Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions

URL SCAN: Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions
FIRST LINE: # Economics > General Economics

The Dissection

This paper is really about bottleneck relocation, not creativity. Generative AI makes idea production cheap; human screening becomes the scarce, failure-prone resource. The experiment shows that identical AI access produces worse outcomes when delivered after the team’s selection process has begun: late-access teams generated 73% more proposals, but added breadth instead of quality. Even language-model judges detected the output increase while missing the quality penalty.

The important finding is mechanical: AI injected at the wrong point does not augment judgment. It overwhelms it.

The Core Fallacy

The paper treats screening as a durable human function that can be repaired through scheduling and workflow design. Under the Discontinuity Thesis, screening is not a permanent moat. It is the next layer to be automated.

The integrated end state is obvious: systems generate, compare, test, rank, prune, and iterate before human attention is required. The paper identifies a temporary integration tax, not a defense of human creative labor. “Expert-rated creativity” in a bounded product-design experiment is also not the same thing as durable economic indispensability.

This supports P1 at the task level and exposes an early form of P2: human coordination becomes the cost center once generation is abundant. It does not, by itself, prove P3. It does show why P3 becomes easier once the screen itself is absorbed into the machine pipeline.

Hidden Assumptions

• Human teams remain the final legitimate arbiters of quality.
• Expert ratings accurately measure economic value.
• AI access is episodic rather than continuous and autonomous.
• The model cannot perform its own search-selection loop.
• More effective screening remains a human skill instead of a software capability.
• The product-design task generalizes to the wider economy.
• AI is used by teams rather than owned and operationally integrated by capital controllers.
• The main problem is timing, not the eventual replacement of the screen.

These assumptions freeze the decisive variable—the automation of judgment—at the exact moment it is becoming cheaper to automate.

Social Function

This is a partial truth with a transition-management function. It correctly exposes a real failure mode: badly integrated AI creates option floods and makes human judgment less effective. But it packages a structural displacement problem as a managerial scheduling problem. That gives institutions a lever to pull while preserving the comforting fiction that the human team remains the center of production.

It also functions as prestige signaling: the organization can claim sophistication by optimizing AI workflows while avoiding the harder question of who will still be economically necessary when the workflow is fully automated.

The Verdict

The paper is useful because it catches the screen collapsing under cheap abundance. Its title is accidentally more prophetic than its conclusions: eventually, pruning will not be “too late” because the machine will prune continuously before the human team sees the options.

This is not evidence that AI cannot replace creative teams. It is evidence that poorly integrated AI temporarily makes those teams look necessary. The paper documents transition friction and institutional incompetence, not a surviving human moat.