What does this research mean for the field?

Generative AI compresses the creative spectrum, boosting weaker performers while reducing idea diversity among creators. Novelty: ClaimNovelty.SYNTHESIS. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The aim is to understand how AI compresses the creative spectrum and to propose alternative architectures that could enhance creativity.

February 26, 2026Open Access

The B+ Trap How AI Compresses the Creative Spectrum - And What a Rebel Architecture Could Change

Key Points

The aim is to understand how AI compresses the creative spectrum and to propose alternative architectures that could enhance creativity.
Analysis of the B+ Trap concept as it relates to AI creativity performance.
Review of 70+ language models and their homogenization effects on creativity.
Meta-analysis of 28 studies quantifying human-AI collaboration outcomes across 8,214 participants.
Identification of metacognitive engagement as a differentiator among highly creative individuals.
Proposal of a 'Rebel AI' architecture aimed at enhancing creative divergence.
While AI boosts average creative performance, it significantly reduces idea diversity.
Meta-analysis shows a moderate boost (Hedges' g = 0.27) in human-AI collaboration but a large negative effect on diversity (g = -0.86).
Reinforcement Learning from Human Feedback narrows output distributions and reinforces majority preferences.

Abstract

Generative artificial intelligence now surpasses average human performance on standard creativity benchmarks, yet the most creative humans still outperform every AI system tested. This asymmetry defines what we call the B+ Trap: AI compresses the creative spectrum from both directions, pulling weaker performers up while thinning the top of the distribution as fewer creators push beyond AI’s suggestions. A 2025 NeurIPS Best Paper studying 70+ state-of-the-art language models confirmed the effect at ecosystem scale, documenting an “Artificial Hivemind” of extreme inter-model homogenization. A meta-analysis of 28 studies (8,214 participants; Holzner et al. 2025) quantifies the trade-off—human-AI collaboration yields a moderate boost to individual creative performance (Hedges’ g = 0.27) but inflicts a large negative effect on idea diversity (g =−0.86). The mechanism is structural: Reinforcement Learning from Human Feedback (RLHF) provably narrows output distributions and amplifies majority preferences, and alignment training has been shown to systematically reduce language models’ conceptual diversity. Yet some creators consistently escape the trap. The research identifies a key differentiator: metacognitive engagement— the capacity to critically evaluate, actively diverge from, and strategically redirect AI suggestions rather than accept them. This paper argues that what these exceptional humans do cognitively, a different class of AI architecture could do computationally. We present a speculative but empirically grounded “Rebel AI” framework—synthesizing adversarial creativity networks, novelty-maximizing reward functions, curiosity-driven exploration, and open-ended evolutionary search—as alternative training paradigms that optimize for divergence rather than convergence. We conclude with implications for small and medium enterprises, where creative distinctiveness is often the primary competitive advantage. Keywords: artificial intelligence, creativity, RLHF, homogenization, adversarial train- ing, novelty search, metacognition, SME strategy, divergent thinking, generative AI

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper