One of the biggest misunderstandings about AI is that it is simply a glorified Google. Why? Because that’s essentially how most people use AI most of the time. We’re not all AI researchers and scientists who might use AI to push beyond the known.
That “heavily accentuated search” is certainly part of the AI picture. Google now goes beyond just find pages containing an answer, even placing an AI-generated summary at the top. That AI-generated summary is usually good enough so you don’t need to peruse a number of pages to figure out your answer.
But real intelligence does something more fundamental. Intelligence makes inferences. It takes what is known, observed, remembered, or assumed and makes a move toward something that is not explicitly there.
You do this constantly. You see dark clouds and bring an umbrella. You hear a strange noise downstairs at night and wonder whether the cat knocked something over. You notice that one route to work has been congested three mornings in a row and try another route. You discover that the screw you need is missing and figure out that a paper clip might hold the thing together temporarily.
These are all forms of inference. In the age of AI, four words are especially good to know: inference, deduction, induction, and abduction. They describe much of what both human and artificial intelligence actually do. They are usually used interchangeably.
Pondering the differences between these words isn’t just an exercise is semantics. The differences reveals much about how we think. And how we can most effectively utilize (not submit to) AI.
Inference: Drawing a Conclusion
Inference is the broadest word. An inference is a conclusion reached from some collection of information.
Suppose you walk outside and see that the street is wet. You might infer that it rained. That conclusion was not contained in the observation itself. All you actually observed was a wet street. Your mind drew a conclusion from the wet street toward an explanation for it.
Maybe it rained. Maybe a sprinkler was running. Maybe a water main broke.
That move beyond the immediately observed facts is inference. Inference is not one particular method of reasoning. It is the larger category. Deduction, induction, and abduction are different ways of making inferences.
We infer a deduction and infer an induction. So rather than thinking of all four words as four siblings, it is better to think of inference as the family, with deduction, induction, and abduction as members of that family.
Deduction: What Must Be True?
Deduction is probably the easiest to recognize because it is the form most associated with formal logic—something we consciously do.
For example, we can start with two premises:
All mammals breathe air.
A whale is a mammal.
Therefore:
A whale breathes air.
Assuming the premises are true, the conclusion must be true.
The same structure appears everywhere:
| Premise | Premise | Deductive conclusion |
|---|---|---|
| The office is closed on Sundays. | Today is Sunday. | The office is closed today. |
| All mammals breathe air. | A whale is a mammal. | A whale breathes air. |
| Employees need a badge to enter the lab. | Maya does not have a badge. | Maya cannot enter the lab. |
| Any number divisible by 4 is even. | 28 is divisible by 4. | 28 is even. |
| All files in this folder are encrypted. | report.pdf is in this folder. | report.pdf is encrypted. |
| If the temperature is below freezing, this sprinkler system shuts off. | The temperature is 25°F. | The sprinkler system shuts off. |
| Anyone under 18 is not eligible for this contest. | Leo is 16. | Leo is not eligible for the contest. |
Deduction works from what we already accept toward consequences that logically follow from it. Computers are extremely comfortable with this kind of reasoning.
Consider a business definition: an account that has been inactive for more than 90 days is classified as dormant. If a particular account has been inactive for 120 days, we can deduce that the account is dormant.
This kind of reasoning appears throughout databases, software, rules engines, expert systems, theorem provers, and knowledge graphs. If you have ever programmed (and this is a Zen site from a software development view), deduction may already feel familiar because it resembles the ubiquitous if-then-else pattern. The connection is not exact—programming rules can also tell a computer to perform an action—but the logical structure is familiar.
For example:
If the temperature is below freezing or the ground is flooded, then the sprinkler system should not run.
We know that the temperature is below freezing.
Therefore:
The sprinkler system should not run.
Boolean logic gives us a familiar way to express the conditions. A condition such as below freezing OR ground flooded evaluates to true or false. Deduction takes the next step: given accepted premises and rules, what conclusion follows from them?
That is why deductive reasoning fits so naturally into software. Computers have spent decades evaluating Boolean conditions and following logical consequences from them. The same basic pattern appears in SQL predicates, business rules, expert systems, and many forms of symbolic AI.
Deduction asks: Given these premises, what must follow?
Induction: What Usually Seems to Happen?
Induction is probably closer to how we naturally learn about the world. The reason I put deduction first is that we are usually more conscious of deductive reasoning. Induction often happens quietly in the background. We notice repeated patterns, form expectations from them, and sometimes turn those expectations into explicit rules that we later use deductively.
A child does not begin with a formal rule saying that dropped objects fall toward the ground.
- The child drops a spoon. It falls.
- Drops a toy. It falls.
- Drops some food. It falls.
- After enough experience, a pattern forms: things I drop tend to fall.
That is induction. We observe examples and infer a broader pattern.
Then something interesting can happen. What began as an inductive pattern can become a premise or rule that we later use deductively. After repeatedly observing that unsupported objects fall, we may come to accept a rule such as:
Unsupported objects near the surface of the Earth fall toward the ground.
Now suppose I let go of a coffee mug. Given the rule and the fact that I have released the mug, I can reason deductively:
The mug will fall.
So induction and deduction are not isolated kinds of reasoning: We often learn a rule inductively and then reason from that rule deductively.
This is one reason the two can become confused. By the time we consciously use the rule, we may have forgotten that the rule itself originally came from repeated observations.
Induction asks: Given what I have observed, what pattern probably holds?
Deduction can then take that pattern, once we accept it as a premise, and ask: Given this rule, what must follow in this particular case?
Abduction: What Could Explain This?
Now we come to perhaps the most interesting and least known one. Abduction begins with an observation and asks what might explain it.
The grass is wet. Why?
