Using ChatGPT for OSHA and Electrical Calculations? 2026 Test of 450 Safety Scripts

In November 2024, licensed electrical contractors on Mike Holt's Forum opened a thread titled "Using ChatGBT as a tool" to compare notes on apprentices and estimators bringing chatbots onto commercial jobs (Mike Holt's Forum, Nov 2024). One veteran member tested routine overcurrent protection and reported back in eleven words: "I just checked to see a fuse coordination, it was wrong, again." Another practitioner checked National Electrical Code lookups and found "none of the citations were real." In April 2026, construction safety researcher S. M. Jamil Uddin at Colorado State University measured that exact failure mode across 450 AI-generated Python safety calculators for OSHA 29 CFR Part 1926 construction rules (Uddin, arXiv:2604.12311, Apr 2026). While 85.3% of the 450 scripts ran without crashing, more than half got the OSHA safety logic wrong, printing clean numbers that violated federal standards.

Most "AI safety at work" articles focus on leaked memos, which is why our safe use of AI tools at work hub separates office data hygiene from physical environmental, health, and safety (EHS) hazards. When an AI assistant invents a sentence in a slide deck, someone loses ten minutes of editing. When a chatbot misstates an excavation slope or drops a terminal temperature limit on a feeder lug, physics collects the bill immediately. One cubic yard of trench soil weighs 2,700 to 3,000 pounds, roughly the weight of a compact car (OSHA Trenching). Federal OSHA's 2026 penalty schedule sets maximum civil fines at $16,550 per serious violation and $165,514 per willful or repeated violation (OSHA Penalties), before stop-work orders or workers' compensation spikes.

Generic trade blogs warn crews "use AI for paperwork, never code" in two paragraphs without explaining why clean AI math fails or how local code editions trap contractors (Trade-schools.net, Apr 14, 2026; FindSkill). Below is the 450-script dataset, the token-prediction mechanism behind edition blurring, Massachusetts's April 24, 2026 switch to the 2026 NEC under 527 CMR 12.00, two worked trench and fall-clearance calculations, and a gangbox-ready policy card.

What happened when researchers tested 450 AI-generated OSHA safety scripts?

In a controlled benchmark of "vibe-coded" construction safety calculators, Colorado State University's S. M. Jamil Uddin found that 85.3% out of n = 450 zero-shot LLM Python scripts executed without syntax errors, yet only 46.7% of the 450 scripts applied valid OSHA 29 CFR Part 1926 regulatory logic (Uddin, arXiv:2604.12311). Among the 384 scripts that ran cleanly, 45.3% (174 out of 384 runnable scripts) failed silently, printing formatted pass/fail safety numbers built on wrong thresholds.

State the methodological boundary up front. Uddin's April 2026 paper is a single, non-peer-reviewed arXiv preprint (n = 450 generated Python scripts) testing zero-shot prompts across three compact commercial models (GPT-4o-mini, Claude 3.5 Haiku, and Gemini 2.5 Flash), not frontier models wired to verified engineering solvers. Even so, those compact models are the exact fast tiers powering free phone apps, browser sidebars, and low-cost contractor "AI copilot" wrappers used in pickup trucks every morning.

Uddin tested 150 construction safety scenarios across three job-site prompt styles: a Safety Manager using formal 29 CFR 1926 vocabulary, a Foreman using trade shorthand, and a Field Worker asking plain-English questions:

That bottom row exposes an inverse competence trap. In software work, a syntax error crashes with a red warning. On a job site, a silent logic bug prints "Minimum Clearance: 13.5 ft - PASS" in calm text, and plain-English worker prompts trigger invented safety variables in 87.3% of trials. Compare how error rates shift across tasks in our guide on how often AI is wrong by task and denominator.

Why do language models fail deterministic OSHA and NEC code math?

Large language models fail OSHA and electrical code calculations because next-token predictors blend statistical phrasing across web pages rather than solving hard step-function lookups or physical constraint equations. Safety rules are discontinuous: one degree of heat, one foot of trench depth, or one change in soil moisture flips an entire legal and physical regime.

Picture asking a talkative retired counter clerk at an electrical supply house to recite NEC Table 310.16 ampacities and OSHA 1926 scaffold tie-ins from memory in a noisy parking lot without opening a codebook. Every trade term sounds familiar, and the answer arrives in four seconds, yet the clerk's memory unintentionally mixes scraps of the 2014, 2017, 2020, and 2023 codebooks with forum rules of thumb. Where that analogy breaks is self-awareness: a human clerk feels memory blur on a five-factor derating stack and opens the printed table, whereas a token predictor formats a guessed average with calm four-decimal certainty.

