Research: Using AI Can Stifle Innovation. But It Doesn’t Have To.

Leaders love AI because it makes knowledge easy to reuse: instant drafts, instant code, instant analysis. But the research shows that making knowledge easy to reuse has a hidden cost. We study how organizations learn, and our recent Management Science research uses a formal analytical model—a mathematical framework that analyzes how individuals in an organization decide whether to explore on their own or reuse what others have found—to examine what happens when “good-enough” answers become essentially free. The model predicted that as reuse rose, independent exploration fell, and teams started to converge on the same few approaches. Put simply, while productivity went up, but innovation quietly flattened. We see a similar dynamic mirrored in empirical research, such as the field experiment documented in a 2015 Research Policy paper by Kevin Boudreau and Karim Lakhani. They found that when computational biologists were given frictionless access to each other’s intermediate solutions, effort shifted toward refining existing approaches at the expense of independent experimentation. The search space narrowed, consistent with the productivity trap our model predicts. We have also experienced the dynamic in our own daily work. To stay ahead of the AI curve, our institutions set up weekly meetings to share new tools and use cases. Attendance was high. But over time the group generated fewer genuinely new discoveries. The pattern was hard to miss: fewer people were contributing to the discussions, and sessions increasingly focused on explaining what the latest AI tools are about, because others lacked the background to understand novel applications without that primer. A small number of people were doing the exploring; the rest were waiting to benefit from it. The system increased sharing, but it weakened the incentive to do the costly work that produces new insight in the first place. It seems clear, then, that generative AI changes the economics of learning. If a plausible strategy memo appears in minutes, fewer people will spend days doing the harder work that produces original understanding, such as talking to customers, triangulating data, and pressure-testing assumptions. When “good enough” is free, fewer people pay the cost to find what’s better. The result is a firm full of people who can retrieve outputs but struggle to judge, adapt, or falsify them. Avoiding the Productivity Trap The solution is a fundamental shift in how we value effort. In our model, adding a small friction to knowledge sharing at the individual level raised the total knowledge at the collective level because it built what scholars call absorptive capacity: the ability to evaluate, adapt, and improve ideas rather than just copy them. When people must invest some effort to use what others found, they do more independent checking and tinkering. That investment both helps them learn from others and produces new knowledge for others to learn from. The result is higher long-run knowledge in the system and higher average long-run performance. That finding prompted us to pilot a “friction protocol” in our MBA and PhD seminars. We set a simple rule: to participate in the sharing session, you had to show evidence of an independent attempt first. We welcomed success stories, but we especially prized the “failed prompts”: the dead ends and hallucinations. Anyone can copy a success from the internet; documenting a failure requires genuine exploration. While introducing the strategic friction appeared to slow the group down, it actually made everyone better off. The most visible shift: more participants arrived with partial solutions to the same problem, which the group could then combine into fuller approaches. By forcing participants to wrestle with the tools first, we built their absorptive capacity. This friction ensured that when a breakthrough was shared, they were cognitively prepared to understand why it worked and how to adapt it. We didn’t just share prompts; we expanded the group’s total “knowledge commons.” Growing Absorptive Capacity Our study is a formal model, not a field experiment: it demonstrates that when learning from others requires prior understanding, long-run knowledge production rises—meaning more distinct approaches survive, fewer people fully abdicate exploration, and the organization is less likely to get stuck on a few “good-enough” solutions. The mechanism it identifies—absorptive capacity—is grounded in decades of research that existing knowledge provides the conceptual frameworks that help individuals encode and apply new information. Children build language by scaffolding new words onto what they already know; professionals build expertise the same way. In practice, absorptive capacity shows up as the difference between retrieving an answer and being able to do something with it. A manager with absorptive capacity doesn’t just accept an AI-generated competitor analysis: she spots what’s missing, tests its assumptions against her own market observations, and adapts its recommendations to her context. A manager without it forwards the output unchanged. When we piloted the friction protocol in our seminars, we tracked two things before and after. First, the diversity of problems and applications surfaced in sharing sessions—a proxy for independent exploration. Second, the depth of adaptation in what was shared: were participants reporting a prompt they had copied, or a workflow they had modified for a specific purpose? After requiring evidence of independent attempts, the shift was remarkable. Before the friction protocol, students surfaced few original problems and almost no partial solutions. After introducing the rule, each student contributed on average two distinct problems and roughly one partial solution per problem. Because everyone had attempted the work independently, sessions became noticeably more interactive, and the contributions showed genuine adaptation: modified workflows rather than copied prompts. Participants could build on each other’s partial solutions quickly, often combining them into full approaches during the discussion. This pattern is exactly what our model predicts: friction doesn’t reduce learning; it raises the quantity and often the quality of what is shared. For managers looking to track absorptive capacity in their own teams, a handful of observable indicators can serve as practical proxies. Are employees increasingly catching errors, omissions, or wrong assumptions in AI outputs—and can they explain why? Are final deliverables diverging meaningfully from the raw