One possibility is that it rained. But there are others. The sprinkler may have run. Someone may have washed the lawn furniture. A pipe may have broken.
We consider possible explanations and try to find the one that best fits what we see. That is abduction. And this is where Sherlock Holmes should come to mind.
Holmes famously talks about deduction, but much of what Holmes actually does is closer to abduction. He notices mud on someone’s shoes, a particular kind of ash, a scratch on a watch, a train schedule, or some strange behavior. Then he asks what chain of events could explain all of those observations together. He constructs a hypothesis. Then he looks for more evidence that should exist if that hypothesis is correct.
Abduction is not just making up stories.
Suppose you hypothesize that the murderer entered through the window. If that is true, there should probably be consequences. There may be footprints outside the window, damage to the latch, dirt on the carpet, or testimony that the window had previously been closed.
The proposed explanation points toward more things we should expect to find.
Abduction asks: What explanation could make these observations make sense?
And Then There Is MacGyver
Sherlock Holmes mostly uses abduction to reconstruct something that already happened.
MacGyver uses a related form of reasoning in another direction. MacGyver has a problem. There is no proper tool available. But there is some wire, a battery, a spring, a piece of metal, and duct tape. So he asks something like: what arrangement of these things could produce the effect I need?
He is not merely retrieving the official purpose of each object. A paper clip is supposed to hold paper. But under the right conditions, it can become a hook, an electrical contact, a pin, or part of some improvised mechanism.
This is an especially interesting form of abduction because the answer may never have existed before.
Holmes asks: What explanation could have produced these observations?
MacGyver asks: What configuration could produce the result I want?
We can call the first explanatory abduction and the second innovative abduction. One tries to discover a hidden explanation. The other tries to discover a hidden possibility.
That second form is one of the most creative things intelligence does.
Three Simple Questions
The three forms can be remembered with three questions.
| Type | Basic question | Example |
|---|---|---|
| Deduction | What must follow? | All birds have feathers. This is a bird. Therefore it has feathers. |
| Induction | What usually happens? | Every crow I’ve seen is black, so I expect the next crow to be black. |
| Abduction | What could explain this or make this possible? | The kitchen floor is wet. Maybe the dishwasher leaked. |
All three are forms of inference. Real intelligence constantly moves among them.
A doctor sees symptoms and proposes possible causes. That is abduction. Past experience helps show which causes are common. That is induction. Medical knowledge tells us that if a disease is present, certain other signs should appear. That is deduction.
Then new observations strengthen some explanations and weaken others. Round and round it goes.
That is much closer to intelligence than looking something up.
Why These Words Are Worth Knowing Now
When an AI tells us something, one good question is: what kind of inference did it just make?
Did it retrieve something it had seen before? Did it follow a rule from known premises? Did it generalize from patterns? Did it propose an explanation? Did it imagine a possible way to get from here to there?
With an LLM, there is an interesting answer hiding underneath all of this. At its foundation, an LLM is trained by learning patterns and predicting what comes next. Given the words that have already appeared, what word—or more precisely, token—is likely to come next?
That makes the basic mechanism of an LLM surprisingly close to induction.
It has encountered an enormous number of examples and learned extraordinarily rich patterns from them. From those patterns, it predicts what should come next in a new situation. Saying that an LLM is “just predicting the next word” understates what can emerge from such a rich model, but the statistical prediction is still there underneath it.
In fact, much of learning in general begins this way. A child encounters dogs and gradually learns what a dog is. We learn that stoves can be hot, that certain facial expressions signal anger, that a particular road is congested around five o’clock, that certain sounds in an engine mean trouble. Nobody initially hands us a book of formal rules. We accumulate experiences and begin noticing regularities.
Much of learned behavior is therefore inductive: this happened before under conditions like these, so something like it will probably happen again.
Humans, however, can take another step. We can turn a pattern we learned inductively into an explicit rule and then use that rule deductively.
After enough experience, a child may learn inductively that unsupported objects fall. Later that knowledge can become an explicit premise:
Unsupported objects fall toward the ground.
Now I release a cup.
Therefore:
The cup will fall.
The rule may have been learned inductively, but once we have the rule, we can reason deductively from it. This happens constantly. Experience becomes a rule. The rule becomes part of our knowledge. We then apply that rule to new situations.
And then there is abduction. Abduction is harder because it does not merely ask what usually happens or what follows from an existing rule. It asks us to propose something that is not yet known. Sherlock Holmes sees the evidence and asks what unseen sequence of events could have produced it. MacGyver looks at the objects around him and asks what arrangement of them could produce an outcome that does not yet exist.
AI systems can be built to perform forms of abductive reasoning, and modern LLMs can certainly generate plausible hypotheses. But this is where the question becomes especially interesting. Is the system merely recombining patterns it has learned, or has it actually found a new explanation or possibility?
Of course, humans are recombining things we have learned too. MacGyver cannot improvise with electricity if he knows nothing about electricity.
So perhaps the boundary is not nearly as clean as we would like it to be. That is exactly why these words are good to know now.
A retrieved fact may have a source. A deduction depends on whether its premises are correct. An induction depends on whether past examples really tell us what will happen next. An abduction depends on whether we have found a good explanation—or merely one explanation among many possibilities.
Real intelligence moves back and forth among all of them. We observe. We notice patterns. We turn some patterns into rules. We apply those rules. Something unexpected happens, so we invent possible explanations. We test those explanations against new observations. Those observations become more experience from which we learn new patterns.
It is a loop. That is inference. Once you know the words deduction, induction, and abduction, you begin seeing that loop everywhere: in science, troubleshooting, business, detective stories, AI, and ordinary thought.
Faith and Patience,
Reverend Dukkha Hanamoku