Two independent benchmarks show why field prompts break LLM math:

• Distractor sensitivity (GSM-Symbolic): Apple researchers Iman Mirzadeh and colleagues tested models across 50 symbolic math templates (Mirzadeh et al., arXiv:2410.05229). Adding a single realistic, non-operative detail (GSM-NoOp, like mentioning yesterday's pipe delivery count alongside today's load) dropped accuracy by up to 65% across tested model cohorts. Real field texts are packed with non-operative weather, crew, and delivery notes.
• Conditional table stacks (SpreadsheetBench): Across n = 912 real spreadsheet workflows (SpreadsheetBench, arXiv:2406.14991), single-turn models failed on 75% to 85% of benchmark tasks requiring conditional cross-table lookups, the exact structure of OSHA and NEC tables.

How does NEC edition blurring create hidden electrical code hazards in Massachusetts?

Using a chatbot for National Electrical Code (NFPA 70) lookups causes edition blurring: models mix rules from the 2017, 2020, 2023, and 2026 NEC cycles without knowing which edition or state amendment your local Authority Having Jurisdiction (AHJ) enforces (NFPA 70).

Effective April 24, 2026, Massachusetts adopted the 2026 Edition of NFPA 70 (NEC) with Massachusetts-specific amendments codified in 527 CMR 12.00 (Mass.gov Electrical Code). Drive forty-five minutes into a neighboring state and the inspector may enforce the 2023 or 2020 NEC with different state amendments. Because most public web text quotes older 2017, 2020, or unamended 2023 articles, chatbots routinely hand Massachusetts electricians outdated rules that fail inspection under 527 CMR 12.00 or create fire and shock hazards.

Licensed electricians repeatedly flag four chatbot code traps (Mike Holt's Forum):

1. Terminal temperature caps (NEC 110.14(C) vs Table 310.16): Asked to size THHN/THWN-2 copper for a 100-amp feeder, chatbots read straight from the 90 deg C column (#3 AWG at 115A) while forgetting 110.14(C), which caps equipment lugs rated 100A or less (or unmarked) at 60 deg C (#3 AWG = 85A) unless terminals are listed and marked for 75 deg C (100A). The 90 deg C value is a starting point for derating, never an automatic pass at the lug.
2. Dropped continuous-load and bundling multipliers: Continuous loads (3+ hours, such as EV chargers or commercial lighting) require sizing at 125% of the continuous load rule (NEC 210.19 / 215.2) plus raceway bundling (310.15(C)(1)) and rooftop ambient corrections. Chatbots often apply one multiplier and silently drop the second.
3. Fabricated fuse coordination and arc-flash values: Selective coordination (NEC 700.32) and incident-energy math (NFPA 70E) require OEM time-current curves, transformer impedance, and bolted fault current. Chatbots guess ratios and invent fake subsections.
4. GFCI and 527 CMR 12.00 blind spots: 250V receptacle, outdoor HVAC, kitchen, and energy-storage GFCI rules shifted across the 2020, 2023, and 2026 cycles, and 527 CMR 12.00 amends specific articles in Massachusetts.

Two worked job-site calculations where a plausible AI answer turns lethal

Step through two OSHA 29 CFR Part 1926 calculations with every assumption stated so your safety officer can verify each number against federal text.

Worked Example 1: 8-foot utility trench beside active road traffic (29 CFR 1926.652)

• Assumptions: A utility crew digs an 8-foot-deep, 3-foot-wide trench for a leaking water lateral beside an active two-lane road. The soil is previously disturbed sandy-clay backfill with visible water seepage and continuous heavy-truck vibration 12 feet away.
• Deterministic OSHA Subpart P rule: Cave-in protection is mandatory at 5 feet depth (OSHA 1926.652). Under Subpart P Appendix A, previously disturbed, vibrating, and seeping backfill cannot be Type A and defaults to Type C, capping the maximum allowable slope at 1.5H : 1V (34 degrees). At 8 ft depth, each wall cuts back 8 * 1.5 = 12.0 ft, requiring a 27.0-foot top width (12 + 3 + 12 = 27 ft) or a rated trench box / shield.
• How a casual prompt fails: Texting "8 ft deep trench, 3 ft bottom, firm sandy clay, what top width for OSHA slope?" omits the disturbed-fill and seepage triggers a Competent Person checks on site. Pulling Type A (0.75H : 1V -> 15.0 ft top width) or Type B (1H : 1V -> 19.0 ft top width) leaves 8 to 12 feet of saturated soil wedge unshored above a worker when soil weighs 2,700 to 3,000 lb/cu yd.

