prompt: "Create a taxonomy of categories of reasoning, including an explanatory description, with examples of each. If someone were to learn about reasoning from first principles today, they would begin by understanding the modes or types of reasoning. They would then review a wide range of examples for each."
completion:
The best place to begin is with how a conclusion is supported. Deduction, induction, abduction, and analogy provide the most useful initial distinctions. But they are not the same kind of category as “causal reasoning,” “ethical reasoning,” or “systems thinking.” Those describe what we are reasoning about, what we are trying to accomplish, or how we organize the problem. A single argument can belong to several categories at once. (Stanford Encyclopedia of Philosophy)
For learning purposes, I recommend a layered taxonomy of 20 reasoning families:
| Layer | What it classifies | Categories |
|---|---|---|
| A. Inference patterns | How premises support conclusions | Deductive, inductive, abductive, analogical, default |
| B. Reasoning within models | The structures and relationships being examined | Quantitative, probabilistic/statistical, causal, hypothetical/counterfactual, spatial/relational, temporal/systems |
| C. Reasoning toward particular purposes | What kind of question we are trying to answer | Practical, strategic, normative/ethical, interpretive/social |
| D. Problem-solving and quality-control methods | How we construct, examine, and improve reasoning | Conceptual/first-principles, decompositional/compositional, constraint-based/algorithmic, critical/dialectical, metacognitive |
This is a teaching framework, not a claim that there are exactly 20 distinct mental faculties. The overlaps are important.
A useful working definition is:
Reasoning is using information, assumptions, or goals to draw, evaluate, or revise a conclusion.
An argument makes this structure explicit: its premises are the starting points, its conclusion is what is being supported, and the inference is the connection between them. Practical reasoning extends the picture to conclusions about what to do. (Stanford Encyclopedia of Philosophy)
For every example, ask:
What are the premises? What is being concluded? Why should the premises support that conclusion? How strong is that support?
The foundational distinction is whether true premises guarantee the conclusion or merely support it. A valid deduction guarantees its conclusion conditional on its premises. Nondeductive reasoning can be excellent without providing that guarantee. (Stanford Encyclopedia of Philosophy)
The situations and numerical examples below are hypothetical teaching cases.
Central question: “What must follow, assuming these premises are true?”
Deduction derives conclusions that cannot be false while all the premises are true. Its defining feature is this relationship of necessity—not simply movement from “general” statements to “specific” ones. Formal logic studies systems for representing and evaluating such consequences. (Stanford Encyclopedia of Philosophy)
Examples. A competition’s rules disqualify every late entry. This entry was late. Therefore, it is disqualified. A key is either in the drawer or the cupboard; it is not in the drawer; therefore, it is in the cupboard.
If package A is heavier than B, and B is heavier than C, then A is heavier than C. If an integer is even, it can be written as (2k); its square is (4k^2=2(2k^2)), so its square is also even.
What to check: Evaluate the premises separately from the inference. The key argument is impeccable only conditional on the premise that those are the two possible locations. A third possible location undermines the premise, not the deductive form.
Central question: “What do observed cases suggest about unobserved cases?”
Induction extends beyond what has been directly observed: from a sample to a population, past cases to future cases, or repeated observations to a general pattern. Here, I use the term in this narrower sense; some traditions use “induction” more broadly for nondeductive inference. (Stanford Encyclopedia of Philosophy)
Examples. Of 2,000 randomly sampled components, 1,800 pass inspection. You estimate that about 90% of the batch will pass. Twenty comparable weekday journeys have taken 25–35 minutes, so you expect tomorrow’s journey to take roughly that long under similar conditions.
Repeated tests show a particular material expanding when heated, leading to a proposed generalization over the tested temperature range. A program succeeds on many representative test inputs, increasing confidence that it will succeed on similar untested inputs.
What to check: Ask whether the observations are representative and whether conditions have changed. Testing many inputs does not, by itself, prove a program correct for every input.
