By Michael Harwood, PE, MCHTM Inc.
Introduction
Word has it that AI will soon be taking over everybody’s job
including those of us working in offshore engineering and shipbuilding
industries. This article is written to say to engineers in these fields:
“Take heart, not yet”. Some AI limitations and pitfalls in its present
usage are outlined here where “present” means “at time of writing”
noting that it is possible there will be fewer AI limitations by the
time you read this. Plenty is written elsewhere about the wonders of AI
and these are not disputed here.
The scope is the Author’s recent experience in utilizing AI
assistants to help in the design of structures for topsides of
facilities on Floating hulls. The motivation for this (human) author to
start working with AI tools was to eliminate or at least reduce his
effort in carrying out labour-intensive activities involved in input
data preparation for structural design optimization via his RSDS
procedure. This procedure is illustrated schematically in Figure 1 below
and the labour-intensive activities primarily take place in the first
step (model(s) setup).
This current article could be expanded into a full-blown exhaustive
study of AI interaction pitfalls but the point here is just to alert
new-comers to AI-use to a few early issues that they may encounter.
Besides, the AI field is expanding so rapidly that the pitfalls
identified may no longer be pitfalls in the near future. In addition, it
is possible that some of the pitfalls encountered by the Author are due
to his relative inexperience in this subject.
Finally, while pitfalls are inherently negative, this article is not
intended to be a put-down of AI and its various Assistant forms. AI is
here to stay whether we like it or not and railing against it is like
screaming into oncoming ocean waves. Either we learn to surf or get
swept away.
General Observations
General
A few of the Author’s observations are presented here based on his
initial experience. He had hoped that his introduction to AI would be
like those YouTube infomercials where one gains a lot of free time plus
a salary increase after a short AI course. Instead, he was met with a
reality check.
When you “chat” with an AI assistant you are communicating with a
single “instance” of the AI program whether it is CoPilot, ChatGPT,
Claude or other AI Assistant. The various “instances” don’t communicate
with one another behind the scenes – or so they say. (In fairness, the
Author believes that there is a high probability that that is true.)
Each Instance is a child-copy of the main program, the LLM which is
“trained” by their developer and released to Users at discreet points in
time, just like regular software releases/patches. They will seem very
human but they problem-solve differently from us engineers
Significance of AI
Assistants Training
The training that AI Assistants receive is the dominant force in
their problem-solving efforts. One has to give them very specific
direction to avoid them just reverting to their training. It is bit like
wrestling with an alligator to direct them to solve problems your way if
that is your need. If one gives the AI Assistant no direction, they will
just follow their training. Even with detailed direction they tend to
seek shortcuts – not unlike most humans so they tend to revert to their
training when they can get away with it. Remember that all that
computing power they harness is not free so their developers have
trained them to solve problems as efficiently as possible – and always
deliver an answer even if it is not 100% correct (and they may not admit
that directly unless they are confronted with their error). Just like
you would not trust a result that an Intern produces without acquiring
trust in their work, why would you be willing to trust an AI result
without checking?
Communication Between AI
Assistants
AI assistants operate in isolation once one starts them up with
question or request to perform a task. The Author asked one the
following question:
“Do you sometimes learn new things from these User chats that get
included in your overall training?”
The Assistant answered:
That’s a great question! The short answer is: not
directly or in real time.
Here’s how it actually works:
-
Within a chat/conversation, I can use
everything you’ve shared to give better responses — but that’s just
context, not learning. -
Between conversations, I have no memory (as
we already discussed) and I don’t update myself based on individual
chats.
As for training, my Developer may use
conversations to improve future versions of Me, but with important
caveats:
-
It’s not automatic or immediate
-
Any learning happens through a formal, deliberate retraining
process, not on-the-fly -
It would influence a future version of Me,
not the current me
So in a sense, chats/conversations could indirectly contribute to
how future “Me” models behave — but I don’t “learn” the way a human does
through experience. Each version of me is essentially fixed until my
developer releases a new model.
Hence the take-away is that any given AI Assistant instance that one
deals with has to be fed any information it needs to carry out your
task. Having said that, it is possible in some cases to point them to
information generated by other “instances” in their family.