Worked Example 2: Foot-level tie-off fall clearance (29 CFR 1926.502(d))

• Assumptions: A 6-foot-tall carpenter wearing a full-body fall-arrest rig (dorsal D-ring 5.0 ft above boots) works on an unprotected 15.0-foot mezzanine deck and considers tying a 6.0-foot heavy-duty 12-ft-free-fall shock lanyard (4.0-foot deceleration pack) to a deck-level foot anchor (0.0 ft).
• Deterministic clearance math: Anchored at foot level, the worker falls 5.0 ft (D-ring to deck) plus 6.0 ft lanyard length (11.0 ft free fall, requiring a 12-ft-FF rated lanyard under 1926.502(d)(16)(iii)). Required clearance below the 15.0 ft deck equals 6.0 ft lanyard + 4.0 ft deceleration + 5.0 ft worker height below D-ring + 1.5 ft webbing stretch/shift + 2.0 ft safety margin = 18.5 feet. At 15.0 ft, the worker strikes concrete 3.5 feet before arrest completes.
• How a chatbot fails: Prompted with "6 ft shock lanyard clearance on 15 ft deck," chatbots output the textbook overhead-anchor sum (14.0 ft < 15.0 ft -> SAFE), missing foot-level geometry.

What can tradespeople and EHS teams safely use AI for? Algorithmic hygiene on site

CDC NIOSH's January 18, 2026 bulletin "Practical Strategies to Manage AI Hazards in the Workplace" (CDC NIOSH) and the National Safety Council's Work to Zero research (NSC PDF) define algorithmic hygiene: use language models only where errors are linguistic and visible to a human reader, and ban them wherever errors hide inside physics, chemistry, or code tables.

Before adopting any green-zone admin AI tool, test it on 50 past job files with our 50-example spreadsheet method to test an AI tool, add this Red/Green split to your one-page AI use policy for a 10-to-50 person business, and keep our first-hour AI incident response plan in the job trailer so bad submittals get pulled before work is energized.

Copy and post: the ed.cafe Job-Site AI Red/Green Card & 3-Step Code Verification Protocol

Tape the ed.cafe Job-Site AI Red/Green Card & 3-Step Code Verification Protocol inside your gangbox lid, job trailer, and apprentice packet. The ed.cafe 3-Step Code Verification Protocol gives foremen a sixty-second field check.

================================================================================
THE ED.CAFE JOB-SITE AI RED/GREEN CARD & 3-STEP CODE VERIFICATION PROTOCOL
Post in: Job Trailers, Gangboxes, Dispatch Desks & Apprentice Packets
Publisher: ed.cafe by No Worker Left Behind (Boston 501(c)(3), EIN 88-3268312)
Core Rule: AI may draft words. AI NEVER calculates loads, clearances, or code.
================================================================================

1. RED ZONE - ZERO-EXCEPTION BAN (NO CHATGPT, COPILOT, CLAUDE, OR AI OVERVIEWS)
   Why: In a 2026 CSU test of 450 AI-generated OSHA Part 1926 scripts, 85.3% ran
   without crashing, yet 53.3% got the OSHA safety logic wrong.
   [X] Trench soil typing, sloping, benching, or trench-box sizing (1926.652)
   [X] Fall-arrest / SRL total fall clearance, anchor loads, or swing-fall math
   [X] NEC / MA 527 CMR 12.00 wire sizing, lug temp limits, derating, or GFCI
   [X] Fuse/breaker coordination, fault current, or NFPA 70E arc-flash boundaries
   [X] Crane load charts, sling-angle tension math, rigging, or scaffold loads
   [X] Lockout/Tagout (LOTO) isolation steps, confined-space air, or SDS chemicals
   [X] "Vibe-coded" custom Python/Excel macros for any life-safety calculation

2. GREEN ZONE - ALLOWED ADMIN USES (WITH MANDATORY HUMAN READ-THROUGH)
   [OK] Turning rough voice notes into clean daily logs, RFIs, or change orders
   [OK] Translating ALREADY-VERIFIED safety handouts (bilingual lead checks text)
   [OK] Finding which page number in an uploaded OEM PDF covers a part or torque
   [OK] Drafting customer scheduling emails (never paste worker SSNs or injuries)

3. THE ED.CAFE 3-STEP DETERMINISTIC CODE VERIFICATION PROTOCOL
   STEP 1 - PIN THE AHJ & EDITION: Record the town AHJ, code year, and state
            amendments (e.g., MA 2026 NEC + 527 CMR 12.00, eff. Apr 24, 2026).
   STEP 2 - OPEN THE PRIMARY TABLE BY EYE: Read the printed codebook, licensed
            NFPA / eCFR page, OSHA Subpart appendix, or OEM load chart directly.
   STEP 3 - COMPUTE DETERMINISTICALLY & SIGN: Run math only on a calibrated trade
            calculator, OEM tool, or PE-locked sheet. Sign the JHA/permit log.