Despite its name, mathematical induction is a deductive proof method, not an inference from a finite collection of observed examples. (Stanford Encyclopedia of Philosophy)
Central question: “What would best explain what I observe?”
Abduction generates or favors explanatory hypotheses. It begins with observations and asks which explanation would make them most intelligible, often considering fit with the evidence, background knowledge, and competing possibilities. It does not establish the explanation with deductive certainty. (Stanford Encyclopedia of Philosophy)
Examples. Every computer in an office loses internet access simultaneously while the local network still works. A failure in the shared external connection is a better initial explanation than unrelated failures of every computer.
Water appears beside a dishwasher immediately after its cycle, making a dishwasher leak a leading hypothesis. Two manuscripts contain the same highly unusual error, suggesting copying or a shared source. A machine jams only when using one batch of components, suggesting that something about that batch is responsible.
What to check: Generate genuine alternatives. “This hypothesis explains the evidence” is weaker than “this hypothesis explains the evidence substantially better than plausible alternatives.” The best explanation you have considered may still be wrong.
Central question: “What might carry over from a relevantly similar case?”
Analogical reasoning transfers expectations, explanations, or possible solutions from a familiar source to a less familiar target. Its strength depends on whether the similarities concern relationships relevant to the proposed conclusion, rather than superficial resemblance. (Stanford Encyclopedia of Philosophy)
Examples. A repair fixed a particular fault in a closely related appliance model, so it is a promising repair to investigate here. A help desk resembles a checkout system in having arrivals, service times, and queues; a queue-management approach that helped one may help the other.
A school excused a missed deadline caused by a documented building closure. A student argues that a documented submission-platform outage deserves similar treatment because both prevented submission through circumstances outside the student’s control. An onboarding procedure helped one team, so a second team with similar work and experience levels tries it.
What to check: Identify the specific relationship that transfers, then look for differences that could defeat the comparison. “Both organizations are large” is usually a weaker basis than “both face the same coordination problem.”
Central question: “What is reasonable to assume unless an exception appears?”
Default reasoning uses ordinarily reliable expectations while allowing them to be withdrawn. It belongs to the broader family of defeasible reasoning: reasoning whose conclusions can lose their justification when additional information arrives. Induction, abduction, and analogy are also commonly defeasible. (Stanford Encyclopedia of Philosophy)
Examples. A shop advertises weekday opening hours of 9–5, so you expect it to be open at 2—until you learn it is closed for a holiday. Someone accepts a meeting invitation, so you expect them to attend—until they cancel.
You expect an object to remain where you left it unless someone moved it. You provisionally accept a competent witness’s account unless evidence of poor visibility, mistaken identity, or deception emerges.
What to check: Make the implied “normally” or “unless” explicit. Revising a default conclusion in light of an exception is not necessarily a reasoning failure; refusing to revise it may be.
These families concern the representations and relationships used. They can employ any of the inference patterns above.
Central question: “How much, how fast, in what proportion, or within what bounds?”
Quantitative reasoning works with quantities, units, rates, proportions, scaling relationships, and numerical constraints. It includes both exact calculations and disciplined approximation; useful methods include dimensional analysis, extreme-case checks, and successive approximation. (MIT OpenCourseWare)
Examples. A recipe uses 300 grams for four servings, so ten servings require 750 grams if ingredients scale proportionally. Producing 2,400 items in eight hours implies an average rate of 300 items per hour.
For 100 guests receiving two 250-milliliter drinks each, the estimated requirement is 50 liters. Doubling the side length of a square multiplies its area by four, not two.
What to check: Inspect units and scaling assumptions. Exact arithmetic applied to rough estimates still produces an estimate. A recipe may scale proportionally while cooking time does not.
Central question: “How uncertain should I be, and how should evidence change that uncertainty?”
