Gains:
- Ability to distinguish on a risk spectrum where AI is a safe accelerator and where it is a safety risk in construction workflows
- Ability to explain why the approval of a competent engineer is legally and professionally mandatory in calculations and decisions affecting building safety.
- Ability to position AI output as a blueprint input to be subjected to three-layer control, rather than as a replacement for engineering judgment
Civil engineering is a profession that is responsible for keeping the structures in which people live, work and pass safely standing. Miscalculating the moment of a beam, the bearing capacity of a column, or the overturning safety of a retaining wall; It seems like a small number error on paper, but on the field it can lead to loss of life, collapse and destruction. When artificial intelligence (AI in short: software that learns patterns from data and produces text, code or calculations) enters this high-risk area, it should be seen not as a "magic solution" but as a powerful but must-control assistant. In this unit, we will discuss where AI is accelerating and where it is dangerous in construction works; We will see why the responsibility always remains with the engineer and how to maintain ethical-confidentiality boundaries.
A large language model (LLM: a type of AI trained on very large chunks of text that generates responses by predicting the next word) does not "understand" a reinforced concrete code; It completes the pattern he sees in similar texts. So it can produce a very fluent, very convincing but inaccurate account. The engineer's job is to test the output against independent evidence without being fooled by this fluency.
Where Does AI Accelerate, Where Is It Risky?
It is useful to think of construction jobs on a “risk spectrum.” At one end are low-risk, retrievable tasks: drafting a technical paper, summarizing text, writing an Excel formula, editing a meeting memo. At the other extreme are high-risk, irreversible tasks: determining the cross-section of a load-bearing member, confirming a foundation depth, signing off a scaffolding calculation.
AI is a real accelerator at the low-risk end of the spectrum: it reduces what used to take minutes to seconds, and even if there are errors, they are easily fixed. At the high-risk end, AI is just a draft generator; Its output is of no value without verification by independent calculation and signature of a competent engineer.
business type
Risk level
The role of AI
Mandatory check
Draft report/specification
low
spelling accelerator
Content and citation control
Preliminary calculation of quantity
medium
draft generator
Unit and rank control
Standard/item finding
medium
research map
Confirmation from official text
Static/reinforced concrete calculation
high
Draft entered
Hand account + software + signature
OHS risk assessment
high
first draft
Expert/OHS approval
Attention: A sentence "AI calculated this way" does not take away any legal or professional responsibility. Responsibility remains with the engineer who signed. Think of AI like an intern ramping up work: you get the draft, but you oversee and sign every line.
Three-Layered Control Mindset
Never accept a security-critical AI output in one go. Establish a three-layer control habit:
- Order and unit check: Is the result physically plausible? Do the units hold? If the capacity is 12,000 kN for a 30x30 cm column, there is an error somewhere.
- Independent cross-validation: Reproduce the same result with a closed formula, hand calculation, or different software (e.g. static program, Python). If the two paths overlap, trust increases.
- Standard and judgment confirmation: Have the coefficients, formulas and limit values used been verified verbatim from the relevant standard (TS 500, Eurocode, TBDY)? The final decision lies with the competent engineer.
These three layers are the common backbone of all subsequent units.
Weak Prompt / Strong Prompt
How you ask the AI to do the same task determines the auditability of the output.
WEAK: "Is this beam safe?" (Span, load, section, material, support are uncertain; AI produces an estimate, it cannot be controlled.) STRONG: "For a simply supported reinforced concrete beam with a span of 6 m: - Uniformly distributed load g+q = 25 kN/m (design load) - Section 30x60 cm, C25/30 concrete, B500C reinforcement Mid-span "Calculate the design moment with a closed formula, show the formula and units SEPARATELY. Write the intermediate steps so that I can verify the result by hand calculation. If you have chosen the regulation coefficient, specify which item it is based on (I will confirm)."
Powerful prompt; It requires input, constraints, and verifiable intermediate steps. So you can test the AI's output line by line.
Professional Ethics and Confidentiality
The use of AI is not just a technical issue, it is an ethical issue. Three principles are critical:
Privacy. Project plans, cost files, contracts and personal data may contain trade secrets and personal data (within the scope of KVKK). Uploading a client's project to a public AI service without corporate approval and a proper data handling policy is a serious violation. Anonymize data when necessary (project name, location, remove names), share only the numerical data required for the account to function.
