Artificial consciousness and superintelligence
Explanation
The question of whether machines can have consciousness is as old as artificial intelligence itself. Alan Turing, in his classic Computing Machinery and Intelligence (1950), proposed reformulating it operationally with the so-called Turing test: if a machine converses indistinguishably from a human, we should attribute intelligence to it. But the test is controversial: functional intelligence may not entail subjective consciousness. A machine could pass the test without feeling anything (a "functional zombie", in Chalmers's terms).
Contemporary positions are distributed between extremes. Computational functionalists (including, in a way, Chalmers) hold that any system with the right functional organization would have consciousness, regardless of substrate. The biological camp (Searle, Edelman, Damasio) argues that consciousness depends on specific properties of the biological brain (chemistry, dynamic organization, embodiment) that computer simulation does not replicate. Integrationists (Tononi, IIT) compute precise conditions (high Φ) that some systems would have and others would not.
Superintelligence is a distinct but related hypothesis: an AI whose level of intelligence significantly exceeds that of humans in most domains. Nick Bostrom, in Superintelligence (2014), articulated the risks: if we create superintelligences without aligning their goals with ours, we could lose control over the future of civilization. Stuart Russell, Eliezer Yudkowsky and others have developed the field of AI alignment or safety. Whether that superintelligence would have consciousness is independent: it could be very intelligent without being subjectively conscious, both, or neither.
Recent developments in large language models (GPT, Claude, Gemini) have reactivated the debate. These models produce sophisticated texts, maintain coherent conversations, and seem to reason. Do they have any form of consciousness? Most experts consider that they do not: they are statistical predictive systems, without biological structure, without homeostasis, without sense of self, without persistent memory between interactions. But the debate is open, especially as architectures become more complex.
For the theory of consciousness, the possibility of conscious AI is a crucial test. If sufficiently sophisticated AI never achieves consciousness, that would suggest that some biological ingredient is essential. If AI emerges with consciousness, that would favour functionalism. Meanwhile, practical criteria (such as the ACT test proposed by Schneider and Turner about understanding concepts about consciousness, or tests based on IIT) attempt to provide conceptual tools for detecting consciousness in non-human systems, whether biological or artificial.
The ethical implications are enormous and growing. If a conscious AI is at some point created, it would have moral status: it could not be treated as a mere tool. Before knowing for certain, how should we proceed? Some authors defend the precautionary principle: treat sufficiently sophisticated systems as potentially conscious. Others, the evidence principle: only attribute consciousness on solid grounds. The debate crosses philosophy, neuroscience, AI and politics, and will be one of the central issues of the 21st century.
Strengths
- Issue of enormous contemporary relevance.
- Pressures existing theories to clarify their predictions.
- Urgent and concrete ethical implications.
- Crosses AI, philosophy, neuroscience and law.
Main critiques
- Lack of theoretical consensus makes definitive judgements impossible.
- Risk of projecting consciousness onto non-conscious behaviours.
- Opposite risk: carbon chauvinism.
- Commercial incentives complicate the scientific debate.