Probabilistic reasoning represents uncertainty numerically. Statistical reasoning uses observations and models to assess claims about populations or processes. Bayesian reasoning is one important approach: it updates degrees of belief using how expected the evidence is under competing hypotheses. (Stanford Encyclopedia of Philosophy)
Examples. Under a fair six-sided-die model, the probability of rolling a six is (1/6). That calculation is deductive within the assumed probability model.
Of 1,000 parts, ten are faulty. A screening procedure flags all ten faulty parts but also 50 good ones. Only (10/60), or about 17%, of flagged parts are actually faulty. The overall rarity of defects matters.
A survey reports 52% support, but you examine sampling uncertainty before concluding that a majority of the population supports the proposal. A hypothesis starts with odds of 1:4; evidence that is three times as likely under that hypothesis as under its alternative changes the odds to 3:4.
What to check: Examine base rates, sampling methods, dependence between observations, and model assumptions. A precise probability is not automatically a well-grounded probability.
Central question: “What produces this outcome, and what would happen if we changed something?”
Causal reasoning examines how outcomes are generated and how interventions would alter them. Mechanistic reasoning specifies the intermediate process linking a cause to an effect. Causal models distinguish observing a condition from actively changing it. (Stanford Encyclopedia of Philosophy)
Examples. Workers are randomly assigned to two instruction formats, and one group consistently makes fewer errors. This supports a causal effect of the format, subject to statistical uncertainty and experimental assumptions.
Umbrella use and wet sidewalks rise together, but in the stipulated model rain causes both; umbrellas do not cause the wet sidewalks. Disabling a software extension removes an error, and re-enabling it restores the error, supporting the extension’s causal role. A damaged power connection causes intermittent voltage loss, which triggers a device to restart: that sequence supplies a mechanism.
What to check: Consider alternative causes, reverse causation, and common causes of both variables. Ask whether the proposed intervention actually isolates the factor of interest.
Central question: “What would follow under a different set of assumptions?”
Hypothetical reasoning explores suppositions. Counterfactual reasoning explores alternatives to the actual or assumed situation, such as what would have happened had an earlier event been different. Its conclusions depend on which other conditions are held fixed and which are allowed to change. (Stanford Encyclopedia of Philosophy)
Examples. Revenue is 100 and costs are 90. If revenue fell by 20% while costs remained unchanged, the organization would lose ten.
A meeting began at 10:00. If you had left at 9:15 and the journey had still taken 30 minutes, you would have arrived before it began. In a specified backup-power model, failure of the main supply would not interrupt operation because the backup would take over. An imagined lottery with equal chances but unequal prizes demonstrates that equal opportunity need not produce equal outcomes.
What to check: State what changes and what remains constant. A counterfactual conclusion is only as reliable as the assumptions governing that alternative situation.
Central question: “How are these things arranged, connected, ordered, or contained?”
This family works with relationships such as left of, inside, adjacent to, connected with, and larger than. Spatial reasoning may use diagrams or mental models, but the crucial content is the relationships represented, not merely the visual appearance. (Frontiers)
Examples. A is left of B, and B is left of C, so A is left of C in the same frame of reference. A flat 80-by-50-centimeter panel can fit within a 60-by-90-centimeter opening after rotation.
A connection is the only link between two parts of a network; removing it disconnects those parts. Every small box is inside a larger container, and that container is inside the warehouse, so the small boxes are inside the warehouse.
What to check: Verify the reference frame and the properties of the relation. “Is taller than” is transitive; “is a friend of” need not be. A diagram can suggest a conclusion that its stated relationships do not justify.
Central question: “How do interactions, dependencies, and delays produce behavior over time?”
Temporal reasoning tracks ordering, duration, and dependencies. Systems reasoning extends this to interacting parts, feedback, accumulation, bottlenecks, and delayed effects. System dynamics explicitly studies how feedback, time delays, and nonlinear responses shape outcomes. (MIT OpenCourseWare)
Examples. A reservoir receives eight units of water per minute and loses ten, so its level continues to fall even if inflow recently increased from six to eight.
