By: Colin Caret

Valid inference

Remember from Chapter 1. Logic and AI that logic is the study of valid inference. In this chapter, you’ll learn more about the concept of validity.

In particular, we’ll go into more details of what it means for an inference to be correct, we’ll describe abstract methods for modeling valid inferences, and we’ll dig into important differences between deductive and inductive validity.

Correctness

When we talk about validity, we are talking about a good feature of inferences, but this is not the only good feature an inference can have. For example, it is good for inferences to be simple, clear, precise, economical, etc. Logic does not deal with all of these topics. Most of these topics are part of rhetoric, the study of persuasive writing style. Logic deals with validity.

Validity is a standard of correctness for inferences. To really fix this concept, it might help to think about it from the other direction. What happens if an inference lacks validity, when it is invalid? Well, that shows us that something went wrong. The inference made a mistake. As we mentioned in Chapter 1.1.2 Validity , some of these logical mistakes or fallacies are so famous that they have their own names.

Here is an example of a fallacy called affirming the consequent:

If it is sunny, Jan is cycling. Jan is cycling. Therefore, it is sunny.

The mistake should be clear. The conditional premise (“if…, then…”) says that sun leads to sport. However, the other premise of this inference is not about sunny weather, so those two premises don’t really “add up” to anything useful. They certainly do not support the conclusion that it is sunny today.

Here is an example of a valid inference called “modus ponens”.

If it is sunny, Jan is cycling. It is sunny. Therefore, Jan is cycling.

This inference involves a premise about sunny weather, which connects with the conditional premise in the right way. This inference is valid because it uses all of its premises correctly. These two examples show that it can be easy to confuse valid and invalid inferences if we don’t pay attention to the details. At a glance, the two inferences look pretty similar.

This is why logic aims for a systematic definition of validity. This notion should be applicable not only to humans but to any information-processing system. It can tell us what counts as intelligent behavior. Without a definition of correct reasoning, how do we even know what we want AI to achieve? So, we would like to say, in general, what makes an inference valid or invalid. The standard idea is that valid inferences preserve truth from their premises to their conclusion.

Hypothetically

To test whether an inference is valid, we ask if the conclusion is true when the premises are true. This is a hypothetical question. Answering this question does not require us to know that the premises really are true. In fact, we can make a stronger point: it is possible to have an inference that is valid even though it has false premises. Here is an example.

All pigs fly. Maddy is a pig. So, Maddy flies.

The reasoning behind this inference is perfectly correct. The conclusion follows from the premises. That makes the inference valid. But we obviously know that this inference also involves some false premises. Pigs can’t really fly.

The important point is that validity is not determined by the actual truth or falsity of statements. What we care about is the connections between statements.

When we test for validity, we do not look at the actual truth of premises and conclusion, instead we look for a relationship between their truth-values. You can think about it like this: imagine that the premises and true and think about whether the conclusion is true under this assumption.

Think about inference (3) this way. When we do this, we imagine a world that is slightly different from the actual world, a world where pigs do fly. In that kind of world, it has to be true that Maddy flies.

This is the basic idea of truth-preservation.

Valid Inference: an inference whose conclusion is true under the (hypothetical) assumption that all of the premises are true.

Now, we can ask a series of follow-up questions. How tight is the truth-preservation relationship? How often does it have to hold? How reliable does an inference need to be in order to call it “valid”? Different answers to these questions take us in two directions: deductive and inductive logic.

Always

With deductively valid inference, truth-preservation always holds. We want 100% reliability in all situations whatsoever. No exceptions allowed. Necessarily, if the premises are true, the conclusion is true. Otherwise, its not deductively valid. This is a high standard we are asking for but it is very nice when we can identify this kind of air-tight reasoning.

Here are some common examples of deductively valid inferences.

If this superintelligent AI system is dangerous, then it gives a lot of bad advice. It does not give a lot of bad advice. So, this superintelligent AI system is not dangerous.
Either it rained last night or the car is dry. The car is not dry. Thus, it rained last night.
All dogs are friendly. Some dogs are chubby. So, some chubby animals are friendly.

