Creation by Definition
Hereby I stipulate that "fb13" is to denote the first human born in the 13th century. Hence it might seems that "fb13 was born in the 13th century" is analytically true, true by definition. But if analytic truths are closed under logical implication, "somebody was born in the 13th century" would also be analytically true. Which it is not.
I don't think tinkering with closure under logical implication will help. Hereby I stipulate that "fb23" is to denote the first human born in the 23rd century. However, if recent progress in civilization continues, there might well be no humans in the 23rd century. And if no humans are born in the 23rd century, "fb23 is a human born in the 23rd century" is false. So it cannot be true by definition.
The upshot is that definitions of singular terms carry ontological commitment. They fail when the intended entities don't exist. This is rather obvious when this entity is supposed to be an ordinary, concrete, contingent being. It is less obvious when it is God, as in some versions of Anselm's famous proof, or when it is the 'Null Individal', which some mereologists introduce to turn formal mereology into a complete boolean algebra. (The Null Individual is defined to be a thing that is part of everything.)
An interesting case are mathematical entities. Consider Hume's Pinciple, which is Frege's partial definition of "the number of --":
HP) For all F,G: the number of Fs equals the number of Gs iff F and G are equinumerous.
Is this definition true by definition? If so, the existence of infinitely many numbers is also true by definition, since (HP) logically implies the Peano Axioms. But can the existence of numbers just be stipulated? I don't think so, though for some reason, in the abstract realm intuitions are much less clear. The only case where everyone agrees that some definition of an abstract term failed is where it is actually inconsistent (like Basic Law 5, Frege's partical definition of "the extension of --"). On the other hand, there are several reasons to believe that mathematical truth and consistency do not coincide. Two of them are Gödel's incompleteness theorems, another is that, say, the negation of the Peano Axioms is also consistent, but still false. Or so I think. (If you favour a structuralist interpretation of arithmetic, take the negation of the axioms of your favourite set theory instead.)
Another interesting case are properties. Hereby I stipulate that "Pb23" is to denote the property of being a human born in the 23rd century. This definition doesn't fail if nobody is born in the 23rd century. Properties may be empty. Could it fail simply because there is no such property as being a human born in the 23rd century? This sounds odd. But still, on any reasonable account of properties, failures like that are possible. For instance, if properties are universals and universals are sparse, it is rather likely that the definition of "Pb23" in fact failed. Even if properties are abundant, the inconsistency of Basic Law 5 illustrates that there cannot be a property corresponding to every condition on things. Let C be one of the conditions for which there is no property. There is no property of being a thing such that C. Hence it is futile to try to name that property.
Here is a passage from On the Plurality of Worlds (p.77), where David Lewis makes that mistake.
[W]hat about the relation of non-identity? [...] we may fairly deny it a place in our select inventory of the natural relations. It would be superfluous to include it if we have the resources to introduce it by definition; and so we do, since X and Y are non-identical iff there is a class that one of X and Y belongs to and the other does not.
What's wrong here is that, on Lewis' own account, there is no relation of non-identity: On this account, properties are classes, and relations classes of ordered tuples; but every class is non-identical with some other thing; so the relation of non-identity would have to contain every class as a member (or a member of ... a member); which is impossible -- at any rate impossible on Lewis' account. So there is no relation of non-identity, and the assumption that it exists leads to contradiction. Then how could we possibly "have the resources to introduce it by definition"? We don't. What Lewis goes on to define is not the term "non-identity", but the predicate "-- and -- are non-identical".
Definitions of predicates really do appear ontologically innocent to me. When I introduce a predicate for every human born in the 23rd century, or for every property not applying to itself, there is nothing that could possibly fail, since there is no ontological commitment in these definitions. (Is this intuition just due to Quinean indoctrination?) I would even prefer not to call alleged definitions of terms or functors "definition" at all. But then, other people write papers on "How to Define Theoretical Terms".
Doesn't the definition of a predicate applying to every human born in the 23rd century presuppose that there is a 23rd century? I'm not sure, but I hope not. Hereby I stipulate that for all x, x is a Bkof iff x is a brother of the present king of France. I don't think this stipulation failed. I think a predicate definition can fail only if it is inconsistent (or very ambiguous). The same is true for operators: Arthur Prior's definition of "tonk" fails because it is inconsistent.