THE RESEARCH

Bootlegging comes from high-tech R&D. The transparency paradox comes from Chinese phone-factory line workers. Algorithmic management and knowledge-hiding come from gig workers. CEO overconfidence comes from M&A in tech and finance. Lumpy investment comes from US census plants. None of them studied the actual case — a manufacturing engineer deciding whether to surface an AI-assisted increment. Stacked, the inferential risk compounds.

VerdictNot fatal, but the load-bearing weakness. Bernstein leads because it is closest — an actual factory. The rest are presented as convergent analogies, and the book's contribution is a hypothesis the pattern warrants, not a proven law. Tyre & von Hippel and the NUMMI study put two more bricks on a real assembly line. Two case studies is still thin.

The evidence on gatekeeping shows a bias against high novelty. Increments are low-novelty by definition. "Gates reject bold ideas" and "gates ignore small improvements" are different mechanisms, and an earlier draft blurred them.

VerdictA real logical flaw, and territoriality fixes it. An increment doesn't need to be novel to be suppressed — it only needs to sit on someone's patch. Radical ideas die from novelty-aversion; increments die from territorial defence plus the scale threshold. Two mechanisms, each independently evidenced.

Lumpy, big-bang investment is often genuinely optimal — economies of scale, indivisibilities, coordination. You cannot half-retool a line, and some increments genuinely cannot be decoupled from the capital decision.

VerdictThis bounds the claim, and the boundary is stated in the book. Boards are right about capital indivisibility. They are wrong to route reversible, two-way-door increments through the same one-way-door gate.

Hidden improvements don't scale, aren't documented or maintained, create single points of failure, duplicate effort, and can breach traceability and certification. On a safety-relevant line, productive deviance is a liability and a recall risk, not a life-hack. Some gates exist precisely because unvetted process changes can kill people.

VerdictThe most serious substantive objection in this list. The book distinguishes domains where deviance is cheap to reverse and low-stakes from regulated ones where the gate is load-bearing and surfacing is mandatory. Romanticising the submarine without that distinction would lose the manufacturing reader, and would deserve to.

The argument bets on the branch where surfacing means surveillance and surveillance means concealment. But AI could instead make increments cheap to document, validate and de-risk — removing the fixed cost that made the gate impassable, so increments clear it easily and never need the submarine at all.

VerdictA genuine fork, and the book names it rather than hiding it. What decides the branch is whether the framing is developmental or evaluative. If build costs stay low while organisations get better at surfacing cheaply, the pattern broke and the book is a period piece.

Bernstein's own prescription is bounded privacy with four specific boundaries — not a surveillance-free culture. Executives have legitimate reasons for visibility: capital allocation, risk, compliance, coordination.

VerdictRhetorically strong, analytically overstated, and softened to Bernstein's actual finding. Not zero observation — observation decoupled from immediate evaluation and control.

Post-it and Gmail are the bootleg projects that won. The graveyard is invisible. And the widely quoted finding that shadow AI beats formal pilots compares self-selected power users on easy tasks against formal programmes attempting hard integration.

VerdictSo the book claims only what survives this: underground activity is where a lot of real value first appears. Not that it beats formal effort on average.

The shadow-user-innovation paper and the ambidexterity frameworks are theory, not tested findings. The revelation-threshold model offers testable propositions and no test results. Sophistication reads as proof if you let it.

VerdictThey are cited as scaffolding that names and structures the phenomenon. The empirical weight stays on Bernstein, Criscuolo, Doms & Dunne, Grijalva, Liu and Allen.

The psychological tax of concealment rests on one survey of 399 employees, published very recently with no replication. The shadow-IT penalty rests on one behavioural experiment. Neither was conducted in manufacturing.

A second study has since landed pointing the same way — Xie and colleagues on generative AI, burnout and bootleg innovation — and it would be easy to bank that as corroboration. It isn't. Two recent self-report surveys of similar design in similar settings are the same instrument pointed twice, not two independent confirmations, and neither of them was run in a factory either.

VerdictPresented as consistent signals, not settled findings. If either fails to replicate, the book's account of what concealment costs the person doing it gets weaker — and nothing else in the argument moves.

Territoriality and voice-extinction are strongly evidenced, but neither was studied on hidden AI-assisted increments in a factory, so they join the stack in attack one. The hollowing-out is weaker still: the literature documents firms losing capability and says nothing about how individual employees feel about it.

VerdictThe mechanisms are sound. The chain linking them is mine. The step from lost organisational capability to individual fear of failing is an inference from experience, and it is doing real work in the Foreword — so it is marked as such there, and in the list above.