In deductive logic, we use the symbol $$\vDash$$ to stand for a deductively valid inference. So if we have a deductively valid inference going from premise $P_1,P_2,\dots$ to conclusion $C$, we can abbreviate this with symbols:

$$P_1,P2,\mathellipsis \vDash C$$

In inference 5, for example, we have that $P_1$ is the sentence “Either it rained last night or the car is dry”, $P_2$ is “The car is not dry”, and $C$ is “It rained last night”. So, we can write the inference as:

$$\text{Either it rained last night or the car is dry}, \text{The car is not dry}\vDash \text{It rained last night}$$

Mostly

With inductively valid inference, truth-preservation mostly holds. If the premises are true, the conclusion is probably true. The likelihood could be slightly increased (weak induction) or greatly increased (strong induction). Either way, assuming that the premises are true gives us some reason to believe the conclusion is true. Inductive inference is not air-tight. There is a chance of going from truth to falsity, but that is a risk we have to take when we are dealing with uncertain information.

Here are some common examples of inductively valid inferences. In each of these examples, the premises give us good reason to believe the conclusion, but none of these ‘risky’ inferences qualifies as a deductively valid inference (all of them are deductively invalid).

It looks like a duck. It quacks like a duck. So, it is a duck.
Strict AI laws have a support of around 80% in a randomly selected sample of 20.000 voters. Therefore, support for strict AI laws in the general public is around 80%.
GPT-4 improved upon GPT-3, which improved upon GPT-2, which, in turn, improved upon GPT-1 in terms of coherence and relevance. Therefore, the next generation of GPT models will further improve in this respect.

In inductive logic, we use the symbol $$\mid\approx$$ to stand for an inductively valid inference, writing $$P_1, P_2, \dots\mid\approx C$$ to say that the inference from premises $P_1, P_2, \mathellipsis$ to conclusion $C$ is inductively valid. To say that an inference is inductively strong, we write a $!$ on top, like so: $$\stackrel{!}{\mid\approx}$$. So if we have a inductively strong inference going from premises $P_1, P_2, \dots$ to conclusion $C$ we can abbreviate this with symbols: $$P_1, P_2, \dots\stackrel{!}{\mid\approx} C.$$

Formalization

By steps, we can make things even more precise. One of the steps we mentioned in Chapter 1. is formalization. We can build formal languages and study their properties. We can also develop theories about how these kind of formal patterns map onto ordinary human communication and thinking.

One powerful idea in logic is that different inferences have the same logical form. This allows us to identify many valid inferences all at once, just by studying their shared form. Reasoning that can be captured in a finite set of symbols and rules can also be programmed into a computer. This was essentially the idea that sparked the first wave of AI research in the 1950s.

The next chapter will delve into more details of formal syntax and how to interpret the logical forms of sentences. We will mention just a few examples of formalisation that are useful for this chapter.

The symbol “$\neg$” can be read “not” in English. $\neg A$ says “Not $A$.”
The symbol “$\land$” can be read “and” in English. $A\land B$ says “Both $A$ and $B$.”
The symbol “$\rightarrow$” can be read “if, then” in English. $A\rightarrow B$ says “If $A$, then $B$”.

Let"s see how this formalisation is used:

The inference form of “modus ponens” looks like this: $ A\rightarrow B, A\vDash B$. This sequence of symbols says that whenever you put together two separate pieces of information, a premise “If $A$, then $B$” and a premise $A$, you can validly infer the conclusion $B$. We said that example 2. follows this pattern. Here are some more examples. These are all instances of “modus ponens”:
1. If programming is difficult to learn, then some people cannot do it. Programming is difficult to learn. Therefore, some people cannot do it.
2. If ELIZA ever beat GPT-3.5 in the Turing Test, then it is cutting-edge AI technology. ELIZA did beat GPT-3.5 in the Turing Test. So, ELIZA is cutting-edge AI technology.
3. If the sun will rise tomorrow, then you will win the lottery next week. The sun will rise tomorrow. So, you will win the lottery next week.
In logic, we represent inferences 10.-12. the same way. They have the same form but different content. As we will show later, the form is enough to explain why all of these inferences are valid. Obviously, this does not guarantee that the conclusion is actually true!
The fallacy of “affirming the consequent” looks like this: $ A\rightarrow B, B\nvDash A$. This sequence of symbols says that if you put together two separate pieces of information, a premise “If $A$, then $B$” and a premise $B$, it is invalid to infer the conclusion $A$. If you use this reasoning, you are making a mistake. We said that example 1. commits this mistake. Here are some more examples:
1. If ZFC is consistent, it cannot settle the value of BB(8000). ZFC cannot settle the value of BB(8000). So, ZFC is consistent.
2. If Strong AI is possible, then there is a flaw in the Chinese Room thought experiment. There is a flaw in the Chinese Room thought experiment. Therefore, Strong AI is possible.
3. If Topsy was an elephant, then Topsy had a trunk. Topsy had a trunk. So, Topsy was an elephant.
Again, we can represent 13.-15. the same way and this common structure can explain why all of these inferences are invalid. This means that they use reasoning. It does not necessarily mean that all of these inferences have false conclusions! For example, the conclusion of 15. is true, but it would be a mistake if you came to believe this conclusion by using this fallacious reasoning.

We have now said a lot about the concept of validity. But what do the tools of logical analysis look like? How do we use logical form to identify whether an inference is valid? That is our next topic.

Deduction

The study of deduction goes back to ancient philosophy and mathematics. Aristotle considered it to be the pinnacle of reasoning. Euclid used deductive reasoning to prove things about basic geometry. He also introduced the axiomatic method. This method has two parts: axioms and rules. An axiom is just a definition.

The first Euclidan axiom says “between any two points there is a line”, which partly defines the abstract concepts of point and line by telling us how these things relate to each other. The other part of this method is a set of rules for reasoning. In other words: logic.

Rules for deductive logic are indefeasible and suited to belief accumulation. These rules do not vary between contexts. They do not require special justification. There is nothing that can make these rules fail. Nothing can defeat a deductively valid inference.

The optimal way of using deduction is to take a set of existing beliefs (or knowledge) and then add more beliefs (or knowledge) by applying logical rules. A common, real world implementation of deductive reasoning occurs whenenever we apply some general pattern to a specific case:

Vixens are female foxes. That is the definition of the concept vixen. Suppose you know this and you hear someone say “there is a vixen living in the forest!”. In that case, you might use deduction to infer that there is a fox living the forest.
Suppose that you live in a village where the bus does not run on days when there is a football match. You know about this policy. You also know that there is a football match today. If you put together these existing beliefs, you might use deduction to infer that there is no bus running today.

Like the examples of “modus ponens”, we have here two examples of inferences with the same logical form. It is actually closely related to “modus ponens” but also involves the quantifier “all” or “every”. In example 16. there is an implicit premise “all vixens are female foxes”. In example 17. there is an implicit premise “all days with footbal are days without the bus”.

Both of these inferences simply apply general pattern to a specific instance. That reasoning is correct as long as the word “all” really means “all”. If we assume that a totally general pattern is true “all Ps are Qs” and “this is P” is also true, it has to be true that “this is Q”. The conclusion follows necessarily.

The concept of deductive validity is specialized. It makes sense to use deductive reasoning for specific tasks: reasoning in mathematics or other axiomatic theories, reasoning with precisely defined concepts or pattern that are truly general. Notice how this air-tight reasoning is not quite the same thing as what Sherlock Holmes calls “doing a deduction”:

“From a drop of water… a logician could infer the possibility of an Atlantic or a Niagara without having seen or heard of one or the other. So all life is a great chain, the nature of which is known whenever we are shown a single link of it. Like all other arts, the Science of Deduction and Analysis is one which can only be acquired by long and patient study…"

A study in scarlet, 1887

Sherlock uses the word “deduction” for any reasoning that is careful, systematic, and reliable. However, the examples in this passage sounds a lot like inductive reasoning. Take a small sample and make an educated guess about the larger collection that it belongs to. That kind of reasoning is not indefeasible, it does not make a necessary connection. We do not call that deduction.