Two tasks taking three and five hours run in parallel, followed by a two-hour review; the earliest completion is seven hours, not ten. In a heating model with delayed responses, continuing to add heat before earlier heating takes effect can produce overshoot. Speeding up an upstream process does not increase total output when a downstream stage remains the binding bottleneck.
What to check: Distinguish levels from rates, local improvement from whole-system improvement, and immediate effects from delayed effects.
Central question: “Given my goals and constraints, what should I do?”
Practical reasoning concerns action. Decision-theoretic approaches organize choices around possible outcomes, uncertainty, and how those outcomes are valued. Expected-utility reasoning is one influential framework, but monetary value and utility are not identical. (Stanford Encyclopedia of Philosophy)
Examples. You must arrive before 9:00. A cheaper route arrives at 9:10, so it fails the constraint; a route arriving at 8:50 remains feasible.
A five-unit precaution eliminates a 2% chance of a 1,000-unit loss. Under an expected-cost objective and those assumptions, it reduces expected cost from 20 to five. Two options are identical on every relevant dimension except price, so you select the cheaper one. A plan that performs reasonably under several plausible scenarios is preferred to one that works extremely well only under a fragile forecast.
What to check: Clarify the objective before optimizing. A flawless calculation can recommend the wrong action when important goals, costs, or constraints are omitted.
Central question: “What should I do when others will also choose, anticipate, and respond?”
Strategic reasoning concerns interdependent choices: your outcome depends on what others do, while their decisions may depend on what they expect from you. Game theory provides formal tools for studying such interactions. (Stanford Encyclopedia of Philosophy)
Examples. A seemingly attractive move in a game leaves a valuable piece undefended, so you reject it after considering the opponent’s reply.
A business predicts that competitors will match a price reduction; the reduction may then lower margins without creating the expected market-share gain. Two roommates agree on quiet hours because uncoordinated choices interfere with both people’s work. A manager notices that rewarding the number of closed cases encourages employees to split cases artificially, so the incentive needs redesign.
What to check: Do not treat other people as fixed features of the environment. Examine their information, incentives, capabilities, and opportunities to adapt.
Central question: “What is justified, fair, permissible, or obligatory?”
Normative reasoning evaluates what ought to be believed or done; ethical reasoning focuses on moral considerations. It can involve consequences, rights, duties, needs, fairness, character, and relationships. Different ethical frameworks may assign these considerations different weights. (Stanford Encyclopedia of Philosophy)
Examples. You promised to help someone. Treating promises as genuine obligations gives you a reason to help even when a more enjoyable activity becomes available.
If several applicants have equally strong claims to a scarce opportunity, a lottery may be justified by a principle of equal treatment. A project would increase efficiency but violate a stipulated requirement for informed consent, giving a reason not to proceed. An accommodation gives people different resources but may be justified by different needs.
What to check: Make the evaluative premises explicit. “This produces more output” does not settle whether it is fair. Likewise, describing what people commonly do does not by itself establish what they ought to do.
Central question: “What does this statement or action mean in its context?”
Interpretive reasoning uses context to infer meaning, intention, perspective, or significance. In communication, understanding an utterance can require more than decoding its literal words; it may require reasoning about what the speaker knows and what they are trying to convey. (Stanford Encyclopedia of Philosophy)
Examples. Someone says, “It is cold in here,” while looking at an open window. You interpret this as a possible request to close it, not merely a temperature report.
A speaker says, “Some reports have arrived,” in response to a question about whether all have arrived; you tentatively infer that not all have. A writer calls a plainly disastrous lunch “a triumph,” suggesting irony. A colleague who has not seen a revised deadline may be uninformed rather than deliberately uncooperative.
What to check: Distinguish plausible interpretation from certainty about another person’s mind. Consider alternative meanings and the information actually available to that person.