AI output, incorporating new facts, constraints, or counterexamples? Are people contributing original observations—customer quotes, numbers from internal systems, competitor moves—rather than generic prompts? And can they apply what they learned in one context to a new problem without starting from scratch? When these behaviors increase, absorptive capacity is growing. Unfortunately, corporate AI is often designed to remove the need for understanding. Leaders optimize for adoption and convenience, unintentionally eroding the organization’s absorptive capacity. To protect long-run innovation incentives, therefore, leaders should reconsider optimizing solely for ease of use and start designing ‘strategic friction’ back into their systems. Introducing Strategic Friction The key isn’t to make work painful. The goal isn’t to create bureaucratic hurdles, but to ensure that reuse depends on capability. It’s about maintaining enough effort in the loop so people keep learning. We demonstrate in our study that there’s an optimal level of absorptive capacity to inject into organizations. Too little friction undermines exploration. Too much slows everything down. And not every task warrants friction: for routine work with low judgment stakes, such as formatting a slide deck, summarizing meeting notes, speed is the point. Strategic friction should be reserved for the tasks where the organization needs people to think and learn, not just produce. In the corporate context, this means moving from an oracle model—where the user asks and receives—to a reciprocal model, where the user must contribute understanding to withdraw wisdom. There is a spectrum of ways to do this, ranging from simple policy changes to more ambitious system-level design. You can: Set a rule requiring documented independent efforts. For example, before using AI to draft a market analysis, an analyst must submit their own rough assessment—even if incomplete—showing what they already know and where they are stuck. It is the easiest friction to implement as it requires no system changes. We took this approach in our academic seminars: no sharing session without evidence of a genuine attempt. Design the AI to ask for the user’s input before generating a full output. For example, if an analyst asks for a competitor analysis, the AI first prompts for three specific observations about the competitor’s recent pricing moves. The AI then incorporates this fresh intelligence into its output. This forces the analyst to investigate before they retrieve: building absorptive capacity while also capturing front-line data that prevents the AI from recycling stale information. The user, however, can provide input through prompts, so the friction remains relatively minor. Build gated interfaces that unlock assistance only after the user deposits their own context. For example, before an analyst can ask the internal AI to generate a risk assessment, the interface would require them to upload a baseline hypothesis and tag three key variables or constraints. The “generate” button stays locked until the human context is deposited. The AI then builds on the analyst’s input rather than replacing it—ensuring that every output is a collaboration, not a delegation. This is the most demanding friction and requires custom tooling of the AI. The policy-level approach is something any team leader can start on Monday. The interface-level approach requires modest design effort but pays a double dividend: it builds the user’s absorptive capacity and captures fresh intelligence from the front lines, preventing the AI model from collapsing into a closed loop of its own outputs. The system-level approach is a longer-term investment, but it is the direction firms should be heading: designing AI tools that make people smarter, not just faster. How to Spot the Builders The defense against stagnation can begin before the employee even logs in. Many firms still hire for “the right answer.” In an AI-saturated environment, that’s less differentiating. What matters is “the right process.” In interviews and performance reviews, therefore, observe how candidates and employees use AI tools. Do they accept outputs as finished work, or do they interrogate them, test assumptions, and add original observations? Recent research makes these patterns visible. A field experiment, led by Hila Lifshitz-Assaf and colleagues at Harvard, MIT, and Warwick, analyzed nearly 5,000 human-AI interactions by 244 BCG consultants and explored three distinct collaboration styles, differentiated by two questions: who selects what needs to be done, and who decides how to do it. “Centaurs” retained human control over both dimensions, using AI only for targeted assistance. “Cyborgs” let the human drive the “what” but gave AI significant control over the “how,” engaging in continuous, critical dialogue. “Self-automators” ceded control of both to AI and experienced what the researchers call “no-skilling,” developing neither domain expertise nor AI fluency. Centaurs and cyborgs are the builders in our framework; self-automators are the free-riders our model warns about. These findings suggest two observable behaviors that distinguish builders from free-riders: Selective delegation. Builders know the frontier of what AI does well and what it doesn’t. They use AI for tasks inside that frontier—drafting, brainstorming, pattern-matching—and do the judgment-heavy work themselves. In prompt logs, their queries are targeted and specific, not broad delegations. Interrogation depth. When builders do use AI, they challenge it. They ask the AI to explain its reasoning, point out inconsistencies, and feed it their own data. They also use AI in the opposite direction, providing their own analysis and asking AI to stress-test assumptions. Free-riders accept and forward. The direction of information flow tells the story: builders push context into the AI; free-riders pull finished outputs out of it. The goal isn’t surveillance; it’s coaching. Just as a code review helps developers improve, an AI-use review helps employees develop better absorptive capacity and collaboration habits with the tool. And it allows organizations to invest in the people who are building the knowledge commons, not just consuming it. . . . AI can make organizations faster immediately. The leadership challenge is preventing that speed from eroding absorptive capacity. Calibrated strategic friction—designed into tools, workflows, and hiring—helps ensure the organization doesn’t just produce answers but keeps producing insight.

Research: Using AI Can Stifle Innovation. But It Doesn’t Have To.

Summary.

Partner Center