Many programming languages are essentially formal languages with deductive rules. This is a perfect way to apply logical methods. Consider the code snippet below. This code is supposed to take an input that is a whole number and identify whether it is a positive number or not:

num = int(input('Enter a whole number: '))

if num > 0:
     print(f'{num} is positive.')
else:
     print(f'{num} is not positive.')

Imagine what would happen if the computer did, effectively, not obey deductive rules like “modus ponens”. We would have no idea what to expect when we run this program. Sometimes when you ran the program and entered the number 1, you might get the correct answer that 1 is a positive number, but other times you might you get no answer at all. That would be useless and frustrating. In order to make the behavior of programs predictable, we want them to effectively follow deductive rules of reasoning.

Semantic methods for deduction

Let’s now take a brief look at semantics. This is the part of logic where we give a model of meaning and truth. A semantics for a language delineates which inferences are deductively valid in that language. To keep things simple we will talk about a stripped-down formal language that only represents the logical form of “and” sentences like “$A$ and $B$”, represented $A\land B$.

Even with a stripped-down formal language like this, we can learn a lot about deductive logic. The basic tool we will use is called a semantic model. A semantic model is like a picture of a possible reasoning scenario. If we look at a model, we can ask whether a specific sentence $A$ is true or false in that model. Every model gives an answer. It assigns a definite truth-value to each sentence.

A formal language $\mathcal{L}$ has more than one model. Different models assign different values to the same sentences. In one model perhaps $A$ is true and $B$ is false. In another model $A$ is false and $B$ is true. That variation allows models to picture different scenarios. Once we have defined all possible models for a language $\mathcal{L}$, how do we use them?¹

For any sentence $A$ in our formal language, we will write $[A]$ to refer to the set of models where that sentence is true. With this notation we can talk about relations between sets of models $[A]$ and $[B]$ for any two sentences $A$ and $B$. When the elements of one set are also contained in another set, we say that there is a subset relation between them.

For example, the set of triangle is a subset of the set of figures. The set of cats is a subset of the set of animals. All of the things in the first set are also in the second set. If $[A]$ is a subset of $[B]$, we write $[A]\subseteq[B]$.

This relation is a good first step to defining deductive validity:²

$$A\vDash B \text{ if, and only if, } [A]\subseteq[B]$$

Why does this definition work? Let’s analyze what this says. In the background we assume that there is well-defined notion of all possible models for our language. From there, we define $[A]$ and $[B]$ for any two sentences $A$ and $B$. Suppose that $[A]\subseteq[B]$ holds. That means for absolutely any model $\mathcal{M}$ of our language, if $\mathcal{M}$ makes $A$ true, then $\mathcal{M}$ also makes $B$ true. That means that an inference from premise $A$ to conclusion $B$ would preserve truth in every possible scenario. It would always preserve truth. It is necessarily truth-preserving. And that is exactly the concept of deductive validity $\vDash$.

The approach we have just described is very abstract. It can be used for any formal language. If we can explain how to define all possible models for a language, the definition of deductive validity just ‘falls out’ of that collection of semantic models.

To give an illustration of this method, let’s look at some inferences with “and” sentences. When we write $A\land B$ it is supposed to represent a single complex sentence that combines two simpler thoughts. For example, we could formalise English sentences like this.

$P$ formalises “Putnam is happy.”
$Q$ formalises “Quine is happy.”
$P\land Q$ formalises “Putnam and Quine are both happy.”