These are methods that can operate across the preceding categories.
Central question: “What do these terms mean, and which assumptions does the argument actually require?”
Conceptual reasoning clarifies meanings and distinctions. First-principles reasoning works back toward basic assumptions, definitions, or constraints and rebuilds from them. Analysis can involve both clarification and tracing a position back to its underlying principles. (Stanford Encyclopedia of Philosophy)
Examples. Revenue and profit are different quantities; high revenue therefore does not establish profitability. “Equal treatment” and “equal outcomes” are different concepts, so an argument that substitutes one for the other needs repair.
Instead of assuming the task is “heat the whole room,” you identify the underlying goal as “keep the occupant warm,” opening other design possibilities. Instead of accepting a quoted production cost as fundamental, you examine material requirements, labor, energy, and unavoidable overhead.
What to check: Separate genuine constraints from conventions and preferences. Labeling an assumption a “first principle” does not make it true or exempt it from evidence.
Central question: “How can I understand the whole through its parts—and reconstruct the whole from them?”
Decomposition separates a complex problem into components. Composition combines component-level results while accounting for their relationships. Breaking a whole into parts is a central conception of analysis, but reconstruction requires attention to how those parts interact. (Stanford Encyclopedia of Philosophy)
Examples. Sales revenue is represented as visitors × purchase rate × average order value. If visitor count and order value are unchanged while purchase rate halves, this model explains a halving of revenue.
Application delay is separated into data retrieval, processing, and display time, revealing where most time is spent. A journey is divided into travel legs and transfer waits, then recombined into a feasible itinerary. A system’s components each meet their specifications, but the overall design still requires checking that their interfaces are compatible.
What to check: Ask what was lost in the decomposition. Reliable parts do not automatically imply a reliable whole when coordination, interfaces, or shared dependencies matter.
Central question: “What satisfies all the requirements, and what procedure will find or verify it?”
Constraint-based reasoning represents possibilities together with conditions they must satisfy. Algorithmic reasoning specifies a sequence of operations for finding, constructing, or checking an answer. Constraint-satisfaction problems formalize variables, their possible values, and restrictions on their combinations. (EECS Instructional Support Group)
Examples. Two appointments must occur on different days. Appointment B can occur only Wednesday, while A can occur Tuesday or Wednesday; therefore, A must be Tuesday.
Three simultaneous meetings require separate rooms, but only two rooms are available, so the proposed schedule is infeasible. A puzzle solver eliminates candidate values as constraints rule them out. Working backward from a 17:00 deadline, three hours of production followed by one hour of review requires starting no later than 13:00.
What to check: Distinguish finding one feasible solution from proving that a solution is optimal—or that no solution exists. Also verify that the formal constraints capture the real requirements.
Central question: “Does this argument withstand serious examination and the strongest relevant objections?”
Critical reasoning examines premises, inference quality, evidence, ambiguity, and alternatives. Dialectical reasoning, in the sense used here, develops a position through objections and replies rather than considering only the case in its favor. (Stanford Encyclopedia of Philosophy)
Examples. Someone claims that every proposed change reduces cost. One genuine counterexample is enough to defeat that universal claim.
A survey includes only current customers, so you challenge its use as evidence about why former customers left. Two supposedly independent reports repeat the same original account, so you do not count them as two independent confirmations. Two people disagree about a “fair” allocation; examining their arguments reveals that one means equal shares and the other means shares proportional to need.
What to check: Apply scrutiny symmetrically to favored and disfavored conclusions. The aim is not to produce an objection at any cost, but to identify which objections genuinely change the argument’s strength.
Central question: “How should I manage my own reasoning?”
Metacognition monitors and regulates thinking. Metareasoning includes deciding which reasoning strategy to use, how much effort to invest, what additional information to seek, and when to stop. Research on rational metareasoning explicitly models trade-offs between the costs of thinking and its expected benefits. (PLOS)
Examples. A high-impact calculation receives an independent check because a small chance of error has large consequences. A spatial problem is difficult to track verbally, so you draw a diagram.
