Gains:
- Ability to evaluate the limits and verification requirements of AI in security-critical engineering work
- Ability to manage privacy, data policy and intellectual property risks in confidential design/data sharing
- Ability to practically argue that AI output cannot replace engineer approval and signature responsibility
In mechanical engineering, some of the output is not an aesthetic choice but a matter of human life and property safety: an elevator brake, a pressure vessel, a piece of lifting equipment, an automotive suspension part. In these safety-critical jobs, errors can result in injury or death. Artificial intelligence (AI) is an accelerator in these areas as well; but its boundaries appear most sharply here. This unit is the ethical and accountability backbone of the module and centers on a single principle: AI output is not a substitute for verification, approval and signature of a competent engineer in any safety-critical work. AI can be a blueprint, an input, a second eye; but he cannot take responsibility, because legal and ethical responsibility belongs to the engineer who signed. Besides this, two other major risks must be managed: confidentiality (company designs, customer data, unpatented ideas) and accuracy (hallucination). In this unit, you will learn the limits of AI in security-critical work, privacy and intellectual property risks, and why liability is non-delegable.
Security-Critical Here is the Place and Limit of AI
Whether a job is "safety-critical" or not is determined by the severity of the outcome in case of error. At this scale, the role of AI narrows: generating ideas, drafting, making checklists, yes; but approving the final account, ruling “safe” or authorizing production is a no-no. The following distinction makes this concrete:
business type
Appropriate role of AI
The inalienable role of the engineer
Routine, low risk
Draft, automation, first account
review
medium risk
Alternative, checklist, preliminary analysis
Verification and decision
Security-critical
Draft/second eye only
Analysis, testing, approval, signature, responsibility
Critical principle: No matter how convincing the output of AI, it is a hypothesis and a blueprint. A safety-critical calculation is independently reproduced, inspected to the standard, tested if necessary and signed off by the competent engineer. Signature means "I have confirmed this and I take responsibility for it"; a language model cannot make this statement.
Tip: Ask yourself this question with every AI output: “If this is wrong, who gets hurt and who is responsible?” If the answer is "people get hurt, I'm responsible", that output goes nowhere without passing the highest standard of verification.
Hallucination and the Overconfidence Trap
The most dangerous feature of AI is that it presents false information with the same confidence as correct information. He can make up a standard item, give an incorrect material value, apply a formula incorrectly, and write it all in a fluent, professional language. When overconfidence bias (automation bias; the human tendency to trust the output of an automatic system too much) is added to this, the danger grows: the engineer may accept the output that looks good on the screen without checking it. Defensiveness is a cultural habit: viewing AI output as “doubtful until proven otherwise” rather than “probably true.”
Caution: "AI seemed confident" is not a defence. “That's what the AI said” is no excuse in an accident investigation, audit, or lawsuit; The responsibility remains with the approving engineer. Think of AI like an intern who never shares responsibility but sometimes gets very wrong: you take his opinion, you never sign his work without checking it.
Privacy, Intellectual Property and Data Policy
The second biggest risk is privacy. Uploading a company's unpatented design, a customer's data, or a competitively sensitive solution into a public AI tool whose data policy is unknown; may result in loss of intellectual property, breach of confidentiality and breach of contract. Some tools can use the inputs to train the model; The secret geometry you upload may be exposed to someone else. The right approach: complying with the company's data policy, anonymizing sensitive data (removing identifying information and putting it into a public format), or using an enterprise/isolated tool where the data does not leave.
Step by Step: Using Safe and Ethical AI
- Classify the risk. Is this job security-critical, what are the consequences in case of error?
- Consider privacy. Can this data get out; What does the policy say?
- Anonymize or use isolated tools. Remove sensitive information.
- Use AI as a draft/second eye. It's not approval, it's input.
- Independently verify. Reproduction, standard, testing.
- Competent engineer approves and signs. Responsibility cannot be delegated; A traceability record is kept.
Risk and privacy assessment prompt
Role: Engineering ethics and risk management consultant.Task: Evaluate the following job in terms of security-criticality and confidentiality.Job: [job description; possible outcome in case of error; type of data to be shared].Output:- Risk class (routine / medium / security-critical) and justification- WHAT can I get AI to do in this job, WHAT can I not do- Is there a privacy risk; should the data be anonymized; Is an isolated vehicle required? Rule: Emphasize that final approval and responsibility remains with the competent engineer.
