Artificial intelligence systems found to excel at imitation, but not innovation
Claims that today’s AI systems excel at imitation but lack true innovation prompt debate over what “creativity” and “understanding” actually mean for machines trained largely on text and images. Commenters contrast probabilistic next-token prediction with human experience, embodiment, will, and feedback loops, while others argue that recombination, statistical pattern-finding, and self-play can already yield novel, useful outputs. Many expect future systems with richer world interaction and explicit incentives for originality to narrow the innovation gap, but disagree on whether fundamental limits or only current training setups are to blame.
Scope of the study and benchmarks
- Several commenters say the paper implicitly equates “AI” with current LLMs and uses a benchmark that leans heavily on reasoning and constrained tasks, where stock models are known to be weak.
- Some point out the raw numbers (near-human innovation rates in parts of the task) make the “AI lacks innovation” headline feel like moving the goalposts.
Imitation, composition, and “real” innovation
- Broad agreement that LLMs excel at imitation and especially at composition: recombining existing ideas (code, themes, styles) into coherent new artifacts.
- View: most human knowledge work is also composition with small tweaks, so LLMs are strong competitors.
- Counterview: AI output often feels superficial and distinguishable from high‑quality human work; it mostly threatens lower-end content jobs.
Innovation as feedback, will, and randomness
- One side: innovation requires will or intrinsic goals; AIs are just powerful goal-seeking tools defined by human designers.
- Opposing view: what really matters is feedback and evaluation, not “will.” Self-play and scientific methods are cited as examples of how feedback drives creativity; future systems that generate their own experience could innovate more.
- Another line of argument: all innovation is some mix of composition and (pseudo-)random variation; critics respond that this ignores the richness of human sensory experience and deliberative abstraction.
Creativity tests, abstraction, and allegory
- Some propose decoding genuinely novel allegories and deep abstract patterns as a meaningful creativity test; they report LLMs doing poorly on such tasks.
- Others counter with anecdotes of LLMs producing nontrivial analogies and bespoke GUI code, arguing skeptics systematically overweight failures.
Hallucinations, reasoning, and planning
- Hallucinations are framed as models “talking ahead of their knowledge” without planning or admitting ignorance.
- Suggested mitigations include reasoning add-ons, tree-of-thought rollouts, and automated self-critique loops.
Consciousness, sentience, and definitions
- One camp sees debates about “understanding,” “sentience,” and “consciousness” as fuzzy-vocabulary disputes that distract from concrete capabilities.
- Another argues definitions still matter for ethics, rights, and law, even if they don’t change near-term technical assessments.
Demand for innovation vs sameness
- Several commenters note that many markets (menus, Hollywood franchises, trendy design) actually reward safe, familiar output more than innovation, making LLM-style imitation commercially attractive.