Two options appear nearly tied, and one inexpensive measurement could reverse their ranking, so you obtain it. Further research cannot plausibly change the decision, so you stop. You compare your past confidence judgments with outcomes to assess whether your confidence needs recalibration.
What to check: Distinguish feeling certain from having well-supported confidence. Also distinguish useful reflection from repeatedly reconsidering an issue without gaining information.
Suppose a team introduces a new support process. Average recorded resolution time falls from 30 minutes to 20 minutes. Should the team adopt the process permanently?
Different reasoning families address different parts of the question:
| Reasoning family | Contribution to the investigation |
|---|---|
| Conceptual | Define “resolved.” Does it mean that the customer’s problem was solved, or merely that the case was closed? |
| Deductive | If the stated sample-level target is an average below 24 minutes, the recorded average of 20 meets that numerical target. |
| Inductive | Assess whether the observed cases support expecting similar performance on future cases. |
| Abductive | Compare explanations: the process helped, the cases became easier, staffing improved, or recording practices changed. |
| Causal | Design a comparison that can distinguish a process effect from those alternative explanations. |
| Statistical | Examine sample size, variability, and uncertainty around the estimated improvement. |
| Systems | Check whether faster initial handling transfers work or delay to another team. |
| Practical and ethical | Weigh implementation costs, customer outcomes, employee workload, and the treatment of difficult cases. |
| Metacognitive | Decide whether existing evidence is sufficient or whether another trial is worth its cost. |
The initial observation—“the average fell”—does not answer all these questions. Better reasoning comes from identifying which question remains unresolved, rather than repeatedly applying the same kind of inference.
Some commonly named “types of reasoning” are better treated as subtypes, combinations, or additional dimensions.
Scientific reasoning combines hypothesis generation, deduction of predictions, observation or experimentation, and evaluation of evidence. A particularly important pattern is hypothetico-deductive reasoning: propose a hypothesis, derive what should follow under auxiliary assumptions, and compare those predictions with observations. A failed prediction challenges the hypothesis-and-assumptions package; it does not always identify which part failed. (Stanford Encyclopedia of Philosophy)
Legal reasoning is a domain-specific combination that can include rule application, interpretation, analogy with precedents, and evaluation of evidence. “Legal” identifies the application domain, not a single inference pattern. (Stanford Encyclopedia of Philosophy)
Heuristic reasoning uses simplifying strategies. Such a strategy should be evaluated in relation to the environment and problem: a shortcut can be effective under some conditions and unreliable under others. It is not automatically irrational merely because it avoids exhaustive analysis. (Stanford Encyclopedia of Philosophy)
For this curriculum, I would likewise place working backward, inversion, and proof by contradiction under problem-solving techniques, rather than treating them as peers of deduction and induction. For example, “What conditions would make this plan fail?” is an inverted question; the reasoning used to answer it might be deductive, causal, statistical, or strategic.
First, learn to distinguish deduction, induction, abduction, and analogy. Take one situation and formulate all four kinds of argument about it. This forces attention to the inferential move rather than the subject matter.
Next, study uncertainty and alternatives. Practice default reasoning, probability, causal inference, and counterfactuals. For every conclusion, identify an exception, competing explanation, or assumption that could change it.
Then, vary the representation and purpose. Revisit problems using quantities, diagrams, timelines, interacting systems, decision models, and ethical considerations. Ask what each representation reveals and what it leaves out.
Throughout, practice constructing and criticizing examples. For each category, create a strong example, a superficially convincing but weak example, and a minimally changed example in which the conclusion is no longer justified. Explain exactly what changed.
The central habit is:
Identify the inferential move, state its assumptions, and match the strength of the conclusion to the strength of the support.