Your Apprentice Already Asked ChatGPT About That Wire. It Was Wrong 60% of the Time.

Tom Kriger runs research and education for North America's Building Trades Unions, which means he spends a lot of time in rooms full of apprentices asking them how they learn. Sometime last year, the answers started changing. Instead of citing the codebook or their instructor or the journeyman they shadow on Tuesdays, apprentices began mentioning something else. They were asking ChatGPT. Specifically, they were using it to interpret electrical codes and keep current with changing regulations. Kriger told Construction Dive in April 2026 that focus groups confirmed this pattern across multiple training centers.

Nobody taught them to do this, no training center added it to the curriculum, and no contractor approved it. The apprentices just did it, the way twenty-somethings adopt any tool that gives them an answer faster than flipping through a 1,000-page codebook printed in 8-point type. That would be fine if the answers were right.

Dakota Prep, which runs NEC electrical exam preparation courses and has helped thousands of electricians get licensed, tested ChatGPT's newest model on 100 National Electrical Code questions, and the results were unambiguous: more than 60% of the answers were wrong. Not ambiguous, not debatable, not "well, it depends on your jurisdiction," but flatly, provably incorrect according to the published code text. Dakota Prep's assessment was blunt: "Prepping with it is a fast track to failing your exam."

A New York electrical inspector put it differently, writing on the NYEIA website that they had personally reviewed "countless AI responses to code questions, and more often than not, they are incorrect or dangerously incomplete." Real job sites had installations partially or fully redone because contractors relied on AI-sourced code interpretations, and some pushed back during inspections, insisting their work was compliant, only to later admit the interpretation came from a chatbot.

Three Programs Racing Toward the Same Problem

Washington noticed, and so did Redmond, and so did Stanley Black & Decker, though none of them coordinated with each other before launching separate responses.

In April 2026, NABTU and Microsoft announced an expanded partnership to bring free AI literacy courses to skilled trades workers nationwide, building on earlier work that had already trained 1,500 instructors. The new effort launched no-cost courses on LinkedIn Learning for apprentices and journey-level professionals, with an industry-recognized credential on completion, and through TradesFutures, NABTU's affiliated nonprofit, the program extends into apprenticeship readiness networks in 34 states.

Separately, the U.S. Department of Labor launched an initiative to integrate AI skills into Registered Apprenticeships nationwide, with a multi-year contract solicitation to align workforce development strategies with AI competencies across traditional trades.

And DeWalt, after commissioning a survey that found only 9% of construction workers use AI day-to-day while 86% claim they feel prepared to, committed $75,000 to ABC's Trimmer Construction Education Fund and launched a pilot AI training program. Their survey exposed the learning gap plainly: workers are teaching themselves AI through YouTube (40%), Coursera (39%), and video tutorials (42%), with no curriculum, no verification protocol, and no one checking whether what they learned is correct for their jurisdiction.

Shadow AI Leaves No Trail

SMACNA, the Sheet Metal and Air Conditioning Contractors' National Association, coined a term for what's happening: "the Bring Your Own AI dilemma." A foreman uses a free AI app to generate a materials list, unknowingly uploading proprietary assembly details to a public model. An estimator asks an AI chatbot about code requirements and gets guidance that hasn't been verified for local jurisdictions, then builds a bid around it. These are individual actions, well-intentioned, compounding into liability exposure nobody can quantify because shadow AI leaves no audit trail.

Consider the math, rough as it necessarily is. According to BLS data and ABC estimates, roughly 370,000 construction apprentices are active in the U.S. at any given time, and the industry needs 456,000 additional workers by 2027. NABTU's own focus groups confirm informal AI use for code interpretation among this population, though no quantitative adoption rate has been published. If even a fifth of those apprentices use ChatGPT for code lookups once a week, and 60% of NEC answers from general-purpose AI are wrong, that's roughly 44,000 potentially incorrect code interpretations per week entering the construction pipeline with no record of the query, no documentation of which answer was used, and no way to trace a wiring decision back to the prompt that produced it.

Purpose-built alternatives exist and they work measurably better. The International Code Council's AI Navigator, launched in 2023, has answered over 140,000 code questions using models trained on amended and specialized variants of the International Codes, and unlike ChatGPT, it rejects speculation, provides linked sources, and refuses to guess. An ASCE-published study from the Computing in Civil Engineering 2025 conference found that a prototype chatbot built for electrical construction contractors showed improved jurisdiction-specific accuracy over general-purpose tools, confirming that the problem isn't AI itself but unsupervised, unverified AI being used for safety-critical decisions.

The Counterargument Nobody Wants to Hear

Workers have always used unreliable sources. They asked a buddy. They misread the codebook. They worked from an outdated edition of the NEC because the new one costs $175 and their employer didn't buy it. AI isn't introducing a new category of error. It's a new medium for an old problem.

Fair point, and it deserves its full weight. But there's a difference between a peer who says "I think it's a 14-gauge minimum for that run" and a machine that states "NEC Article 210.3 requires a 14 AWG minimum conductor for 15-amp branch circuits" in the same authoritative tone it would use to tell you the capital of France. One carries the implicit uncertainty of a human guess. The other carries the false authority of a machine that sounds exactly like an expert even when it's fabricating article numbers that don't exist in any published edition of the code.

And unlike a dog-eared NEC copy with highlighted annotations that document a decade of an electrician's evolving understanding, ChatGPT leaves no trace. When an inspector flags a violation three months later, there's no way to determine whether the interpretation came from the code, the journeyman, or a prompt typed into a phone at 6:47 AM.

What You Should Do

If you're a general contractor running residential jobs, ask your subs whether their apprentices are using AI for code lookups, because they probably are, and establish a policy before an inspection forces the conversation.

If you're a homeowner with a project underway, you can't ban your electrician's apprentice from using their phone, but you can ask your contractor whether they have an AI use policy and require that code interpretations come from verified sources.

If you're an apprentice, NABTU's new AI literacy courses on LinkedIn Learning are free and carry an industry-recognized credential, and ICC AI Navigator is purpose-built for code questions and doesn't hallucinate article numbers. Both are better than asking a machine that gets the answer wrong six times out of ten.

Limitations of This Analysis

NABTU's focus group data is qualitative, with no published sample size or methodology. Dakota Prep's 60% error rate comes from their own internal testing and has not been independently replicated. Shadow AI usage is by definition untracked, so all prevalence estimates are extrapolations. The DeWalt survey reflects self-reported responses from a population with selection bias toward tech-engaged professionals. The 44,000-per-week figure is illustrative, not measured. The ASCE prototype chatbot study is a proof-of-concept, not a production deployment, and ICC AI Navigator's accuracy has not been independently audited against a standardized NEC question set.