AI improves learning performance – but destroys what education is really about. Why future skills remain difficult to teach despite AI.

The new meta-analysis by Wang & Fan (2025), based on 51 studies, shows: ChatGPT has a strong effect on learning performance (g = 0.867) as well as moderate effects on learning perception (g = 0.456) and higher order thinking (g = 0.457). At first glance, it's a positive result – yet, from a psychological perspective, it reveals a structural dilemma.

The authors recommend employing ChatGPT specifically as a tutor, partner, or tool over four to eight weeks to achieve stable effects. But this is precisely where the problem starts: When it comes to complex thinking and deeper understanding, the effects remain limited. Why?

Higher order thinking (such as reflection, transfer, or synthesis) is 𝐧𝐨𝐭 a function of information processing but a psychologically demanding act: metacognitive, ambiguity-tolerant, emotionally involved. Such skills are not developed through reactive responses to prompts but through active thinking, friction, and self-direction. This cannot be standardised using AI.

In addition, the studies show a high heterogeneity (I² = 89.2% in learning performance), which means the benefit strongly depends on the 𝐜𝐨𝐧𝐭𝐞𝐱𝐭: course type, learning model, role of AI. Without a 𝐝𝐢𝐝𝐚𝐜𝐭𝐢𝐜𝐚𝐥-𝐩𝐬𝐲𝐜𝐡𝐨𝐥𝐨𝐠𝐢𝐜𝐚𝐥 𝐟𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤, the effect of ChatGPT is hard to control. Paradoxically, the very skills that AI aims to foster, namely critical thinking and self-regulation, are those that are necessary to use the technology effectively.

The third, often overlooked point is the 𝐩𝐬𝐲𝐜𝐡𝐨𝐥𝐨𝐠𝐢𝐜𝐚𝐥 𝐢𝐧𝐞𝐫𝐭𝐢𝐚 of educational cultures. Innovations are introduced quickly on a technical level but are scarcely integrated mentally. Transforming one's learning attitudes (towards more autonomy, reflection, and intrinsic motivation) requires more than tools. It needs leadership that is itself capable of learning.

The numbers sound good, yet they obscure the real problem: The crucial competencies for the future are not technical, but psychological. They cannot be 'installed' but must be developed. Ignoring this will see us overwhelmed by the AI-driven learning revolution; not because it is too powerful, but because we are internally unprepared.
𝐁𝐞𝐜𝐚𝐮𝐬𝐞: 𝐓𝐡𝐨𝐬𝐞 𝐰𝐡𝐨 𝐨𝐮𝐭𝐬𝐨𝐮𝐫𝐜𝐞 𝐦𝐞𝐭𝐚𝐜𝐨𝐠𝐧𝐢𝐭𝐢𝐨𝐧 𝐥𝐨𝐬𝐞 𝐭𝐡𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨 𝐟𝐨𝐫𝐦 𝐣𝐮𝐝𝐠𝐞𝐦𝐞𝐧𝐭𝐬. 𝐖𝐞 𝐨𝐩𝐭𝐢𝐦𝐢𝐬𝐞 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐨𝐮𝐭𝐩𝐮𝐭 - 𝐚𝐧𝐝 𝐢𝐧 𝐝𝐨𝐢𝐧𝐠 𝐬𝐨 𝐥𝐨𝐬𝐞 𝐭𝐡𝐞 𝐡𝐮𝐦𝐚𝐧 𝐚𝐬𝐩𝐞𝐜𝐭.