Why do informal fallacies have to be detected by humans according to my textbook? It claims that Artificial Intelligence of any kind is unable to do this on its own.
Informal fallacies don't have exact definitions --they are just structural similarities between different poorly constructed arguments in natural language. They overlap with each other, there are edge cases where determining whether a fallacy is being committed is highly subjective, and they can manifest in widely different arguments. This isn't to say that a highly advanced real-world artificial intelligence (such as Watson) couldn't conceivably do a credible job, but would be at least as difficult as identifying any other fuzzy concept.
Formal fallacies, on the other hand, are exact, objective structural flaws in artificial, constructed languages. They are as easy for a computer to find as a mistake in a math calculation.
It might help if you quoted your textbook in more detail, but I take the point in part to be that informal fallacies are not inherently machine-checkable.
The key to understanding this is to understand both what a fallacy is and the distinction between formal and informal fallacies.
A fallacy is a mistake in argument or maybe by extension an error in logic in some slightly broader domain (like a computer program). These fallacies are then divided into formal fallacies and informal fallacies.
Here, the word formal has a technical meaning -- namely, that the form of the argument has an error of inference. Informal, by contrast, is any other error of reasoning that occurs at any stage in an argument.
To illustrate the point, the list of formal fallacies is quite short and easily checked by a computer. For propositional arguments, there are only three: affirming the consequent, denying the antecedent, and affirming a disjunction. In each case, the symbolization is assumed to be valid. And there's a pure error in application of the standard laws of logic.
By contrast, informal fallacies make mistakes in formalization and other aspects that make them difficult if not impossible to spot programmatically (it's not my domain but I would assume this issue is fundamentally identical to the difficulties faced by machine translation). Here, there's always a question of whether the formalization is in fact wrong.
Followers of Xenu might see statements by Xenu as axiomatic truths worthy of inclusion with the Peano axioms. But the rest of us would view that as an appeal to authority.
Pro-abortion people might call calling a fetus a human person an equivocation where as pro-life advocates would disagree.
The list of such disagreements about whether a fallacy is correctly applied is never-ending.
In theory, software can check for informal fallacies, just as in theory, software can simulate human beings. However, in practice this is an immense undertaking.
Our intelligence is built on a huge body of contextual information. Even detecting fallacies can call upon apparently unrelated pieces of information. So first, you'd need to encode all of this information at the right level of representation for the software. One attempt to do this (at a high level) was The Cyc Project.
Then you have to simulate the robust pattern matching abilities of the human brain. If you try doing this by simulating the brain, then you'll need to build a neural network with a quadrillion connections. Further, since individual neurons don't deal with concepts but with far finer grained inputs, your database would have to be much lower level.
Then there's the question of embodiement. Is it enough to simulate a brain? What does input, output and representation even mean? We're embodied beings, so must intelligence in the computer be somehow embodied? Rodney Brooks is a roboticist with this view and has built some robots and written eye-opening articles in this regard.
Once you've done that, you have to deal with possible side effects. For instance, is our incredible pattern finding ability one reason why we make mistakes? If so, being able to do what we do well may necessarily entail making the same types of mistakes that we do. So, would we have ended up with software that would make fallacies? Can we have our cake and eat it too?
One possible shortcut is to build this system with very little contextual information and expose it to the same inputs that people get in their formative years, in an attempt to "grow" this intelligence. Even that is full of problems, but at least you can bypass the problems of trying to encode higher level concepts in very low level neurological terms.
So while the question of fallacy checking is solvable in principle, it may be an error to treat it as an isolated act of intelligence, separable from the other acts of intelligence. Philosophy has a long history of trying to detach reason from bodily concerns, and more and more people are coming to realize that this may be a fallacy. Many conclude that reason cannot be separated from passions or even embodiement; in short, there may be no such thing as "pure detached reason" and one might have to take a holistic approach even to such apparently isolated domains as informal fallacy checking.
Your textbook is wrong.
To detect an informal fallacy requires intelligence. Whether this is natural intelligence or artificial intelligence doesn't matter much. An Artificial Intelligence that deserves the second part of the name has no problems detecting informal fallacies.
Your textbook may find arguments why creating an Artificial Intelligence may be difficult. I doubt that it can find valid arguments why creating an Artificial Intelligence would be impossible.