Sticky Information

In the AI Story

The surgeon's knowledge of what is wrong with a particular instrument exemplifies stickiness. She has held the instrument inside a human body a thousand times. She knows, not from reading a specification but from accumulated memory of her hands, that the jaw angle is wrong by three degrees for her specific approach, that the ratchet mechanism creates vibration at a critical moment, that the grip diameter is too large for her sustained hold. The knowledge lives in the intersection of her particular technique, her particular anatomy, and the particular procedures she performs. Transferring it to a manufacturer requires translating embodied expertise into formal specification — a translation that is expensive and lossy.

The stickiness is not a market failure in the traditional sense. It is a feature of how knowledge is distributed in the world. The person closest to the problem knows the most about the problem, and the most important things she knows are things she cannot easily tell anyone else. This is true in surgery, in teaching, in architectural design, in marketing, in every domain where the quality of a solution depends on understanding the specific context in which the solution will be used.

Before the language interface, users with sticky information had two options. They could attempt to transfer the information to a manufacturer through surveys, focus groups, or customer support — accepting the degradation that the transfer process imposed. Or they could act on the information themselves, building a solution that leveraged their direct access to the sticky knowledge, but only if they possessed the technical skills to build. The language interface introduces a third option: the user describes her need in natural language, and a machine translates the description into a working solution without requiring the user to fully codify her knowledge.

The stickiness itself has not been eliminated — the surgeon's tacit knowledge remains tacit, the teacher's embodied understanding remains embodied. What has changed is the cost of acting on that knowledge. The user no longer needs to either codify her knowledge or acquire technical skills. She needs only to describe her need in the language she already uses to think about it. The technical skill previously required to act on sticky information has been absorbed by the machine, rebalancing what matters in innovation toward domain expertise and contextual understanding — toward the sticky information itself.

Origin

Von Hippel introduced the sticky information concept in a 1994 paper in Management Science titled 'Sticky Information and the Locus of Problem Solving: Implications for Innovation.' The paper formalized an intuition that had emerged from two decades of empirical research on why user innovation persists despite manufacturer incentives to eliminate it. The concept gave a name to what the data had already shown: that information costs, not capability or incentive, determined where innovation actually occurred.

The paper's central argument — that problem-solving locates itself where the sticky information is — became one of the most cited contributions in innovation economics. It reframed the question of whether users or manufacturers should innovate as a question about where the relevant information resides and what it costs to move. The answer, across the industries von Hippel studied, was consistently that the information was expensive to move and that user innovation was therefore structurally favored for problems where context mattered.

Key Ideas

Cost of transfer, not concealment. Sticky information is not hidden; it is expensive to extract, codify, and transmit without loss of fidelity.

Contextual and embodied. The stickiness derives from the information's residence in specific contexts — tacit knowledge, workflow adaptations, embodied experience — that resist abstraction.

Structural explanation for user innovation. Users innovate because acting on their sticky information is often cheaper than transferring it to a manufacturer and waiting.

Stickiness persists with AI. The language interface does not eliminate stickiness; it reduces the cost of acting on sticky information without requiring its codification.

Rebalancing of what matters. When technical skill is absorbed by the machine, domain expertise and contextual understanding — the sticky information itself — become the scarce and decisive inputs to innovation.

Debates & Critiques

Whether AI systems can eventually acquire sticky information through extended observation of users — via telemetry, conversation logs, and behavioral data — is an open question. Von Hippel's framework suggests that stickiness is resistant to this kind of extraction because the information is constituted by context and embodiment rather than by patterns of observable behavior. Empirical investigation of how much user-specific context large language models can absorb through sustained interaction will determine whether the sticky information barrier is permanent or merely expensive.