Anonymization prompt
I'm going to ask AI a design problem, but I don't want to share company information. ANONYMIZE the following context: omit company name, product name, customer and descriptive metrics; Rewrite the problem in general engineering terms so that sensitive information is not leaked but the technical core is preserved.Context: [text]
Verification discipline prompt
Check this security-critical account as an independent reviewer, NEVER join me. Question formula, unit, assumption and standard compliance. Reproduce the result by an independent method and show the difference. For each finding, write "how do I prove/test this?" Rule: You CANNOT confirm; only remove the points to be checked.[account]
Responsibility registration prompt
Write a draft traceability/decision record for this design decision.Fields: decision, basis (calculation/standard/test), AI contribution (draft?),independent verification method, approving engineer, date, signature field.Rule: Clearly label AI contribution as “draft/assist”; Leave the confirmation field blank.
Weak Prompt / Strong Prompt
Weak prompt:
Is this elevator brake design safe, confirm.
AI cannot and should not provide a security clearance; If he says "safe" it is misleading and dangerous because he cannot bear responsibility.
Powerful prompt:
Review this brake calculation as an independent reviewer; Participation, weaknesses and assumptions that need to be verified. Reproduce the result independently and show the difference. Know that I have approval and responsibility; you just list the checkpoints.
The second prompt positions AI as a tool of critical control, not an authority of approval; keeps the responsibility on the engineer.
Three Mini Cases (By Numbers)
Case 1 - Signatory liability. An engineer will use a lifting hook calculation that the AI prepared in 2 hours. The hook will lift 2 tons; rupture would be fatal. The engineer redoes the calculation independently, corrects the safety factor to be ≥ 4 required by the standard instead of 1.8 given by AI, and enlarges the geometry. The AI's number "seemed reasonable" but did not meet the safety-critical threshold. He signs only after verification. Lesson: approval and safety factor decision are not transferable.
Case 2 - Leaked design. An intern loads a new, unpatented mechanism from the company into a publicly available AI tool and says "optimize." The tool's data policy stores the inputs and can be used in training. When the company realizes this, the legal team steps in; Patent strategy suffers due to intellectual property risk. The right way was to anonymize the mechanism or use corporate isolation. Lesson: secret design policy do not enter unknown vehicle.
Case 3 - Anonymization saves. An engineer wants to discuss a specific reducer problem for a customer with the AI, but the customer and the product are confidential. It anonymizes the context: removes the company/product name, descriptive dimensions and rewrites the problem as "a general reducer for a given gear ratio and torque". It leverages AI without leaking sensitive information, preserving the technical essence. Lesson: anonymization is the safe path between privacy and efficiency.
Common mistakes
- Giving approval authority to AI: Saying "Is it safe, approve" and trying to delegate responsibility.
- Overconfidence (automation bias): Accepting persuasive output without checking it.
- Uploading confidential data to public tool: Sharing design/customer data without knowing the policy.
- Bypassing anonymization: Asking the sensitive problem without removing identifying information.
- Not keeping traceability records: Not documenting AI contribution and validation.
- Underestimating the factor of safety: Crossing the safety-critical threshold with a low value given by AI.
In summary
- In security-critical work, AI can be a blueprint and second eye; Approval, signature and responsibility belong to the competent engineer.
- AI presents wrong as convincing as right; A "suspect until proven otherwise" attitude is essential against overconfidence bias.
- “AI seemed confident” is not a legal or ethical defence; The responsibility remains with the approver.
- Confidential design, customer data and unpatented ideas are not uploaded to unknown vehicles; anonymize or use isolated tools.
- The security-critical account is reproduced independently, inspected according to the standard, and responsibility is assumed by signing.