If a person asserts $P$ and then asserts $Q$, they make two separate statements. If a person asserts $P\land Q$, they make one single statement that has a complex, inner structure made from a combination of simpler statements with the logical word “and”. Semantics can model the meaning of this logical word as an operation on sets of models. The relevant operation is called set intersection:

In the diagram, we see that there is a set $X$ and there is a set $Y$. Maybe these are sets of people or sets of numbers. It doesn’t really matter. We just want to see what the intersection of two sets refers to. The intersection $X\cap Y$ is the overlapping part of those two sets. The elements of $X\cap Y$ are all of the things that are shared in common between $X$ and $Y$.

For example, if $F$ was the set of all fruits and $O$ was the set of all orange things, then $F\cap O$ would be the set of orange fruits. This intersection set contains mandarines, apricots, and mangoes but it does not contain bananas (not orange) and it does not contain carrots (not fruits).

We can apply this operation to any two sets. What if we apply it to $[A]$ and $[B]$? What does $[A]\cap[B]$ refer to? For absolutely any model $\mathcal{M}$ in the intersection set $[A]\cap[B]$, that same model $\mathcal{M}$ has to make $A$ true and $\mathcal{M}$ has to make $B$ true. This operation essentially takes the models of two separate sentences and cuts them down to the set of models that make both true at once. It is natural to identify the meaning of “and” with this operation: $$[A\land B] = [A]\cap[B].$$ This gives a very precise meaning to ‘$\land$’ which we call the logical conjunction.

Using these methods, we can establish some useful facts about valid inferences. The way we verify these facts is to just think carefully about the definitions of formal symbols like ‘$\land$’ and ‘$\vDash$’. In other words, we are going to put the axiomatic approach into action. We take definitions of basic concepts and see where they lead. In this chapter, we will only sketch the explanations or proofs of these facts.

$A\land B \vDash A$

Why is this form of inference valid?

Sketch of an Explanation: $[A\land B]=[A]\cap[B]$ and since the intersection of two sets has to be part of each set, we know that $[A]\cap[B]\subseteq[A]$.

If $A \vDash B$ and $A \vDash C$, then $A \vDash B\land C$

This says that if there are two valid inferences using the same premises, you can also use those premises to validly infer a conjunction of the previous conclusions.

Sketch of an Explanation: Assuming that $A \vDash B$ and $A \vDash C$ hold we have $[A]\subseteq[B]$ and $[A]\subseteq[C]$. So if any model $\mathcal{M}$ is in $[A]$, the same model $\mathcal{M}$ also has to be in $[B]\cap[C]$. We have $[A]\subseteq[B]\cap[C]$.

If $A \vDash C$, then $A\land B \vDash C$

This principle says that if it is valid to infer a conclusion from some premises, then it is valid to infer the exact same conclusion after ‘adding new information’ to those premises.

Sketch of an Explanation: We showed that $A\land B \vDash A$. So if we asume $A \vDash C$ then we have both $[A\land B]\subseteq[A]$ and $[A]\subseteq[C]$. This implies $[A\land B]\subseteq[C]$ because the subset relation is transitive: if first set is part of second, and second set is part of third, the first one has to be part of the third.

The last fact is called monotonicity and it gets to the heart of what makes deduction special. If there is an element of your beliefs that implies a conclusion, then as long as you hold on to that original element your beliefs will always imply the same conclusion – no matter how much new information you aquire! This is why we say that deductive rules are indefeasible.

John McCarthy was famously optimistic about the power of logical methods. He even thought we could use them to capture the kind of commonsense reasoning that most adults are capable of doing. This led to the field of knowledge representation where engineers try to formalise vast amounts of information. These are like axioms of a massively complicated theory, but the method has its limitations.

Induction

The study of induction also has roots in ancient philosophy and psychology. The Buddhist scholars, and brothers, Asaṅga and Vasubandhu focused on inferences like “where there is smoke, there is fire”. They considered associative thinking, tracking regularities in experience, and drawing structural analogies to be the cornerstones of real human belief-forming practices. This was an attempt to describe how the mind really functions to help us navigate a world where evidence is imperfect.

Rules for inductive logic are defeasible and suited to belief modulation. These rules can vary between contexts. They often require special backing. These rules can fail, or perhaps a better way to put it is that an inductive rule can be better in some environments, worse in others.

