Relevance guides effective action

The authors show that relevance determines which information and actions are meaningful in a given context, and that without it systems cannot select what is pertinent. This implies that decision makers must clearly define the spaces of semiotic action so that results are useful.

Amid the rise of language models, the question arises of what really sets human thought apart from machines. Rauterberg's paper in Discover Artificial Intelligence argues that relevance and the assignment of meaning are essential properties of any intelligent act, and that without them neither humans nor artificial systems can select what is pertinent in a given context.

To develop this thesis, the author contrasts formal and computational languages with natural language, showing that the former, by their syntactic nature, cannot express reasons of belonging or justify why an element is part of a whole. He draws on examples of combustion engines and parliamentary institutions to illustrate how relevance answers the question of why something makes sense within a whole, and argues that this capacity for justification is what current AI lacks when it limits itself to manipulating symbols with no reference to a semiotic context.

The argument is aimed at the reductionist view that equates brain processes with digital computation and assumes that increasing processing power is enough to reach general intelligence. It points out that, even with a practically unlimited number of devices, building a general AI remains out of reach as long as the spaces of semiotic action where operations acquire meaning are not explicitly defined. The tension remains open between the need to formalize those spaces so that machines can operate on them and the risk that formalization eliminates precisely the dimension of relevance it is meant to capture.

Defining those spaces of semiotic action means that those who lead content or technology product projects must clearly specify the purposes, contexts of use and criteria of pertinence that guide the selection of information and actions, so that the results the systems generate are useful and not merely statistical.

Conclusions

• Relevance is necessary for a system to select pertinent information and act appropriately.
• Formal languages, by their syntactic nature, cannot express the notion of relevance.
• Building useful intelligent systems requires specifying concrete, bounded spaces of semiotic action.
• Human limits of time, computing power and communication shape what is considered relevant in each situation.