Purpose: This paper aims to contribute to the futurology of a possible artificial intelligence (AI) breakthrough, by reexamining the Omohundro–Bostrom theory for instrumental vs final AI goals. Does that theory, along with its predictions for what a superintelligent AI would be motivated to do, hold water?Design/methodology/approach: The standard tools of systematic reasoning and analytic philosophy are used to probe possible weaknesses of Omohundro–Bostrom theory from four different directions: self-referential contradictions, Tegmark’s physics challenge, moral realism and the messy case of human motivations.Findings: The two cornerstones of Omohundro–Bostrom theory – the orthogonality thesis and the instrumental convergence thesis – are both open to various criticisms that question their validity and scope. These criticisms are however far from conclusive: while they do suggest that a reasonable amount of caution and epistemic humility is attached to predictions derived from the theory, further work will be needed to clarify its scope and to put it on more rigorous foundations.Originality/value: The practical value of being able to predict AI goals and motivations under various circumstances cannot be overstated: the future of humanity may depend on it. Currently, the only framework available for making such predictions is Omohundro–Bostrom theory, and the value of the present paper is to demonstrate its tentative nature and the need for further scrutiny.
An artificial general intelligence (AGI) might have an instrumental drive to modify its utility function to improve its ability to cooperate, bargain, promise, threaten, and resist and engage in blackmail. Such an AGI would necessarily have a utility function that was at least partially observable and that was influenced by how other agents chose to interact with it. This instrumental drive would conflict with the strong orthogonality thesis since the modifications would be influenced by the AGI’s intelligence. AGIs in highly competitive environments might converge to having nearly the same utility function, one optimized to favorably influencing other agents through game theory. Nothing in our analysis weakens arguments concerning the risks of AGI.