Gains:
- Ability to recognize that AI can produce fictitious coefficients, rule clauses and technical values and apply verification in the context of life/property safety
- Ability to manage cyber security, data privacy and regulatory limits in ship and OT (operational technology) systems
- Understanding that the ultimate responsibility for class approval, navigational safety and ethics lies with the engineer and the captain.
Throughout this module, we used artificial intelligence (AI) as an accelerator at every stage of ship and marine engineering: from form design to structural analysis, predictive maintenance to voyage optimization, emissions compliance to shipyard production. In this final unit we step back and look at the whole picture: what are the key limits of AI in maritime, what risks (hallucination, cybersecurity, data privacy, bias) are on constant alert and who bears responsibility? Because the prerequisite for using a vehicle safely is to know where it can go wrong.
The module's abiding thesis applies here and governs all others: in safety-critical maritime work, AI output is no substitute for approval by the competent engineer, master or class surveyor. AI suggests, drafts, calculates and reminds; But the owner of the decision that determines the safety of life, property and the environment is the person who signs that decision and is legally responsible.
Concepts: Hallucination: AI fabricating a value, rule, or resource that does not exist. OT security: Cyber protection of operational technology (machine and cruise control systems). Data poisoning: Intentional corruption of training/input data to mislead a model. Model bias: The systematic and unfair bias of the model that comes from the data on which it is trained. Chain of responsibility: The legal and technical ownership of a decision is clearly evident.
Permanent Limits of AI
AI has four permanent frontiers in shipping, and these will not disappear even as technology advances. First, hallucination: the model fills the gap with “most likely” statements when it does not have definitive data; This turns out to be a made-up class clause, an impossible stability value, or a non-existent coefficient. Second, out-of-deployment behavior: the model becomes unreliable outside the situations for which it was trained and tested (unusual ship type, unusual malfunction, unseen traffic). Third, lack of understanding of physics: AI processes word patterns, does not "understand" energy conservation or Newton's laws; It can produce output that seems reasonable but is physically impossible. Fourth, timeliness: legislation, price and technology change; The model's information is frozen at a date.
These limits explain why the verification discipline we learned throughout the module is indispensable: order-of-magnitude control, unit consistency, independent reproduction, and source verification. Especially when having the AI do calculations (e.g. a Python/Excel script generated for a ballast/stability moment or a power calculation), even if the code appears fluid and working, its logic, inputs and edge cases are not translated into a safety decision without independent testing. Just because an account code "works" doesn't mean it's "correct."
Caution: An AI-generated accounting code may work without errors, but may contain the wrong formula, the wrong unit, or an overlooked edge case. Before accepting the code, test it with known input (an example you know the answer to firsthand), test for edge cases, and verify the result in an independent way. In safety-critical calculations such as stability, an approved loading computer and engineer confirmation are essential.
Cyber Security and Data Privacy
The modern ship is a floating computer network: navigation systems, machinery control, cargo management and satellite connectivity are interconnected. This means cyber attack surface. AI systems add two new risks to this picture. First, OT (operational technology) safety: an AI component tied to navigation and machine control poses a direct safety risk if attacked; so critical OT systems are isolated, AI output should never trigger a control action on its own, human approval and safe default behavior must be maintained. Second, data privacy: hull form, performance curves, shipowner-yard contracts and operational telemetry are trade secrets and intellectual property; Entering this data into unapproved external AI services is an irreversible loss.
Additionally, manipulation risks such as data poisoning (misleading the model with malicious input) and spoofing of data such as AIS require not to blindly trust AI's inputs. Rule: critical data is anonymized or not shared at all, only approved and confidentiality-assured tools are used, critical inputs are cross-verified for authenticity.
Risk
How does it occur?
precaution
hallucination
Made-up rule/value/source
Confirmation from the source, rank control
OT cyber attack
Infiltration of control system
Isolation, human approval, secure default
data leak
Entering confidential data into external tool
Anonymization, approved tool
Data poisoning/spoofing
Intentionally corrupted input
Input validation, cross-checking
Model bias
Unbalanced training data
Reviewing the result with human judgment
Ethics and Chain of Responsibility
The legal and technical ownership of a decision in maritime should be clear: the captain is responsible for navigational safety, the chief engineer is responsible for the machinery, the engineer in charge is responsible for design and calculation, and the class surveyor is responsible for rule compliance. AI should not introduce a “liability gap” into this chain. “AI suggested this” is not an excuse; The human who uses the AI and accepts its output is responsible. Therefore, the use of AI must be transparent (it must be recorded which decision it supports, what verification it has undergone) and the final decision must be up to a signed human.
The ethical dimension also includes crew safety and the environment: if a fuel-saving proposal puts the crew at risk, if an automation solution narrows the safety margin, if short-term gain masks long-term risk, engineering ethics requires saying "no". AI is a powerful tool for productivity; But efficiency can never come before life, the environment and honest engineering judgment.
Mini Cases
Case 1 — “AI said so” excuse. In a report, a limit value is used as given by the AI, without verification, and the value turns out to be incorrect. The responsible person tries to defend himself by saying "AI suggested this." Neither class nor liability law recognizes this: the engineer who accepts the output is responsible. Lesson: AI is not an excuse; The responsibility lies with the person using it.
Case 2 — Leakage of confidential data. A team pastes a shipowner's actual performance and fuel data into a publicly available AI tool and requests analysis. This data is now out of control; trade secret and breach of contract arise. There would be no risk if working with anonymized representative values or an approved tool. Lesson: real/confidential data is never entered into an unapproved external tool.
Case 3 — Working but wrong code. AI generates a Python script for a moment of stability; the code runs without errors and returns a reasonable number. When the engineer tests it with a known loading case (the answer to which is known by hand), he finds that the code takes an arm's length term with the wrong sign. The script was "working" but the result was wrong. Lesson: the operation of the account code is not a guarantee of correctness; Testing with known input and independent verification is essential.
Copiable Prompt Templates
Template 1 — Limit and risk self-assessment:
Role: You are a reviewer auditing the AI output I use.Context: [task/output].Task:1) Where is the risk of hallucination high in this output?2) Is there anything questionable in terms of physical plausibility?3) What values that require timeliness (legislation/price) should be verified?4) Chances of this output being an out-of-distribution (unseen) case?List each questionable point and how to verify it.
Template 2 — Account code verification schedule:
Come up with a verification plan before using the following AI-generated accounting code (Python/Excel):[paste code/logic]1) Suggest a test input for which I know the answer offhand.2) List the edge cases that need to be tested (zero, negative, extreme).3) How do I check unit consistency?4) By what independent means do I confirm the result?Constraint: Approved tool/engineer verification is required on the safety-critical account.
Template 3 — Data privacy filter:
Privacy check the following context I would give to an AI tool:[text to share]1) Is there trade secret / confidential / contract data in it?2) What values should be anonymized or extracted?3) Does this task require an approved internal tool?Constraint: Actual ship/performance/contract data does not go into the external tool.
Template 4 — Accountability and transparency note:
Draft a transparency memo to document the following AI-powered decision:[decision].Content: (1) at what stage was the AI used, (2) what verifications did it go through, (3) which human approved it, (4) what sources were verified.Constraint: Make it clear that the final responsibility lies with the human signing.
Weak prompt / Strong prompt
Weak prompt:
I will put this AI analysis in the report, is it ok?
Powerful prompt:
Role: You are the independent reviewer who audits the AI output.Context: [analysis/output and where to use].Task:1) Flag risk of hallucination, physical plausibility, and timeliness.2) Is there data that should not be shared for confidentiality?3) What steps would a human need to verify this output?Constraint: You don't make the "appropriate/inappropriate" decision; State the verification steps and that the responsibility lies with the signing engineer.
Weak prompt asks for confirmation; Strong prompt exposes risks, privacy and human verification chain.
Common mistakes
- It means "AI said so." The responsibility lies with the person who accepts the output; AI is no excuse.
- Entering confidential data into external tool. Trade secret and breach of contract; Once leaked data cannot be retrieved.
- Thinking that the working code is correct. A script that works without errors may contain the wrong formula/unit; Testing and independent confirmation are essential.
- Leaving the OT system to AI. The control action is never triggered by AI alone; Human approval and secure default are maintained.
- Putting efficiency before safety. Life, environment and honest engineering judgment are not sacrificed for any savings.
In summary
The enduring limits (hallucination, out-of-distribution behavior, lack of grasp of physics, out-of-dateness) and risks (cyber/OT security, data privacy, poisoning, bias) of AI in maritime illustrate why the verification discipline learned throughout the module is indispensable. In safety-critical decisions, AI output is not a substitute for competent human approval; The chain of responsibility is clear and “AI said so” is not an excuse. AI is a powerful accelerator; But life, environment and honest engineering judgment always come first.
Application task
Select an AI output (an account, a report draft, or a code) that you produced in this module. Flag the risks of hallucination, physical plausibility, and actuality with the “limit and risk self-assessment” template. If it is an account code, test it with a known input using the "account code verification plan" template and verify independently. Apply a “data privacy filter” and filter out data that should not be shared. Finally, write a transparency memo documenting the decision and making it clear who has ultimate responsibility.
checklist
- [ ] I checked the hallucination, physics, and actuality risks of the AI output.
- [ ] I have tested and independently verified the account code with known input.
- [ ] I did not enter confidential/contracted data into the external tool; I anonymized it.
- [ ] I maintained human approval and secure default in OT/control systems.
- [ ] I transparently documented the use of AI and clarified who was responsible.
- [ ] I have never put efficiency before life, environment and engineering ethics.
Module Exam
1. When working with an AI tool as a marine engineer, the ultimate responsibility for which of the following should always remain with the human (engineer/classman)?
- A) Approval and signing of a structural element's compliance with the class rule ✔
- B) Producing the first draft of a resistance report
- C) Editing the axis labels of a chart
- D) Preparing a glossary of terms for a technical report
Description: AI; It can speed up tasks such as resistance precalculation, report drafting, and data summary. However, approving the compliance of a structural element with the class rule and reflecting this as a signature on the technical file is a decision that creates life and property safety; cannot be transferred to AI without independent verification and authorized approval.
2. You had the AI calculate the resistance of a boat and it reduced the result by 20% with an 'empirical correction factor' that was not in the model experiment. What is the first correct action to take?
- A) Accepting the result as it is because the AI seems confident
- B) Question the source of the coefficient, reject the unfounded correction and recalculate it with the recognized method ✔
- C) Increasing the coefficient even more and staying on the safe side
- D) Switching to propeller selection without changing the result
Explanation: Language models can convincingly fabricate unfounded coefficients and numbers (hallucination). In a calculation that directly affects propulsion power and fuel, such as resistance, it is necessary to question the source of each coefficient used (method, model test, standard) and reject the unsupported coefficient.
3. Instead of using the amount of cargo carried as the denominator in calculating a ship's CII (Carbon Intensity Indicator), AI used only distance. What is the main problem with this output?
- A) CII is calculated only for tankers, this is unnecessary on board
- B) CII result always comes as a percentage, decimal is wrong
- C) The CII denominator should not only be distance, but the carrying capacity (deadweight) multiplied by distance ✔
- D) Speed data is mandatory for CII calculation, it is missing
Description: CII measures carbon emissions per job carried and is based on the offered payload multiplied by distance (e.g. deadweight x distance). Leaving out the term load/capacity completely miscalculates the indicator. The first and mandatory step is for each term in productivity indicators to fully comply with the IMO definition.
4. What is the most valuable output of an AI-powered predictive maintenance system in the engine room?
- A) Summarize the average cost of past maintenance in one line
- B) Transforming all sensor data into a more colorful graph
- C) Color and list vibration/temperature trends
- D) Pre-marking the time window when a bearing or turbocharger is more likely to fail ✔
Description: The purpose of predictive maintenance is to detect the tendency for equipment to deteriorate before it fails and plan maintenance in a timely manner. Thus, the risk of unexpected malfunctions and stoppages is reduced. It is valuable not only to summarize the past but also to produce early warnings about the future.
5. AI compared the stress result in an FEA (finite element analysis) model to the allowable stress in the class code and declared 'appropriate'. Which is the most correct approach?
- A) Verifying the model setup (mesh, boundary condition, load) and confirming the allowable stress from the class rule text ✔
- B) Directly confirming the result because the AI said 'fit'
- C) Randomly lowering the allowable stress to stay on the safe side
- D) Changing only the color scale without tightening the network
Description: FEA result; It is completely dependent on mesh quality, boundary conditions, load definition and material properties. Additionally, the allowable stress value used by the AI may be incorrect. Before accepting the result, it is necessary to verify the model setup and confirm the class rule value from the text.
6. You created a route that minimizes fuel by performing weather routing with AI. Which consideration is most critical to the usability of the result?
- A) The route must be safe in terms of navigational safety (wave, weather, ship restrictions) and must be approved by the captain ✔
- B) The shortest appearance of the route on the map
- C) The route includes the most port calls
- D) The route minimizes ETA regardless of the contract
Explanation: No matter how fuel-optimal the route proposal is, navigational safety and the authority of the captain take priority. If the storm, wave height and ship restrictions exceed the safety limit, fuel gain is meaningless. The final route decision must be filtered through safety constraints and approved by the captain.
7. You asked the AI for a SOLAS requirement and it gave an exact value with reference to a specific 'Chapter/Rule number'. What is the correct engineering behavior?
- A) Applying the value directly because the AI is up to date
- B) Raising the value randomly to stay on the safe side
- C) Verify item number and value from official, up-to-date SOLAS/class source ✔
- D) Adding to the file without checking the item number
Description: Language models may misremember or make up contract/rule clause numbers and values. In binding fields such as SOLAS, MARPOL and class rules, each item and value must be verified verbatim from the official, current source (IMO text, class rule, management circular); The AI output only shows where to look.
