The 19th meeting of the dialogue circle "Coaching meets AI" focused on the so-called alignment problem of artificial intelligence (AI). This issue is regarded as one of the central challenges in AI research. It concerns ensuring that the goals and decisions of AI systems align with human values, intentions, and ethical standards. The key points and discussions from the meeting can be summarized as follows:
-
Definition of the Alignment Problem:
- AI might misunderstand or ignore human values if they are inaccurately defined.
- A well-known thought experiment, the "paperclip maximizer," illustrates how AI, under radical goal orientation, can make harmful decisions.
- Human values are situational, context-dependent, and not entirely algorithmizable.
-
Difference Between Human and AI Rationality:
- Instrumental Rationality: AI’s ability to pursue goals efficiently is based on clearly defined specifications and data analysis.
- Normative and Value-Based Rationality: Human values are abstract and situational, fundamentally beyond AI's comprehension. Concepts like "human dignity" cannot be fully grasped or interpreted by machines.
-
Coaching and the Alignment Problem:
- Human coaching also faces an alignment problem:
- First level: The values of the coach and coachee may diverge.
- Second level: The coachee's conscious goals may not align with their unconscious needs.
- In AI-assisted coaching (triadic coaching), the human coach must continue to handle value examination and goal alignment.
- Human coaching also faces an alignment problem:
-
Solutions and Challenges:
- Approaches like "Human-in-the-Loop" (triadic coaching) emphasize the necessity of human control and oversight in AI applications.
- Transparency in AI systems must be improved to better understand how decisions are made.
- The discussion clarified that values cannot be fully formalized, and responsibility for value coherence must remain with the human coach.
-
Historical and Cultural Context:
- AI is seen as the fourth stage of knowledge transmission (after language, writing, and printing). However, it differs fundamentally as it is not only a carrier of information but also an actor and decision-maker.
- Historically, values and norms have been passed down through grand narratives. The group discussed how AI is embedded in existing cultural narratives and how this might influence its decisions.
-
Reflection and Future Perspectives:
- Participants expressed interest in delving deeper into the alignment problem and exploring possible future scenarios for AI and superintelligence.
- A central thesis was that AI will never surpass all aspects of human intelligence, as it is limited to instrumental rationality.
- The meeting concluded with a reflection round where participants emphasized the significance of the topic for coaching and society and encouraged further discussions on the ethical and practical aspects of AI applications.