»Destructive Alignment« – or: Will AI Make us Smarter?

Abstract Whether AI makes us smarter depends entirely on what we mean by smart. The question is caught in a reflexive loop: how we measure intelligence shapes what intelligence becomes, which in turn shapes how we measure it. Educational frameworks like »constructive alignment« (CA) embody this tension by ensuring students learn what can be measured. The problem becomes acute with GenAI: the one thing that cannot be measured in education is also something GenAI cannot do – what Plessner called »[Sehen] mit anderen Augen«, the capacity to step outside one’s evaluative framework and re-evaluate it. GenAI can, at best, make us »smarter« only within fixed frameworks, but not in the capacity that grounds all intelligence: transcending and evaluating frameworks themselves. When empirical studies measure whether AI makes us smarter using constructive alignment metrics, they engage in what I call »destructive alignment« (DA), i.e. constructive from the standpoint of current measurement, but destructive from the standpoint of genuine intelligence.