Stickiness Without Resistance: Knowledge Transfer Failure in Human-AI Collaboration Without Human Friction

Fuente: Zenodo
Saved in:
Bibliographic Details
Main Author: Hardwick, Spencer
Format: Recurso digital
Published: Zenodo 2026
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866901499431354368
author Hardwick, Spencer
author_facet Hardwick, Spencer
contents <p>Spencer Hardwick, Founder @ Calx Labs</p> <p>spencer@calx.sh<br>@spenceships on Twitter</p> <p>Keywords: knowledge transfer, human-AI collaboration, behavioral learning, correction engineering, organizational learning, agent memory</p> <div> <div>Szulanski (1996) demonstrated that best practices resist transfer within organizations even when the source is willing, the recipient is motivated, and the knowledge is documented. For three decades, the explanation has centered on human friction: ego, politics, low absorptive capacity, relational distance between people. We present evidence from two field studies spanning 43 days, 8 AI agents, 151 documented corrections, and two production systems that knowledge transfer stickiness persists in human-AI collaboration where every human friction barrier is absent. AI agents have no ego, no motivation to hoard knowledge, and no political incentive to resist adoption. The stickiness persists anyway.</div> <br> <div>In Study 1, 237 correction-derived rules transferred from one agent to another produced 44 new corrections in the receiving agent, with 13 falling in categories the transferred rules explicitly addressed. In Study 2, building the correction-tracking system itself, a correction logged for one agent on March 22 was independently reacquired by a different agent through three recurrences on April 1, despite the written rule being available in shared memory. We identify two types of behavioral learning with predictably different transfer properties: architectural corrections, which modify system structure, achieve zero recurrence across both studies; process corrections, which add text-based rules, show recurring failure chains of 3 to 8 entries in which each correction explicitly references its predecessors and the error still recurs. Four domains with authored rules but no active correction history show zero behavioral change, a suggestive pattern consistent with situated learning theory. A preliminary compilation pattern emerges at approximately 3 to 4 recurrences, the point at which process corrections are identified as needing architectural intervention. Meta Superintelligence Labs' HyperAgents framework independently converged on a structurally similar boundary between prompt-level and mechanism-level learning. The correction-tracking tool was itself a site of the phenomenon during construction: 41% of corrections generated while building it were about the correction system itself, the strongest ecological validity test available when the researcher is maximally aware of the effect under study. All artifacts including correction logs, rule files, and commit history are publicly auditable.</div> </div>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19382717
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Stickiness Without Resistance: Knowledge Transfer Failure in Human-AI Collaboration Without Human Friction
Hardwick, Spencer
<p>Spencer Hardwick, Founder @ Calx Labs</p> <p>spencer@calx.sh<br>@spenceships on Twitter</p> <p>Keywords: knowledge transfer, human-AI collaboration, behavioral learning, correction engineering, organizational learning, agent memory</p> <div> <div>Szulanski (1996) demonstrated that best practices resist transfer within organizations even when the source is willing, the recipient is motivated, and the knowledge is documented. For three decades, the explanation has centered on human friction: ego, politics, low absorptive capacity, relational distance between people. We present evidence from two field studies spanning 43 days, 8 AI agents, 151 documented corrections, and two production systems that knowledge transfer stickiness persists in human-AI collaboration where every human friction barrier is absent. AI agents have no ego, no motivation to hoard knowledge, and no political incentive to resist adoption. The stickiness persists anyway.</div> <br> <div>In Study 1, 237 correction-derived rules transferred from one agent to another produced 44 new corrections in the receiving agent, with 13 falling in categories the transferred rules explicitly addressed. In Study 2, building the correction-tracking system itself, a correction logged for one agent on March 22 was independently reacquired by a different agent through three recurrences on April 1, despite the written rule being available in shared memory. We identify two types of behavioral learning with predictably different transfer properties: architectural corrections, which modify system structure, achieve zero recurrence across both studies; process corrections, which add text-based rules, show recurring failure chains of 3 to 8 entries in which each correction explicitly references its predecessors and the error still recurs. Four domains with authored rules but no active correction history show zero behavioral change, a suggestive pattern consistent with situated learning theory. A preliminary compilation pattern emerges at approximately 3 to 4 recurrences, the point at which process corrections are identified as needing architectural intervention. Meta Superintelligence Labs' HyperAgents framework independently converged on a structurally similar boundary between prompt-level and mechanism-level learning. The correction-tracking tool was itself a site of the phenomenon during construction: 41% of corrections generated while building it were about the correction system itself, the strongest ecological validity test available when the researcher is maximally aware of the effect under study. All artifacts including correction logs, rule files, and commit history are publicly auditable.</div> </div>
title Stickiness Without Resistance: Knowledge Transfer Failure in Human-AI Collaboration Without Human Friction
url https://doi.org/10.5281/zenodo.19382717