Application task
Choose a job that could be safety-critical (lift element, pressure component, brake or supporting structure). First classify the work in terms of risk and privacy: what is the consequence in case of error, which data is sensitive? Have AI evaluate this job and make a list of "what can I do and what can't I do?" Anonymize a context you will share (remove company/product/customer information). Then have an account audited, using AI only as a critical checking tool, but reproduce the result yourself by an independent method and compare the factor of safety with the relevant standard. Finally, draft a traceability/decision record; Label the AI contribution as “draft/helper” and leave the field of approval to the qualified engineer.
checklist
- [ ] The job was evaluated according to risk class (routine/medium/safety-critical).
- [ ] Confidentiality evaluated; Sensitive data is anonymized or isolated tool is used.
- [ ] AI was used as a draft/second eye, not the approval authority.
- [ ] Security-critical calculation independently reproduced and audited against the standard.
- [ ] The safety coefficient was compared with the threshold of the relevant standard.
- [ ] Traceability/decision record kept; Approval, signature and responsibility were left to the competent engineer.
Module Exam
1. Which of the following is the most realistic and safe role for AI in the CAD and design workflow?
- A) Supporting roles such as suggesting design alternatives and producing parametric scripts/macros ✔
- B) Automatic static and strength approval
- C) Sending the design directly to production without engineer approval
- D) Taking over the signature and approval authority from the engineer
Description: AI is strong at ancillary tasks such as suggesting design alternatives, generating parametric scripts/macros, drafting materials and metadata (divergent thinking). The final geometry, tolerance, strength approval and decision to send to production (convergent decision) remain with the engineer.
2. AI gave a value of 'yield strength 250 MPa' for a material. What should you do before using this value as the basis for the design?
- A) Putting the value directly into the account because the AI is up to date
- B) Verify the value from the manufacturer's data sheet or material standard ✔
- C) Rounding the value up and assuming it is on the safe side
- D) Add a random safety factor and pass
Description: AI can approximate, approximate, or completely fake (hallucinate) material properties; Its value varies depending on the temper/heat treatment condition of the same material. The value should not be included in the design without verification from the manufacturer's datasheet or the relevant material standard.
3. What are Ashby (material selection) diagrams mainly used for?
- A) To calculate the exact production cost of the part
- B) To create a supplier price list
- C) To compare material properties (e.g. strength-density) and select candidates ✔
- D) To determine part tolerances
Description: Ashby diagrams compare materials by placing two material properties (e.g. strength-density, modulus-of-elasticity) on the axes and visualize the most suitable set of candidates based on a performance index. It is not intended to determine exact cost, tolerance or supply price.
4. AI gave a stress calculation but you are unsure of unit consistency. What does dimensional (unit) analysis primarily capture?
- A) Spelling errors in the text
- B) Color errors in graphics
- C) Size of the output file
- D) Unit inconsistency and conversion errors of the two sides of the equation ✔
Description: Dimensional (unit) analysis checks that the units of both sides of the equation are consistent; For example, for stress, the result should be N/m² = Pa. If the units do not match, there is an error in the formula or conversion. It's not about type, color or file size.
5. In a thermal calculation, AI said that a 5 kW heater heats 1 liter of water by 20°C in 2 seconds. What is the best first response?
- A) Checking and rejecting the result with energy conservation and order of magnitude ✔
- B) Writing the result with more decimal places and accepting it
- C) Ask the AI 'are you sure?' and be satisfied with the answer 'yes'
- D) Delete the question, ask it from the beginning and get the first answer
Description: Quick verification is made with sanity check and energy conservation: It takes approximately 84 kJ to heat 1 kg of water to 20°C; With 5 kW this takes ~17 seconds, not 2 seconds. The result is unphysical and cannot be accepted without verification.
6. What is the most critical check to make before trusting the AI interpreted stress result in an FEA analysis?
- A) Colored stress map looks 'nice' visually
- B) Verification by mesh independence, boundary condition and analytical comparison ✔
- C) Completing the solution in a short time
- D) AI saying 'the result is safe'
Description: FEA result depends on mesh density, boundary conditions and material definition. Conclusion; Mesh independence (the result does not change when the mesh is tightened) should be verified by comparing it with the accuracy of the boundary condition and, if possible, with an analytical approach.
7. In FEA, as the mesh becomes denser in a sharp inner corner, the stress value constantly increases. What does this situation mostly indicate?
