AI’s capability overhang leaves most users underutilizing existing models, argues Ethan Mollick

Ethan Mollick, writing on his Substack One Useful Thing, argues that the current capabilities of AI models are barely being tapped by most users, creating a “capability overhang” between what models can do and what people accomplish with them.

He points to recent examples where frontier models like GPT-6 Astra and Fable 5.1 completed multi-week creative and technical projects in hours, such as converting the 1977 text adventure Zork into a 3D action-adventure game and reconstructing Umberto Eco’s personal library in 3D from fragmented visual and textual sources.

Mollick also describes using GPT-6 Astra to produce a book trailer for his upcoming work, Co-Existence, including scripting, 3D animation in Blender, voice generation, music, and editing—tasks that typically require large teams and specialized skills.

These demonstrations relied on his own expertise to guide the AI, spot errors, and iterate, underscoring that current AI systems still depend on human judgment to achieve high-quality outcomes.

Mollick emphasizes that the gap is not about future models but about how people leverage existing ones, urging users to focus on human-specific advantages: deep knowledge, wide knowledge, taste, and agency.

Deep knowledge enables experts to quickly detect AI outputs that deviate from domain norms, while wide knowledge helps users recognize and request specific patterns or techniques (e.g., avoiding design “eyebrows” on webpages).

Taste, he argues, is the human ability to evaluate and refine creative direction, and agency is the capacity to set goals, provide feedback, and integrate AI outputs into broader projects.

He concludes that competing with AI in output generation is a losing strategy; instead, people should combine their unique human strengths with AI to achieve outcomes neither could produce alone.