← Back to Autonomy

A Plain-Language Field Guide · Part 1 of 7

Logic and Reasoning

Part I · The Foundations and the Human Mind

What logic is, what reasoning is, the many kinds of reasoning we use including cause and analogy, how to build a sound argument, and why the human mind does all of it in its own peculiar way.

Section 01

What is logic?

Logic is the study of what makes a conclusion genuinely follow from what came before it.

Imagine someone tells you two things: every dog is an animal, and Rex is a dog. You instantly know a third thing without being told it: Rex is an animal. You did not check on Rex. You did not need to. The conclusion was already locked inside the first two statements. Logic is the set of rules that explain why that move is allowed, and why other moves that look similar are not.

The single most important idea in logic is this: it cares about form, not content. A good argument keeps working even when you rip out the words and pour in new ones, as long as the shape stays the same. That shape is called the form of the argument.

Interactive · swap the content, keep the form

The skeleton below never changes. Press the button to pour different words into the same three slots and watch the argument stay valid no matter what.

All dogs are animals.
Rex is a dog.
Therefore Rex is an animal.
FORM: All A are B · x is A · therefore x is B  |  VERDICT: valid ✓

This is why logicians often write arguments with letters instead of words. Stripped to All A are B; x is A; therefore x is B, the argument is obviously airtight, and you can see that the truth of the conclusion does not depend on what A, B, and x actually mean.

Validity is not the same as truth

A subtle but vital point: an argument can be perfectly valid (correct in form) while being completely false in the real world. "All fish can fly; a salmon is a fish; therefore a salmon can fly" is a flawless piece of logic built on a false starting point. Logic guarantees that if your starting points are true, your conclusion must be true. It makes no promises about the starting points themselves. Checking those is someone else's job.

In one line Logic is the grammar of good thinking: it tells you which moves preserve truth, regardless of what you are talking about.

The main layers of formal logic

Logic is not one thing but a stack of systems, each adding expressive power:

Layer 1

Propositional logic

Combines whole statements with and, or, not, and if-then. "If it rains, the match is cancelled."

Layer 2

Predicate logic

Adds "all" and "some" and lets you talk about properties of things. "Every student passed."

Layer 3

Modal logic

Adds necessity and possibility. "It is possible that it rains; it is necessary that 2+2=4."

Layer 4

Informal logic

Studies real arguments in ordinary language, including the fallacies that trip people up.

Layer 1 in your hands: the truth table

Propositional logic is built from a tiny number of connectives, and each one is fully defined by a table of every possible case. Pick a connective below and read off exactly what it does. The only inputs are whether P and Q are true or false.

Interactive · truth-table builder

Layer 2 in your hands: drawing the logic

Predicate logic talks about "all" and "some." Centuries before symbols, people checked these arguments by drawing circles. Each argument below is a real categorical syllogism. Watch how the picture either forces the conclusion or quietly leaves a gap.

Interactive · categorical logic, drawn
A B C
Section 02

What is reasoning?

Reasoning is the actual mental act of moving from what you already believe to something new.

Here is the cleanest way to hold logic and reasoning apart in your head:

The key distinction Logic is the standard. Reasoning is the activity. Logic describes how inference ought to work to keep truth intact. Reasoning is what a real mind actually does, which may live up to that standard or fall flat. You can reason badly. Logic is the ruler you hold up against your reasoning to find out.

Every act of reasoning, good or bad, has the same basic anatomy. You begin with some premises (things you already accept), you apply some kind of inference (a mental move), and you land on a conclusion (something new you now accept). The diagram below is the engine that runs underneath every section that follows.

PREMISE 1 what you accept PREMISE 2 what you accept INFERENCE the mental move CONCLUSION something new today's conclusion becomes tomorrow's premise
Figure 1. The anatomy of any inference. The dashed loop is what makes a mind grow: conclusions are fed back in as fresh premises, so reasoning compounds over time.

The whole drama of reasoning lives in that middle teal box. The kind of mental move you make there is exactly what separates the types of reasoning we turn to next. Some moves come with an ironclad guarantee. Most do not, and trade certainty for the ability to say something genuinely new.

Section 03

The types of reasoning

There are a handful of distinct mental moves. Knowing which one you are making tells you how much to trust the result.

The three classical families are deductive, inductive, and abductive reasoning. A few others (analogical, causal, Bayesian, defeasible) are best understood as specialised cousins of these. Tap each tab to explore one, and watch the certainty meter: it is the single most useful thing to track. The first move guarantees its conclusion; the rest only make it more or less likely.

Interactive · the three core moves
From general rule to certain instance

Deductive reasoning

You start with a general rule and apply it to a specific case. If the rule and the facts are true, the conclusion cannot be false. This is the only family that comes with a guarantee. It is also "tight": it never tells you anything that was not already implied.