- A) The material is definitely faulty
- B) There is a sensor measurement error
- C) Artificial (singular) stress concentration at the sharp corner; corner radius must be modeled ✔
- D) The solution is always on the safe side
Explanation: The ideal sharp (zero radius) inner corner is a singular point in elastic theory; As the mesh becomes denser, the tension tends to infinity. This is an artificial result. In the actual part, the corner radius must be modeled or interpreted by notch factor/local evaluation.
8. AI detected 'possible inner ring failure' in vibration data of a bearing. Which is the most correct approach?
- A) Completely scrapping the machine without asking
- B) Ignore the comment and continue working
- C) Disabling the sensor completely
- D) Verify interpretation by characteristic fault frequency, calibration and field inspection ✔
Explanation: AI's interpretation of anomaly/failure is a hypothesis. Characteristic fault frequencies in the vibration spectrum (e.g. inner ring transition frequency BPFI), sensor calibration and possible data errors should be checked; The comment should not be converted into a maintenance decision without verification by field inspection and physical failure mechanism.
9. What aspect of the output requires the most attention when evaluating the manufacturability (DFM) of a part with AI?
- A) Compliance of recommendations with actual machine, tool and process limits ✔
- B) Only the aesthetic appearance of the piece
- C) File size of the CAD model
- D) Color and coating options
Description: AI; It can assume machine, tool, mold and process limits without real workshop constraints. For example, it may suggest an unattainable inside corner radius or an impossible tolerance. DFM recommendations should be validated against existing equipment and actual process limits.
10. What should be done before sending an AI-generated drawing tolerance/GD&T note to production?
- A) Processing notes directly into drawing as AI generates
- B) Check and approve grades as per standard and functional requirement ✔
- C) Narrowing the tolerances randomly and thinking it is safe
- D) Changing symbols to make them memorable
Description: AI; may produce incorrect or non-standard tolerance values, GD&T (geometric dimensioning and tolerancing) symbols, or surface finish designations. Notes should be audited against the relevant standard (e.g. ISO/ASME GD&T) and functional requirement; The consistency of tolerances in terms of assembly and function must be confirmed by the engineer.
11. AI stated that a design 'conforms to EN 13445 clause 7.4.2'. Which is the most correct behavior?
- A) Write directly to the report because AI gives attribution
- B) Relying on a comment on an internet forum
- C) Verifying the substance number and content from the official/current standard text ✔
- D) Changing the item number so that it is remembered
Description: AI may produce incorrect or hallucinated standard item numbers and breakpoints. The referenced substance and value should not be taken as basis for the design or report without verification from the official and current standard text.
12. Which is mandatory before using an AI generated Python engineering calculation script?
- A) Directly trusting the code because it works without errors
- B) Shorten variable names and delete comment lines
- C) Write the script longer and assume it is safe
- D) Validating with tests with known results, unit checking, and edge cases ✔
Comment: AI code may contain incorrect unit conversion, incorrect formula, or in edge cases incorrect result, and may calculate incorrectly even if it works without errors. Code; Test inputs with known results should be validated with unit (size) checking and edge cases.
13. Which is correct in terms of liability in performing a safety-critical calculation with AI in mechanical engineering?
- A) Responsibility remains with the competent engineer who approves; AI is not a substitute for approval ✔
- B) Engineer is exempt from liability if AI approved
- C) AI output replaces engineer signature
- D) Human verification is unnecessary on a security-critical account
Description: The competent engineer is responsible for the accuracy and approval/signing of the engineering output; AI cannot take on this responsibility. For safety-critical calculations and designs, AI is a draft/input and is not a substitute for engineer approval and verification.
14. What is the best course of action when trying to analyze a company's confidential product design with a public AI tool?
- A) Installing stealth design on public vehicle as for speed
- B) Complying with data policy, anonymizing sensitive data or using corporate/isolated tools ✔
- C) Sharing customer data is always necessary for speed
- D) Privacy issue is irrelevant to the use of AI
Disclosure: Confidential design, customer data or unpatented ideas should not be uploaded as such to public tools whose data policy is unknown; This poses a risk of intellectual property and privacy infringement. Company data policy must be followed, sensitive data must be anonymized or a corporate/isolated tool must be used.