A problem is considered AI-complete or AI-hard if solving it is equivalent to creating Artificial General Intelligence.
For example, natural language processing (or machine translation) is often considered AI-complete because understanding arbitrary language constructs seems to require broad general knowledge. The term was coined by the computer scientist Fanya Montalvo as an analogy with NP-complete, a class of problems in complexity theory. Problems labeled AI-complete like graceful degradation or computer vision tend to be framed at human-level intelligence; there may be many problems that AIs can solve that humans cannot. While mathematical formalizations of the class have been attempted, the term is usually used to communicate the qualitative difficulty of a problem.
Roman V. Yampolskiy proposed a classification in AI-Complete, AI-Hard and AI-Easy problems: