- Generative AI can amplify existing biases and thinking patterns within organizations
- Human limitations often pose the biggest bottlenecks in innovation processes
- Proper prompting, review design and human oversight are crucial for effective AI integration
Generative AI is actually changing the way we approach innovation, but perhaps not as evenly as many teams might think. A recent working paper, summarized on SSRN, points out that the outcomes from the same AI model can vary depending on the human and organizational hurdles it is used within. The main takeaway is simple, yet important, AI does not just speed up innovation. It can also tend to amplify the preexisting thinking patterns, screening preferences and blind spots that organizations already have.
The authors of the paper, which include researchers from Harvard Business School, Wharton, Northwestern and Columbia, see innovation as a kind of sequence consisting of four steps: idea generation, screening, understanding consumer insights and then learning from the market. And here’s the thing, they argue that many of the trickiest problems in these stages are not really technical at all. It is more about human limitations. People tend to cling onto familiar ideas, reviewers often favor well-polished proposals, consumers have trouble pinpointing unmet needs until they have actually tried a product, and managers tend to interpret feedback selectively. So even with generative AI in play, you might get increased output but still a narrow set of ideas, and worse, it can deepen the biases teams want to escape from.
That broader outlook really helps explain why the paper’s warning is not just a simple “be careful.” In many organizations, the real bottlenecks for innovation are often hiding right there in plain sight. For instance, a team might think they are falling short because they do not have enough ideas, when maybe the actual problem is that the environment rewards only the familiar ones. Or perhaps another team perceives their difficulty as weak consumer research, but really, it is that customers just can’t clearly articulate their unmet needs until they actually interact with the product. Sure, generative AI can lend a hand at each step, but it’s not a fix-all solution, especially not if used carelessly. In fact, if companies deploy it haphazardly, it might even make these underlying problems less obvious, creating the illusion that their innovation pipeline is all smooth sailing when it’s really not.
This tricky tension shows up especially in the early idea-stage, where teams often turn to AI for quick brainstorming. According to the paper, if you just ask the model to “think outside the box,” it tends to drive teams toward the most statistically common answers. That can end up flattening originality, causing different companies to come up with remarkably similar ideas. Other research into cognitive bias bottlenecks points out that things like default heuristics, the order in which ideas are presented, and typical errors in judgment all shape which ideas get surfaced and which get ignored. The practical solution, the authors suggest, is not just to tell people to “think bigger” after the fact. Instead, you should redesign the prompting process and review methods so the AI is nudged toward divergent ideas rather than more of the same.
Once you move to the screening phase, the challenge shifts from creative generation to evaluation. The working paper highlights that AI-generated proposals tend to seem more convincing because they are fluent, thorough and easy to read. But this can cause reviewers to mistake sophistication for quality, which is not always accurate. That becomes even more problematic if organizations train these models on past approval decisions, they might unintentionally automate the biases already baked into those decisions instead of correcting them. A related study on AI integration echoes this concern, framing it more broadly within management issues: organizations constantly wrestle with balancing autonomy and oversight, efficiency and trust, while also trying to minimize bias without unintentionally amplifying it.
The most compelling use case for AI, the paper notes, comes after the launch phase when teams are flooded with reviews, tickets, social media posts and customer returns, stuff that is just too much for humans to handle effectively. AI can help cluster and summarize all that feedback at scale. But even then, it can bury weak signals and sometimes make it easier for teams to stick to decisions they already favored, rather than truly reassessing. That’s why the authors stress the importance of maintaining human judgment at some level. If companies let AI handle too big a chunk of the loop from concept to customer feedback, employees might lose the critical touch with real users that sharpens judgment. And other research on AI bias points out that flawed data and organizational incentives often sneak into these automated systems, reinforcing existing problems instead of solving them.
All in all, the big message is clear: AI is a powerful tool, but not a magic wand. Its effectiveness depends heavily on how organizations integrate and manage it, and on being aware of the biases and blind spots it can both reveal and hide.
Takeaways
- - Generative AI can boost output, but it may also amplify existing organizational bias.
- - The biggest innovation bottlenecks are often human, not technical.
- - Better prompting, review design and human oversight matter as much as the model itself.
- - AI is most useful when it supports, rather than replaces, judgment in sourcing ideas, screening proposals and learning from customers.
Disclaimer: This article may have been created with AI assistance and reviewed by our editorial team. It is provided for general informational purposes only. Readers should verify information independently before relying on this content.