Rule: every car with a dead battery will not start.
Fact: this car has a dead battery.
↳ Therefore this car will not start. (necessarily)
Certainty of conclusion100% · guaranteed
From many cases to a general expectation

Inductive reasoning

You generalise from repeated observations to a broader pattern or a future case. It can tell you genuinely new things (it is "ampliative"), but the conclusion is only ever probable. A single counterexample can break it. This is the engine of science and of everyday expectation.

Observation: this engine has logged a fault every cold morning for 90 days.
Observation: tomorrow morning will be cold.
↳ Therefore it will probably log a fault tomorrow. (likely, not certain)
Certainty of conclusion~85% · strong but defeasible
From an observation to its best explanation

Abductive reasoning

You see a surprising fact and reach for the hypothesis that would best explain it. This is the logic of diagnosis, debugging, and detective work. It is the most creative and the least safe: a better explanation may always be waiting. Confirming it takes further evidence.

Observation: the floor is wet this morning.
Best available explanation: a pipe leaked overnight.
↳ Therefore a pipe probably leaked. (best guess, pending checks)
Certainty of conclusion~60% · plausible, needs testing

The specialised cousins

Cousin

Analogical

Carries a pattern from a case you understand to one you do not. "The heart is like a pump, so it can fail like a pump."

Cousin

Causal

Works out what makes what happen, separating real causes from things that merely travel together.

Cousin

Bayesian / statistical

Updates how confident you are in a belief as new evidence arrives, using probabilities rather than yes or no.

Cousin

Defeasible

Reaches reasonable conclusions that you stand ready to take back. "Birds fly," until you meet a penguin.

The trade-off that runs through everything Only deduction is certain, and certainty costs you: it can never go beyond what you already knew. Every other kind of reasoning buys you new knowledge by giving up the guarantee. A useful mind spends its day mostly in the uncertain kinds and reserves deduction for tying things down.

The Bayesian cousin, made concrete

Bayesian reasoning is just the disciplined way to update a belief when evidence arrives. The classic surprise: when something is rare, even a good test that "fires" usually fires falsely. Frame it as a diagnostic alarm. A fault is uncommon in the fleet, your detector is mostly accurate, the alarm goes off. How worried should you be? Drag the sliders.

Interactive · updating a belief with evidence
Belief BEFORE the alarm (base rate)2%
Belief AFTER the alarm fires27%

The defeasible cousin, made concrete

Most everyday conclusions are held loosely. You accept them now and stand ready to drop them the moment a better fact shows up, all without any contradiction. Logicians call this non-monotonic: adding information can remove a conclusion rather than only adding to it. Watch a belief get withdrawn cleanly.

Interactive · a conclusion that retracts
Default rule: birds fly.
Known: Tweety is a bird.
↳ Conclusion: Tweety flies.
Section 04

One scene, three inferences

The clearest way to feel the difference: point all three kinds of reasoning at the same fact and watch them behave differently.

The scene: you walk outside and the grass is wet. That one observation can launch any of the three moves, and each answers a different question.

OBSERVATION the grass is wet DEDUCTIVE rule says: sprinkler on ⇒ grass IS wet (certain) INDUCTIVE wet every morning so far ⇒ wet again tomorrow (likely) ABDUCTIVE what best explains wet? ⇒ it rained (best guess)
Figure 2. Same fact, three directions. Deduction runs from a known rule down to a guaranteed result. Induction runs from a track record forward to a likely future. Abduction runs from the fact backward to the cause that would best account for it.

Notice that deduction reasons forward from a cause you already know to its certain effect. Abduction reasons backward from an effect to a likely cause. Induction reasons across time, from a pattern to its continuation. They are not better or worse than one another. They are answers to three different questions, and a skilled thinker keeps track of which question is on the table.

Section 05

Reasoning about cause

One kind of reasoning is so useful, and so easy to get wrong, that it deserves its own treatment: reasoning about cause. It is the kind your instinct reaches for first, and trusts far too much.

The trap is famous. When two things move together, there are three possible reasons, and the data alone cannot tell you which. One may cause the other, some third thing may cause both, or it may be coincidence. A vibration reading and a bearing fault that rise together might mean the wear drives the vibration, or that a load change drives both, or nothing at all. Watching, by itself, never settles it.

Judea Pearl organised causal questions into three rungs of a ladder, each harder than the one below and each demanding more than observation. Climb it.

Interactive · the ladder of causation
Pick a rung to see the question it answers and what it takes to climb there.
1 · SEEING association what is likely? 2 · DOING intervention what if I act? 3 · IMAGINING counterfactual what if I had not? harder; needs more than observation
Figure 3. The ladder of causation. Each rung answers a different question and needs more than the one below. Observation reaches only the first; acting and imagining require a model of what causes what.