Transparency. Don't hide that you're leveraging AI; But don't put the responsibility on AI. It's okay for a report to start with an AI outline, it's okay to present unaudited AI output as your own engineering judgment.
Competence limit. AI can give you a sense of confidence in an area in which you are not an expert. Producing a geotechnical calculation in a fluent text does not make you competent in that field. For work requiring expertise, consult the relevant expert.
Tip: Enterprise-specific, data-free AI solutions (either hosted in-house or contractually guaranteed data confidentiality) are much more secure on confidential projects than public services. Learn your company's data policy before using it.
Three Mini Cases
Case 1 – Accelerating use. Project engineer Elif drafts a 40-page technical specification with AI in 2 hours; Normally it took 2 days. Then, it checks each item according to the standard and finishes it in 6 hours. Total yield: approximately 1.5 business days. The risk is low because the output is fully controlled.
Case 2 – Caught error. Site manager Murat requests a rollover safety calculation from AI for a retaining wall. AI gives the safety number 2.4. Murat does the calculation: AI got the ground thrust coefficient wrong, the real safety number is 1.1. So the wall is at the border. Three-layer control prevents the risk of collapse.
Case 3 – Breach of confidentiality. A technical office employee uploads the cost file of the project they will bid on to a public AI. There are unit prices in the file that should not be shared with the rival company. This is against both the contract and KVKK. The solution was to share only the required numeric fields anonymously or use an in-house tool.
Copiable prompt templates
You can use the three templates below by adapting them to your own business.
RISK CLASSIFICATION PROMPT:"Position the following task on a risk spectrum for construction work: [write the task]. - Say low / medium / high risk, with justification - Evaluate if it is reversible, does it involve safety of life / property - Suggest the appropriate role of AI in this task (accelerator / blueprint / idea only) - List which independent verifications are essential in this task"
PRIVACY / ANONYMIZATION CHECK PROMPT: "Tick if there is any confidential/personal data (project name, location, name, unit price, TR/contact) in the following text that I need to remove before uploading it to the AI and suggest an anonymous version: [paste text]Leave only the numerical data required for the account to work."
THREE-LAYER VALIDATION REQUEST PROMPT: "You will produce a calculation result. In order for me to check this in three layers: 1) Write the formula you use and each intermediate step with units. 2) State whether the result is reasonable in terms of order / size. 3) Tell me from which standard / source I should confirm. Do not present the result as an absolute truth; give it as a draft to be checked."
Common mistakes
- Mistaking AI's fluent and confident language for accuracy.
- Security-critical account acceptance in one step, without verification.
- Trying to put the blame on AI by saying "AI calculated it".
- Uploading confidential project data to an external service without approval or anonymization.
- Substituting AI for engineering judgment in a field in which you are not an expert.
- Setting up the prompt indefinitely and getting uncontrollable output without any intermediate steps.
In summary
- AI is a powerful accelerator in low-risk businesses; In high-risk, safety-critical work, it is merely a blueprint to be inspected.
- The responsibility always lies with the competent engineer who signs; AI cannot take on this.
- Pass each security-critical output through three layers: rank/unit, independent cross-validation, standard, and jurisdictional validation.
- Set up the prompt with input, constraint, and verifiable intermediate steps.
- Confidentiality, transparency and competence boundary are the three cornerstones of ethical use.
- Anonymize confidential project data or use secure in-house tools.
Application task
Choose a real task from your own business (a report section, a bill of quantities, a standards item search). First position this task on the risk spectrum (low/medium/high). Then ask the AI in a “strong prompt” format and subject the output to three layers of checking: (1) rank/unit, (2) hand calculation or independent source, (3) standard confirmation. Write down each error and correction you find in one sentence. Finally, consider whether there is a privacy risk in the data you use.
checklist
- [ ] I positioned the task correctly on the risk spectrum.
- [ ] I set up the prompt with input, constraint and verifiable intermediate steps.
- [ ] I checked the rank and units of the result.
- [ ] I cross-verified the result in an independent way (hand calculation/software/source).
- [ ] I confirmed the used coefficients and values from the relevant standard.
- [ ] I subjected the final decision to the approval of the competent engineer.
- [ ] I evaluated the privacy/KVKK risk in the data I used and took precautions.