think, feel or do differently. It is almost always written on the same pattern:

The learner should be able to VERB + SUBJECT OF LEARNING.

For example: "recall the law of supply and demand", "explain the principle of natural selection in your own words", "construct an original example of an argument by analogy". The verb names the cognitive process, the object names the knowledge being targeted. This combination is what makes the objective verifiable.

Objectives can come from three sources: teachers, learners themselves (what they already know, their interests, the situations they encounter) and subject-matter experts. Ralph Tyler summed up the work of designing a curriculum in four questions that have not aged:

Which learning objectives should be reached?
Which learning experiences make it possible to reach them?
How should these experiences be organized so that they form a continuous whole rather than a series of isolated topics?
How can the effectiveness of each experience be evaluated?

Level 1: knowing, remembering, recognizing

This is the most common category of objectives, and often the least loved. Yet it covers very different kinds of knowledge:

Specific facts: the terminology of a field (the jargon experts use among themselves) and specific facts (dates, people, places, sources).
Ways of dealing with specifics: conventions (usage, style, rules of presentation), trends and sequences, classifications, criteria of judgment and the methodology of a field.
Abstractions: principles that summarize observations, then theories that link those principles into a systematic view of a complex phenomenon.

Three remarks make this level more interesting than it looks. Knowledge that is organized and connected to other knowledge is better retained than isolated knowledge. Remembering depends on the cues available at the moment of recall, which is why it pays to vary the way a question is asked. And knowledge is a prerequisite for everything else: without facts and concepts in memory, there is nothing to analyze or evaluate.

How do you test it? By checking that the person can state or recognize true propositions in answer to questions. The form of the question and the expected level of precision should not stray too far from the form in which the knowledge was learned, otherwise you are testing something other than memory.

Level 2: understanding

To understand is to be confronted with a communication (a text, a diagram, a formula) and to be able to do something with it. The taxonomy distinguishes three comprehension behaviors, in order of increasing difficulty.

Translation means transforming a communication from one form into another: restating a concept in your own words, summarizing a text, illustrating an abstract idea with an example, turning a chart into a sentence or a sentence into a diagram. It is the hinge between memory and the higher levels.

Interpretation means grasping a communication as a configuration of ideas: understanding the relationships between its parts, telling the essential from the incidental, relating it to your own experience. Being able to distinguish warranted, unwarranted and contradictory conclusions drawn from a document is a matter of interpretation.

Extrapolation means extending a communication beyond the limits set by its author: estimating consequences, predicting the continuation of a trend, distinguishing a value judgment from a prediction.

How do you test it? For translation, present the definition of a concept in a form different from the one that was learned, or ask the learner to recognize examples of the concept among counter-examples. For interpretation, give a document and ask for its consequences, its summary or its outline. For extrapolation, ask for a prediction. In every case the context must be new to the learner, otherwise you are back to a memory test.

Levels 3 to 6: apply, analyze, evaluate, create

The higher levels are the ones we mention most and assess worst, because they require open-ended situations.

Apply: execute a known procedure in a new situation, or implement a general method on an unfamiliar problem.
Analyze: break a communication down into its elements, make the relationships between those elements explicit, then recognize the organizing principles that structure it (the author's point of view, technique, bias).
Evaluate: make a reasoned judgment, either from internal evidence (consistency, logical accuracy) or from external criteria (comparison with standards or with other works).
Create: produce an original communication, a plan or an ordered set of operations, or synthesize a set of abstract relations into a theory or a model.

The revised taxonomy insists on a useful distinction: intellectual abilities let you use remembered technical knowledge on a new problem; intellectual skills let you use a general method in a new situation. The former are learned through repetition, the latter through variety.

Other classifications

Bloom is not alone. Gagné proposes a five-step learning hierarchy (verbal association, discrimination, concept learning, rule learning, problem solving). Ausubel and Robinson distinguish representational, concept and propositional learning, followed by application, problem solving and creativity. Gerlach and Sullivan prefer six observable behaviors: name, order, identify, describe, demonstrate, construct. These frameworks overlap widely and confirm an intuition: there is a natural progression from recall to production.

What this changes for a tutor, human or AI

The practical value of the taxonomy is that each level maps to a family of activities. A multiple-choice question or a cloze test checks recall. A restatement, a summary or picking the right example checks understanding. A case study checks application, a critique of an example checks analysis and evaluation, building an original example checks creation.

Level (2001)	Typical verbs	Exercise that checks it
Remember	recall, recognize, name	multiple-choice question, cloze test
Understand	restate, summarize, illustrate, predict	restatement, summary, picking the right example
Apply	execute, use, implement	case study, unfamiliar problem
Analyze	break down, relate, distinguish	critique of an example, spotting a bias
Evaluate	judge, compare, justify	reasoned critique against explicit criteria
Create	produce, design, construct	building an original example, a plan, a model

For the Remember level, well-written flashcards reviewed on a spaced schedule are the most efficient format; we compared the best AI flashcard makers on exactly that criterion.

This is exactly the logic we follow in Fenelon: a learning goal is broken down into finer objectives, each placed at a level of the taxonomy, and that level decides which type of exercise is offered. A learner who fails an understanding objective does not need a longer exercise; they need an exercise at a different level, often one step down, before climbing back up.

Bloom's taxonomy is not a theory of learning. It is a language. But a language shared by the person who designs a path, the person who assesses it and the person who follows it is often worth more than a theory.