Can we classify AI as: subhuman, quasihuman, human, superhumam?

The Inadequacy of Linear Human-Centric AI Classification

The classification of AI into subhuman, quasihuman, human, and superhuman is fundamentally flawed because intelligence is not a linear spectrum. Current AI systems exhibit highly uneven capabilities—superhuman in narrow domains (e.g., AlphaZero in chess, AlphaFold in protein folding) while remaining subhuman in general cognition, common sense, and adaptability.

Key Flaws in the Linear Framework:

1. ·Multidimensionality of Intelligence: AI excels in specific tasks (e.g., arithmetic, game strategy) but fails in others (e.g., physical manipulation, social reasoning). A linear scale cannot capture this fractured performance.
2. ·Generality Gap: No AI today approaches human-level general intelligence (AGI). Systems like GPT-4 produce quasihuman outputs (e.g., coherent text) but lack understanding, intentionality, or world models.
3. ·Moral/Functional Conflation: Labels like "human" or "quasihuman" risk implying consciousness or agency, which are absent in current AI. Moral status (e.g., personhood) must be evaluated separately from functional performance.

A Better Framework: Capability vs. Generality

Instead of a linear hierarchy, AI should be assessed along two axes:

1. ·Capability: Task-specific performance (subhuman to superhuman).
2. ·Generality: Breadth of transferable skills (narrow to general).

Current AI resides in the high-capability, low-generality quadrant—superhuman in narrow benchmarks but unable to generalize or adapt like humans. For example, a chess engine cannot explain its moves to a child, and a language model cannot reason about physical causality beyond its training data.

Practical Implications:

• ·Avoid misleading terms like "quasihuman" or "human-level," which overstate competence and invite overtrust.
• ·Use precise descriptors: e.g., "superhuman at protein folding," "subhuman at common-sense reasoning."
• ·Separate performance from moral status: No current AI qualifies as an agent or conscious entity.

Conclusion: The four-category ladder is obsolete. AI classification requires a multidimensional approach that disentangles task performance, generality, autonomy, and consciousness.