The early work on logic-based AI quickly shifted from deductive logic to inductive logic. The main reason is that induction is immensely important to a complete description of commonsense reasoning.

Suppose that you agreed to meet Karl for lunch in a few minutes. Karl eats lunch every day in the canteen and he always gets there early. You infer that Karl is in the canteen now. This is inductively valid. However, as you are walking to the canteen you see that the building is on fire and the occupants are standing on the pavement watching the fire fighters put out the blaze. You take back the inference and the conclusion you previously drew. Now, you infer that Karl is not in the canteen (at least, you hope not).

The optimal way of using induction is to take a set of existing beliefs (or knowledge) and then tentatively entertain the beliefs (or knowledge) that are most likely to be true based on your initial information. Some common, real world implementations of inductive reasoning occur when we use observations of similarities to infer the existence of a general pattern, or we apply so-called generic information.

I tallied thousands of swans around the river Rhine. I tallied thousands of swans around the IJsselmeer. Every swan that I observed was white. So, I infer that all swans are white.
Birds fly. Tweety is a bird. Thus, tweety flies.

The example of Tweety is a famous example of perfectly good induction, but one that obviously ‘goes wrong’ in some contexts. Generic information like “birds fly” is very different from a truly general pattern. It is not talking about absolutely every bird. It means that most birds fly, typical birds fly. So if you apply this generic information to Tweety, then your conclusion is probably true. But this conclusion could be false if it turns out that Tweety is a penguin or an ostrich.

Semantic methods for induction

One way to approach semantics for inductive logic is to introduce probabilities into the picture of semantic models for deductive logic. The idea is that if we can talk about the probabilities of the premises and conclusion in our semantics, we can directly translate the idea of the premises making the conclusion (more) likely into our model. We will illustrate these ideas by focusing on a stripped-down formal language that only represents the logical form of “and” sentences.

Assume that we have access to absolutely all possible models of our formal language, just like before with deductive logic. Remember from section 2.3.1 Semantic methods for deduction that for a sentence $A$, the notation $[A]$ denotes the set of all the possible reasoning scenarios where $A$ is true. Similarly, we have the same definition of logical conjunction $$[A\land B] = [A]\cap[B]$$ within this background set of all models.

In this setting, we introduce the probabilities as numeric measures of how likely it is that a given sentence is actually true.³ For a sentence $A$, we write $$Pr(A)$$ to denote this measure. You can think of $Pr(A)$ as a measure of how likely it is that the way things actually are is among the possibilities $[A]$ according to which $A$ is true.

There are a number of laws that probabilities satisfy. For example, the value of $Pr(A)$ must always be between $0$ and $1$ (inclusive): $$0\leq Pr(A)\leq 1.$$ We’ll study these laws later in the book, when we return to probabilities and inductive reasoning. For now, a simplified interpretation of the numbers $Pr(A)$ works just fine.

For now, you can think of $Pr(A)$ as measuring the proportion of all possible scenarios that are $A$-scenarios. So, if $A$ is the sentence “Robots can fly”, $Pr(A)$ measures how many scenarios of all scenarios are ones where robots can fly. Assuming that all scenarios are equally likely, then if robots can fly in half of all scenarios, then $Pr(A)=0.5$. And if robots can fly in $\frac{3}{4}$ of all possible scenarios, then $Pr(A)=0.75$. And so on.

Now that we have gotten some idea of probabilities in this way, we can talk about conditional probabilities: how likely it is that something’s true assuming that something else is true. Remember from above that validity if a hypothetical concept. So to properly capture inductive validity, we need a notion of hypothetical probability.

Mathematically, we write $Pr(B|A)$ for the conditional probability of $B$ given $A$. Intuitively, $Pr(B|A)$ is a measure of how likely it is that $B$ is true assuming that the actual scenario is an $A$-scenario. That is, if $A$ is the sentence “Pigs can fly”, for example, and $B$ is “horses can fly”, then $Pr(B|A)$ asks how likely it is that horses can fly, assuming that pigs can fly.