This is why a pure classifier cannot do root-cause analysis. A classifier lives entirely on the first rung. It learns which patterns tend to accompany which faults from observation, and that is association, not cause. Ask it what to change to stop the fault and it has no ground to stand on, because that is a second-rung question and its training data never recorded an intervention. Detection can be statistical. Isolation and correction are causal. That is the half of diagnostics a correlation engine cannot reach, and the reason a fault-isolation system needs a causal model of the machine, not only a pattern matcher.

Why it matters To answer the questions that actually matter, what caused this and what will fix it, you need more than a model trained on observations. You need a causal model of the system, or the ability to run interventions, even if only in simulation. Correlation lines up the suspects; only intervention convicts.
Section 06

Other ways we reason

Deduction, induction, abduction, and cause are the headline modes, but the mind leans on a few others so constantly that we barely notice them.

Analogy: structure, not surface

Analogical reasoning maps the relations of a familiar situation onto a new one. What carries the inference is shared structure, not surface resemblance. The solar system maps onto the atom because the relations line up, a central body with smaller ones held in orbit, not because they look alike. This is how a solution is transferred across domains, as with the fortress and the tumour from the problem-solving section, and it is the basis of case-based reasoning, retrieving a past case whose structure matches the present one. Its danger is the inverse: a surface match hiding a different structure misleads with great confidence.

SOURCE: solar system sun planets orbit the sun TARGET: atom nucleus electrons orbit the nucleus the relations are mapped, not the looks
Figure 4. Analogy works by structure mapping: it carries over the pattern of relations, so it holds only as far as the relations match, not the surface details.

Conditionals and the hypothetical

Much reasoning runs under a supposition: if this, then that. But "if" hides several meanings. The indicative conditional ("if the valve is stuck, the pressure is high") describes the actual world. The counterfactual ("if the valve had not stuck, the pressure would be normal") describes a world that did not happen, and to evaluate it you need the causal model from earlier. Treating one as the other is a common and costly mistake, and the Wason selection task already showed how badly people test even a plain conditional.

Necessity, possibility, and time

Some claims are not merely true but necessarily true, and others merely possible. Modal reasoning handles must, might, and cannot. Temporal reasoning handles before, after, until, and always. These can look like luxuries, but they are the backbone of verification: a safety property is a modal and temporal claim, that nothing bad can ever happen, on any sequence of events, which is a far stronger thing to establish than that nothing bad has happened yet.

Why it matters When you reason by analogy, check the relations and not the looks. When you reason with conditionals, keep the indicative and the counterfactual apart. These quiet modes do much of the work, and most of their failures come from using the right machinery on the wrong kind of claim.
Section 07

How humans actually reason

Here is the twist: people do not reason the way the textbooks prescribe. We are brilliant practical reasoners and surprisingly poor formal ones.

If logic is so clean, why do smart people make logical errors all day long? Because the mind did not evolve to run a proof engine. It evolved to make fast, good-enough decisions in a messy world. Cognitive science has spent fifty years mapping the gap between how we ought to reason and how we do.

Two systems, one mind

The most influential picture, popularised by Daniel Kahneman, is that thinking runs on two gears. System 1 is fast, automatic, effortless, and intuitive. System 2 is slow, deliberate, effortful, and logical. System 1 handles almost everything. System 2 is lazy and only wakes up when forced. The trouble is that System 1 has confident answers even when they are wrong.

Interactive · catch your own System 1
A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball.
How much does the ball cost?

Heuristics and biases

To move fast, System 1 leans on mental shortcuts called heuristics. They are usually helpful, which is exactly why we keep them. But they misfire in predictable ways, and those predictable misfires are called cognitive biases. A few you have certainly experienced: confirmation bias (noticing evidence that fits what you already believe), availability (judging how common something is by how easily an example comes to mind), and anchoring (letting the first number you hear drag your estimate toward it). Amos Tversky and Daniel Kahneman documented dozens of these, and the catalogue keeps growing.

The famous one: the selection task

Nothing shows the gap better than Peter Wason's card puzzle from 1966. Try it yourself before reading on. Four cards each have a letter on one side and a number on the other.

Interactive · the Wason selection task
RULE: if a card shows a vowel on one face, then it shows an even number on the other.

Which cards must you turn over to check whether the rule holds? Click to select, then submit.

Eflip
Kflip
4flip
7flip

Only about one person in ten gets the abstract version right. Yet here is the punchline that changed cognitive science: take the identical logic and dress it as a social rule, and almost everyone solves it instantly. "If a person is drinking alcohol, they must be over 18." Which do you check? Obviously the person drinking, and the person who is 16. Same logical structure, effortless answer. The mind seems to carry special-purpose machinery for catching cheaters that it simply does not bring to abstract symbols.

