Your quality system has the same blind spot. A CAPA closes on time with clean documentation and never touches root cause. A parameter widens because the data said it was safe, and the field failures arrive three quarters later. A supplier gets qualified off an AI-generated summary nobody went behind.<\/p>\n
Every one of those decisions is now faster and cheaper to make. None of them is measured on the timeline where the cost actually lands. That gap is where deskilling does its quiet work.<\/p>\n The mechanism is boring, which is why people ignore it. Any model that learns from outcomes needs a reward signal it can measure now. Nedelcu puts it bluntly: there is no fitness function you can define for maintainable code, “otherwise it would’ve been baked into our linters.” Same in quality. There is no measurable signal that tells you a root cause analysis was actually correct, or that widening a tolerance raised long-term variation, because those answers arrive quarters later, tangled up with a dozen other changes.<\/p>\n So the model optimises for what it can be scored on: completeness, plausibility, format. A five-whys chain that reads well. A deviation write-up that hits every field. None of that is correctness. It is the appearance of correctness, which is the only thing available to train on.<\/p>\n Before you approve any AI use case, ask one question: how fast and how cheaply can someone verify the output was right? If the answer is minutes, and a competent person can check it without special effort, AI will do well there. Document classification, first-pass complaint triage, pulling specification values out of PDFs, drafting a work instruction from an existing procedure. Wrong answers surface immediately and cost almost nothing.<\/p>\n If verification takes months, or requires waiting for field data, treat AI output as a starting draft and nothing more. Root cause determination, supplier risk judgment, process capability changes, deciding which deviations are actually trivial. These are the calls your senior people are paid for, and they are paid for them precisely because the feedback loop is too slow for rules.<\/p>\n Obviously wrong output is harmless. Someone catches it, laughs, moves on. The dangerous output is the one that passes review because it matches the shape of good work.<\/p>\n That shape is exactly what the model learned. Reviewers trained on format rather than reasoning will approve it. Then the expertise to tell the difference quietly erodes.<\/p>\n Nedelcu reaches for the Dreyfus model to explain what AI actually is: a fast, tireless advanced beginner<\/em>. It learns from rulebooks written for novices and from patterns in the wild, and most work in the wild is mediocre. Feed a model a million procedures of average quality and you get average judgment delivered at speed.<\/p>\n His example is AI “simplifying” code by splitting a function into smaller ones that aren’t reusable, where you still have to read the child to understand the parent. Motion that looks like improvement and isn’t. As he puts it:<\/p>\n Experts don’t follow the rules, they make the rules.<\/p><\/blockquote>\n You have seen the manufacturing version. An AI-generated FMEA that is procedurally immaculate, every severity and occurrence score defensible, every column filled. And it misses the one failure mode your 22-year process engineer would name in thirty seconds because he watched it destroy a quarter’s output in 2014.<\/p>\n Ask that engineer how he knew. He will say the sound was wrong, or the fixture felt loose on second shift, or that supplier always slips after a management change. None of it survives being written into a work instruction, because it is context-dependent in ways rules can’t capture. Rules make beginners productive. They cap experts.<\/p>\n That intuition was built expensively, through production failures the person swore never to repeat. It is the most valuable asset in your plant and the only one that never appears on a balance sheet. It is also the thing that quietly disappears when your engineers stop doing the work that builds it and start reviewing machine output instead.<\/p>\n The loss of tacit knowledge is not an HR problem to solve at retirement. It starts the day a junior engineer accepts an AI-drafted risk assessment without arguing with it. He got the document. He did not get the scar.<\/p>\n The split is not about difficulty. It is about how fast you find out you were wrong. AI is excellent at high-volume, verifiable, reversible work: drafting work instructions, pulling values out of scanned inspection records, flagging anomalies in SPC streams, translating between an ERP and a LIMS, triaging inbound complaints into categories a human then confirms. You see the error in minutes and undo it for free.<\/p>\n Then there is the other list. Root cause determination. Process design trade-offs. Supplier risk judgment. Validation strategy. Slow feedback, high consequence, and, critically, these are the tasks that build expertise in the first place. Hand them over and you buy throughput today against a capability bill that lands on someone else’s balance sheet.<\/p>\n Ask how long until a bad output becomes visible. If the answer is measured in weeks or quarters, no model and no reviewer is getting a useful correction signal. Ask what a wrong answer costs and whether you can reverse it. A mislabelled complaint is a five-minute fix. A waived deviation is not.<\/p>\n Third question, the one nobody asks: who learns from doing this manually? If the task is how your next senior quality engineer develops a nose for trouble, automating it is a training decision, not a productivity decision. Treat it that way.<\/p>\n Rubber-stamping is worse than no review, because it manufactures the paperwork of oversight without the substance. Require the reviewer to state, in one line, what they would have concluded independently before they read the AI output. If they cannot, they are not reviewing, they are approving.<\/p>\n Rotate that review across juniors, not just your two most experienced people. Nedelcu’s warning is that people stop being in control and stop learning from mistakes. The bench that catches AI’s errors only stays stocked if you keep putting people in front of decisions with real consequences.<\/p>\n![\"Software](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/09\/ai-deskilling-ai-has-no-wisdo-inline-1.jpg\")
<\/figure>\nWhy AI Can’t Learn What Takes Years to Show Up<\/h2>\n
Fast feedback vs. slow feedback: a sorting test for any AI use case<\/h3>\n
Why ‘it looks correct’ is the failure mode, not ‘it looks wrong’<\/h3>\n
Experts Don’t Follow the Rules, They Make the Rules<\/h2>\n
The tacit knowledge your best people can’t write down<\/h3>\n
![\"Dreyfus](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/09\/ai-deskilling-ai-has-no-wisdo-inline-2.jpg\")
<\/figure>\nWhere AI Genuinely Wins, and Where It Quietly Eats Your Capability<\/h2>\n
A three-question filter before you automate a judgment task<\/h3>\n
The review discipline that keeps humans in the loop without killing the speed gain<\/h3>\n