But how should we define this mathematically? It turns out that assuming the simplified interpretation of probabilities as proportion of scenarios from before gives us a very natural answer. The idea is to say that : $$Pr(B|A)=\frac{Pr(A\land B)}{Pr(A)}.$$ Basically, what this definition says is that $Pr(B|A)$ is a measure of the proportion of $A$-scenarios that are $B$-scenarios (look at the intersection diagram to see this immediately).

With this out of the way, we get a very natural model of inductive validity by saying that: $$ P_1,P_2,\dots\mid\approx C \text{ if, and only if, } Pr(C|P_1\land P_2\land \dots)\geq Pr(C)$$ In words, the inference from $P_1,P_2, \dots$ to $C$ is inductively valid just in case the conditional probability of $C$ given $P_1,P_2, \dots$ is higher than the probability of $C$ not assuming the premises. Rudolf Carnap coined the term “increase of firmness” for this relation.

In fact, in this model we can also measure how much the premises raise the probability of the conclusion. This is simply $Pr(C|P_1\land P_2\land \dots)-Pr(C)$. Using this idea, we can define a strong inductive inference as one where the conditional probability of the conclusion given the premises is much larger ($\gg$) than that of the conclusion: $$P_1, P_2, \dots\stackrel{!}{\mid\approx} C \text{just in case }Pr(C|P_1\land P_2\land \dots) \gg Pr(C)$$

Let’s think about this model for inductive validity. Using it, we can establish some useful facts about inductive inferences. As before, we will only sketch the explanations or proofs of these facts. We start off with some commonalities between inductive and deductive logic.

For example, we can easily show that:

$A\land B \mid\approx A$

Sketch of an Explanation: This is inductively valid because the proportion of $A\land B$-scenarios which are $A$-scenarios is $1$: all $A\land B$-scenarios are $A$-scenarios. That is $Pr(A|A\land B)=1$. Since $0\leq Pr(A)\leq 1$, we know that $Pr(A|A\land B)\geq Pr(A)$.

Question: is the inference always inductively strong?

But, crucially, we get that:

It is possible to have $P_1,P_2,\dots \mid\approx C$, but not $Q,P_1,P_2,\dots\mid\approx C$.

This says that monotonicity fails for induction. In other words, it can be valid to infer a conclusion from certain premises, but invalid to infer the exact same conclusion after “adding new information” to those premises. New information changes the conclusion we can inductively infer.

Sketch of an Explanation: Suppose you live in a town with a few vegans. You know that all those people ride bicycles, they do not own cars.

In the diagram below, $Pr(A)$ is the probability of being a cyclist, $Pr(B)$ is the probability of eating meat, and $Pr(C)$ is the probability of being vegan:

For a randomly chosen person, $Pr(C)$ is not very high since the proportion of vegans, $C$, among the general population is low.

At the same time, $Pr(C|A)$ is relatively high, of the cyclists, $A$, a much larger proportion is vegan than in the general population. This means that $Pr(C|A) > P(C)$, and so $A\mid\approx C$. Or, in words, a randomly chosen person being a cyclist is inductive evidence for that person being vegan.

But if we look at the proportion of bikers that are cyclist and meat eaters, obviously, none of them is a vegan–they’re meat eaters, after all. This means that $Pr(C| A\land B)=0$. But since $Pr(C)$ is positive—there are some vegans, this means that $Pr(C| A\land B)\ngeq Pr(C)$ and so $A,B\mid\approx C$ fails.

The non-monotonicity of induction is very important. Perhaps a certain element of your beliefs implies a conclusion at this moment, but if you acquire new information then your updated beliefs might not imply the same conclusion anymore. This is why we say that inductive rules are defeasible. At one moment, in one context, it might be inductively good to infer a conclusion based on some limited information. At the next moment, based on new information, it is no longer inductively good to infer that same conclusion. You have to ’take back’ the previous ‘outcome’ of your inferences.