4. 90-SECOND MONDAY TOOLBOX TALK SCRIPT (READ ALOUD TO CREW)
   "Quick rule on phone AI apps like ChatGPT or AI search summaries on this site.
   Use them to clean up an RFI email if you want, never to check OSHA rules,
   trench widths, lanyard clearance, rigging, or NEC wire sizes. Researchers
   tested 450 AI safety calculators this year: 85% ran without an error message,
   and nearly half of those working scripts spat out wrong OSHA math. Plus AI
   doesn't know local amendments like Massachusetts 527 CMR 12.00. If a number
   protects a body or goes in a panel, open the real codebook, OEM chart, or ask
   the Competent Person / Master on site."
================================================================================

Crews adopt this boundary fastest after seeing a live AI error in a low-stakes room: run our 75-minute first hallucination lab for adults or pull facilitator guides in the Staff Room. Through the Massachusetts AI Hub, the Google AI Professional Certificate is free for Massachusetts residents (Mass AI Hub), and No Worker Left Behind (Boston 501(c)(3), EIN 88-3268312) is forming Circle #1 at ed.cafe/start/, a free 30-chair cohort closing December 1, 2026 (pre-launch as of October 2026; zero cohorts completed yet).

Where this stops being true

• Preprint scope and model tiers: Uddin (arXiv:2604.12311, n = 450 scripts) is a single non-peer-reviewed preprint testing zero-shot Python scripts on compact models (GPT-4o-mini, Claude 3.5 Haiku, Gemini 2.5 Flash). Solver-backed engineering software differs sharply, though no unvalidated probabilistic chat interface is acceptable for life-safety design.
• Calibrated computer vision vs chatbots: Purpose-built proximity sensors, forklift cameras, or ergonomic monitors audited under CDC NIOSH and NSC algorithmic-hygiene rules are distinct from asking a text chatbot to do code math.
• State-Plan OSHA and local AHJ rules: Twenty-two states and Puerto Rico run private-sector OSHA State Plans (such as Cal/OSHA, MIOSHA, and WA DOSH) with stricter triggers than federal 29 CFR 1926, and local electrical inspectors enforce municipal/state amendments (527 CMR 12.00 in Massachusetts).
• Non-delegable licensed sign-off: Nothing here replaces an OSHA Competent Person, a Registered Professional Engineer (1926.652 trenches over 20 ft), or a licensed Master / Journeyman Electrician.

Questions people ask

Key terms

• Algorithmic hygiene: The CDC NIOSH workplace safety practice (January 2026) of auditing, bounding, and verifying AI tools with the same hazard-control discipline applied to physical or chemical hazards.
• Authority Having Jurisdiction (AHJ): The state board, municipal building department, or local wiring inspector legally empowered to enforce a specific code edition and local amendments on your site.
• Competent Person (29 CFR 1926.32(f)): An on-site individual capable of identifying existing and predictable physical hazards and authorized by the employer to take prompt corrective measures, including stopping work.
• Edition blurring: A language-model failure mode where rules from multiple NFPA 70 (NEC) or OSHA code years are blended into a single plausible-sounding but non-compliant answer.
• Inverse competence trap: The empirical pattern in arXiv:2604.12311 where plain-English prompts from less code-trained workers triggered higher rates of AI-invented safety variables (87.3% of worker prompts) than formal manager prompts (64.7%).
• Silent safety failure: A script or calculation that runs without syntax errors (85.3% execution rate) while outputting unsafe or code-violating specifications (45.3% of runnable scripts).

Sources

1. S. M. Jamil Uddin (Colorado State University), "Is Vibe Coding the Future? An Empirical Assessment of LLM Generated Codes for Construction Safety" (Apr 2026, n = 450 scripts; single non-peer-reviewed preprint): arXiv:2604.12311 PDF
2. Mike Holt's Forum Practitioners, "Using ChatGBT as a tool" (Nov 2024): Mike Holt's Forum
3. Commonwealth of Massachusetts, "Massachusetts Electrical Code (527 CMR 12.00 - 2026 NEC effective Apr 24, 2026)": Mass.gov and NFPA 70
4. Iman Mirzadeh et al. (Apple), "GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in LLMs" (Oct 2024): arXiv:2410.05229
5. Zeyao Ma et al., "SpreadsheetBench: Real World Spreadsheet Manipulation" (Jun 2024, n = 912): arXiv:2406.14991
6. CDC NIOSH, "Practical Strategies to Manage AI Hazards in the Workplace" (Jan 18, 2026): CDC NIOSH
7. National Safety Council Work to Zero, "Predictive Analytics and Machine Learning for EHS": NSC PDF
8. US Department of Labor / OSHA, "Trenching Safety, 29 CFR 1926.652, and 2026 Penalties": OSHA Trenching, 29 CFR 1926.652, and OSHA Penalties
9. Trade-schools.net & FindSkill, "AI for Tradespeople & Electricians" (2026): Trade-schools.net and FindSkill
10. Massachusetts AI Hub, "Free Google AI Certificates" (2026): Mass AI Hub