Can Consciousness Emerge from Training? Limits of Current Practice
Modafar Al-Shouha
Zenodo (CERN European Organization for Nuclear Research) May 13, 2026 preprint DOI: 10.5281/zenodo.20156857 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Argues that current AI training methods lack the evolutionary pressures—real-world stakes, self-maintenance, open-ended environmental change, and energy limits—that the authors propose were necessary for consciousness to emerge in animals, and suggests introducing these pressures into AI training to enable more grounded predictions about artificial consciousness. |
Abstract
Debates about AI consciousness usually focus on external behavior, even though subjective experience cannot be directly observed. This proposal takes a different angle. Instead of trying to measure consciousness in current models, we look at the types of evolutionary and developmental processes that could actually create it. We argue that how we train AI today is only a distant echo of the biological conditions that led to consciousness in animals. Standard approaches—like supervised learning, reinforcement learning, and continual learning—lack the key evolutionary pressures necessary for emergence. Specifically, these models do not face real-world stakes, self-maintenance requirements, open-ended environmental changes, or strict energy limits. We conclude by suggesting new research directions that would introduce these biological pressures into AI training, helping us make more grounded predictions about the potential for artificial consciousness.