8. In an autonomous or decision-assisted navigation system, the AI suggested a maneuver based on radar and AIS data. Which is the correct assessment in terms of COLREG (conflict avoidance)?
- A) Automatically execute the maneuver as AI accesses current data
- B) Transmit the suggestion to the other ship via VHF and wait for its approval
- C) Turning off the radar and relying only on AIS
- D) Consider the suggestion as a decision support and leave the COLREG and final maneuver decision to the officer of the watch/captain.
Explanation: In accordance with COLREG rules and good maritime practice, the responsibility for avoiding collision and the final maneuver decision rests with the authority on the bridge (captain/officer of the watch). AI recommendation is a decision support; blind spots, sensor error, and rule interpretation require human supervision.
9. You have generated a fuel economy recommendation from the main engine SFOC (specific fuel consumption) curve with AI. What is the most appropriate method to test reasonableness?
- A) Apply the recommendation to past trip data and compare with actual fuel consumption ✔
- B) Accept the suggestion as it is and write it into the budget as savings.
- C) Taking the most optimistic savings percentage as the only real value
- D) Just beautify the graph of the SFOC curve
Explanation: The savings claim produced by AI should be tested by comparing it with the ship's actual voyage data (recorded fuel consumption, load, speed, weather). Applying it retrospectively to past voyage data and comparing it with actual consumption shows how realistic the recommendation is.
10. A draft survey/technical report prepared by AI contains technical values that are fluid but of uncertain origin. What is the correct behavior?
- A) Presenting the report as it is because the text is fluent
- B) Match each value to the original measurement/source and remove the unconfirmed ✔
- C) Rounding and glossing the values
- D) Just read the executive summary and skip the rest
Description: Each technical value in the report must be traceable to its original source (measurement, calibrated device record, class rule, manufacturer data). No number should be submitted to class or management without verification of its source, even if it is a value in a fluent sentence; Otherwise, a fictitious value enters the decision base.
11. What is the best step before using an AI-generated Excel/Python calculation (e.g. ballast condition and moment of stability)?
- A) Applying the calculation directly to the actual loading plan
- B) Just looking at the length and layout of the code
- C) Trying edge cases by comparing the calculation with a known test scenario and an approved loading tool ✔
- D) Deleting part of the code to shorten it
Explanation: Stability and load calculations may appear correct but give incorrect results. Running the calculation on a small test scenario where you know the outcome, either manually or with a certified loading computer, comparing it to the expected output, and trying out limit/edge load cases will catch the error before it carries over to the actual loading decision.
12. You want to paste your ship's unshared design and performance data into a public AI tool and request analysis. Which is the most correct approach?
- A) Pasting the data as is for speed
- B) It is enough to just delete the ship name and share all remaining data
- C) Sharing the data and then asking the AI to delete it
- D) Checking the institutional policy and anonymizing the data or using a secure/institutional tool ✔
Disclosure: Hull form, performance curves and shipowner/yard contract data are trade secret and intellectual property sensitive; Additionally, critical infrastructure and OT system information carry cyber risk. Before sharing data, it is necessary to check corporate policy, anonymize it or use a corporate tool that does not share data.
13. AI-powered quality control at a shipyard block assembly flagged a weld seam as 'acceptable'. What is the correct engineering behavior?
- A) Consider the AI evaluation as final acceptance and approve the sewing
- B) Considering the AI preliminary evaluation as a screening tool and linking the final acceptance to the NDT result and class approval ✔
- C) Moving to the next block without examining the seam at all
- D) It is sufficient to simply visually check the appearance of the seam
Description: Welding quality is a safety-critical area that requires non-destructive testing (NDT) results and class approval. AI's preliminary assessment through images or data provides guidance, but final acceptance is given by qualified inspection and approval by the class surveyor.
14. In fleet management, AI produced a definitive recommendation for an expedition saying 'this bunker port and this speed are definitely the most profitable'. Which is the correct assessment?
- A) Implementing the recommendation directly as AI accesses up-to-date data
- B) To convey the proposal to the shipowner as a firm commitment
- C) Locking the entire expedition plan based on this single output
- D) Considering the definitive offer as unreliable and treating it as a scenario that can be verified with the current price, charter condition and safety ✔
Explanation: Bunker price, charter conditions, weather and port delays are rapidly changing and uncertain variables; AI cannot guarantee absolute profitability. The proposal should be treated as a scenario output that must be verified with the current price, charter party terms and safety constraints.