The artificial language Lojban was not expressly designed to be used by machines; it (or rather its antecedent Loglan) was designed as a test of the Sapir–Whorf hypothesis, its overt basis in predicate logic being sufficiently alien that its inventor thought it would serve the purpose.
Lojban is something of a kitchen sink language in its design, but its design has several aspects which are appealing to at least some AI enthusiasts:
- It has spoken syntactic brackets, and it can be parsed syntactically by an LALR parser (defined in Yacc); so its syntax as formally specified is unambiguous.
- It is also morphologically unambiguous, at the cost of some restrictive phonotactics. (I’m seeing that loanword phonotactics are less restrictive than they were in my day.) So a stream of phonemes can be broken up into morphemes only one way.
- It was a well-elaborated list of 1300-odd basic predicates, with their arguments fully specified. It has prepositions supporting a full case grammar, and (unofficially) conventions for deriving compound predicate arguments from their components. (I was involved in the latter.) This does not quite make its semantics as unambiguous as they’d like, but it certainly puts it on a very formal footing.
Syntactic and morphological ambiguity are not the big challenge of natural language processing; stats tends to take care of that. Semantics is always sloppier, but I’m not sure that the new generation stats-mongerers are that fussed about formal semantics either. But yes, Lojban has been attractive to several AI people for that reason. Ben Goertzel, who is on here, has been vocal about this; see e.g. Aspects of Artificial General Intelligence.