We reason with models, not rules

Why the difference? One leading answer is Philip Johnson-Laird's mental models theory. It proposes that people do not apply formal rules at all. Instead we build a small concrete picture of the situation in our heads, imagine how it could play out, and read the answer off that picture. This explains both our strengths and our errors: when the picture is easy to build, we shine; when the situation has many invisible possibilities to juggle, we miss some and go wrong.

The honest summary Humans are superb context-sensitive, practical reasoners and rather clumsy formal logicians. Formal logic is not our native tongue. It is a cultural technology, invented and written down precisely because our built-in wiring is not reliable enough on its own. That is also why we build calculators, checklists, and reasoning machines: to shore up the parts our minds were never optimised for.

Are the shortcuts dumb, or smart?

The heuristics-and-biases tradition treats mental shortcuts as a catalogue of error. A second tradition, led by Gerd Gigerenzer, argues nearly the opposite: under real limits of time, information, and computation, fast and frugal heuristics are often not just adequate but optimal, and can beat elaborate models that overfit noisy, sparse data. A rule as simple as "go with the option you recognize" can outperform regression in the right setting. The resolution is that rationality is ecological. A rule is not good or bad in the abstract, only relative to the environment it runs in. The honest picture holds both views at once: the mind is neither broken nor perfect, but tuned, sometimes well and sometimes badly, to the world it evolved to handle.

Section 08

When reasoning breaks: the fallacy gallery

A fallacy is an argument that feels persuasive but does not actually deliver its conclusion. Knowing the common shapes is the cheapest upgrade to your thinking.

Fallacies are the informal-logic counterpart to bugs. Some are errors of form (the logic simply does not hold) and some are errors of relevance (true-sounding moves that dodge the real question). Tap each card to see an example and, more usefully, why it slips past us.

Affirming the consequent

+

"If the server is down, the page fails. The page failed, so the server is down."

Why it fools us: it borrows the shape of valid deduction but runs it backwards. The page could fail for many other reasons. This is a deductive error of form.

Ad hominem

+

"You can ignore her safety analysis. She has never even worked on the line."

Why it fools us: it swaps the claim for the claimant. Whether the analysis is right does not depend on who made it. An error of relevance.

Straw man

+

"You want more testing? So you want to ship nothing, ever." Then the exaggeration gets attacked.

Why it fools us: knocking down a distorted copy feels like winning. The original, weaker claim was never addressed.

Base-rate neglect

+

"The alarm is 95% accurate and it fired, so the fault is almost certainly real."

Why it fools us: it ignores how rare the fault is to begin with. The Bayesian widget above shows exactly how badly this reasoning fails when the base rate is low.

Correlation as cause

+

"Failures rose the same month we changed suppliers, so the supplier caused them."

Why it fools us: two things moving together is easy to see; a hidden third cause, or pure coincidence, is not. Timing is suggestive, not conclusive.

Slippery slope

+

"If we allow one manual override, soon no one will trust the automation at all."

Why it fools us: it treats a chain of "and then" steps as inevitable without showing any link is actually forced. Each step needs its own evidence.

Section 09

Building and judging arguments

The fallacy gallery showed how arguments fail. The harder and more useful skill is building one that stands, and knowing what to look for in someone else's.

Stephen Toulmin gave the anatomy of a real argument, well beyond the textbook pair of premise and conclusion. It has six parts. The claim is what you want accepted. The grounds are the data you offer for it. The warrant is the principle that licenses the leap from grounds to claim. The backing is what supports the warrant itself. The qualifier states how strongly the claim holds, probably or almost certainly or always. And the rebuttal names the conditions under which it would not hold. Most weak arguments are not outright false. They are missing the warrant, so the grounds do not actually reach the claim, or missing the qualifier, so they are stated with more confidence than the grounds permit, or missing the rebuttal, so they pretend nothing could defeat them.

GROUNDSthe data WARRANTwhy grounds support claim BACKING CLAIMqualifier: probably REBUTTAL: unless...
Figure 5. The anatomy of an argument. Grounds reach the claim only through a warrant, the warrant rests on its backing, the qualifier states the strength, and the rebuttal names what would defeat it.

A worked case in this shape: the grounds are a vibration signature, the claim is a bearing fault, the warrant is that this signature accompanies bearing wear, the backing is the physics plus fleet history, the qualifier is "probably," and the rebuttal is "unless a loose sensor mount is mimicking it." Laid out this way, the weak point is obvious and testable: confirm the warrant and rule out the rebuttal.

Why it matters To judge an argument, find the warrant and ask whether it truly licenses the leap, then look for the qualifier and the rebuttal. An argument that states no rebuttal is almost always more confident than its grounds deserve.
Continue · Part 2 of 7Limits, Machines, and Problem-Solving →The hard problems, mechanized reasoning, choosing under uncertainty and over time, and how humans solve problems.