Abstract

The article presents a theoretical study that highlights applicability boundaries for the contemporary AI algorithms. Past the boundaries the technology presents an existential threat to humanity. A discussion how to extend the safety margin concludes the article. In particular, the article analyzes various AI alignment classes, which are differentiated by algorithmic principles. The applicability constraints are being considered. To quantify the phenomenon, AI alignment limits are compared against cognitive task complexity and mapped onto the same scale. It reveals safe operations ranges for the algorithmic approaches. Another insight is that the AI alignment limits are forming a distinct data row. An improved alignment criterion is being proposed as a result of extrapolation. Respectively, a new class of AI alignment is being identified. It resembles being failsafe for all actual cognitive tasks. An algorithm feature to implement the alignment class is proposed.