AI Deepens Innovation Bottlenecks It's Supposed to Solve

AI tools promise to accelerate innovation, but they often make the process worse

Every innovation team now uses similar foundation models and prompt libraries, yet results vary wildly. Some report creative breakthroughs while others drown in homogenized ideas that sound machine-generated. The difference isn't which model teams choose—it's how AI interacts with the human bottlenecks already embedded in innovation processes.

That's the central finding of new research from Harvard, Wharton, Northwestern, and Columbia, published as a working paper invited by the International Journal of Research in Marketing. Rather than cataloging what generative AI can do, the researchers examined how it affects the cognitive, social, and organizational constraints that have always slowed innovation.

Their conclusion challenges the prevailing optimism: when used without deliberate redesign, AI often deepens the very bottlenecks it appears to solve.

Why it matters

Most organizations treat AI adoption as uniformly beneficial—more ideas, faster testing, cheaper research. This research reveals that assumption is dangerous. AI trained on existing human output tends to reproduce the patterns that created innovation failures in the first place, just faster and at scale. Understanding which bottlenecks AI genuinely dissolves versus which it amplifies will determine whether companies achieve breakthrough innovation or just accelerate mediocrity.

Four stages where AI backfires

Ideation becomes fixation. When teams brainstorm with large language models, two problems compound: the model gravitates toward statistically typical responses, then humans become mentally anchored to those safe ideas. The result is a two-step narrowing that reduces both individual creativity and market-wide diversity. Chain-of-thought prompting can push models toward bolder territory, but telling fixated humans to "think more broadly" typically fails.

Screening rewards polish over substance. Evaluation panels have always been biased against novel ideas. AI supercharges this by generating fluent, well-structured pitches that evaluators mistake for quality. Field experiments show that when screeners receive AI recommendations with written rationales, they rubber-stamp rejections more often without improving accuracy. The fix: strip submissions to common formats and remove AI explanations that substitute for human judgment.

Consumer simulations miss irrational behavior. AI-generated customer personas produce systematically wrong predictions because they behave too rationally. Real customers cling to sunk costs and overweight switching friction—the exact behaviors that determine whether innovations succeed. Digital twins underestimate these adoption barriers, making them unreliable for genuinely new products.

Post-launch feedback enables confirmation bias. AI excels at synthesizing massive volumes of customer feedback, genuinely solving an information bottleneck. But the same fluent summaries become tools for selective citation, allowing teams to find support for decisions they've already made. Whether AI synthesis produces honest reporting or just more convincing rationalization remains an open question.

The automation trap

The research identifies a compounding risk: as AI absorbs ideation, screening, and testing work, junior employees lose opportunities to build judgment through direct experience. Organizations develop sophisticated users of AI output who never learned to recognize when that output is wrong.

Over time, the entire innovation feedback loop drifts from real customers as each stage becomes mediated by models trained on each other's outputs rather than market reality. Everything accelerates while losing ground truth.

The researchers don't advocate slowing AI adoption. Instead, they argue for preserving unmediated contact with reality through lead-user immersion, ethnographic fieldwork, and human accountability for consequential decisions. These aren't nostalgic holdovers but load-bearing structures that prevent automated systems from drifting.

The research was first reported by Harvard Business Review, authored by Julian De Freitas, Ayelet Israeli, Gideon Nave, Artem Timoshenko, and Olivier Toubia.