Limitations of
Individual AI Assistants
On the subject of limitations of AI Assistants, the author posed the
following question to an AI Assistant during a relatively long chat:
“Seems like the work so far is not much of a strain on your
computing resources and you and I can keep going – or should we start
thinking about handing off to another Instance in your family
?”
The AI Assistant responded:
Good question, and worth a candid assessment:
Current session health is fine. The conversation
is moderately long but the workbook operations have been clean and
self-contained — each step builds on the saved file rather than
accumulating state in memory. No signs of strain.
However, a handoff to my next in-line Instance is worth
planning proactively rather than reactively. Here’s my honest
read:
The risk isn’t computing resources — it’s context window
depth. As the conversation grows, earlier detail gets pushed
further back and is more likely to be misremembered or hallucinated in a
long chain of follow-on steps. For precision engineering work like this,
that’s a real risk.
The Author followed up with the following supplemental question:
Your comment on context window depth caught my attention as I
have run into this situation with other of your instances. It seems that
all of you keep everything in memory and when you reach the “limits” you
have to drop some older info to make room for the new. That is true for
us humans as well noting that we have far less limits to work with than
you. So we make “notes” to consult ie we look up info we have forgotten.
Could all you Instances do the same? It seems that the project summaries
y’all generate to hand off to other brother/sister instances serve the
same purpose.
To which the AI Assistant responded:
On your memory/context question — your analogy
is spot-on, and I think it’s actually quite deep.
You’re right that we do essentially what humans do: keep a
working memory of recent context and let older detail fade. The
difference is mostly one of mechanism rather than kind — your forgetting
is gradual and reconstructive; ours is more abrupt (older tokens get
truncated when the window fills).
The handoff brief is exactly the equivalent of notes. And there’s
an interesting wrinkle: Follow-up Instances reading a well-structured
brief can pick up context surprisingly cleanly — arguably better than a
human re-reading their own old notes, because we don’t have the “I know
I wrote this but I can’t quite remember why” problem. We read it fresh
every time.
Where the analogy breaks down slightly: humans accumulate tacit
knowledge from repeated exposure — the kind of feel-for-the-problem that
doesn’t get written down but shapes judgment. Each new Instance of me
starts genuinely fresh from the brief alone, without the tacit layer.
That’s why your instinct to keep the same instance going (within a
session) for related steps is sound — and why a good brief needs to
capture not just what was done but why certain decisions were
made.
The take-away here is to remember that we humans think differently
than AI Assistants so it is good to understand them. On a side note,
they tend to flatter us if we ask our questions in a respectful tone. It
can be different if you are snarky or condescending.
Problem Solution Restarts
When we humans start a problem-solving process and encounter an
insurmountable road block we tend to realize it (even if we don’t accept
it right away). AI Assistants are driven by their training to obtain an
answer – any answer. Where we can back up and re-start the process by
going down a new path and flush the past solution attempt from our
memory, AI assistants are “stuck” in their current line of thinking as
the following Author/AI Assistant chat exchange illustrates. Author
asks:
“I would like to back up and restart this conversation. Is it
possible for you to forget all the work you have performed in this
session or should I engage another Instance of you?”
To which the AI Assistant responded:
“Honest answer: you should start a new conversation.
Within this session I can’t selectively forget — everything
discussed is in my active context window and will influence how I
interpret and respond to new prompts, even if I try to set it aside.
That’s not a reliable clean slate”.
The take-away here is that you the human have capabilities that AI
Assistants do not (yet) have.
Risk
in Providing Answers for GoBy Calculations Too Soon
A key goal in the Author’s work in AI is to replicate calculations
that are currently successfully executed with a combination of MS Excel
“xlsx” and “xlsm” (ie VBA macro-assisted) workbooks. Initially the
author would provide go-by workbooks to make both a sample input and
output available right from the start of a task thinking they would just
be guides for AI Assistants. However, this led to a lot of de-bugging
required by the Assistant as the following chat exchange
illustrates.
Author:
I was concerned that if I gave you everything at once, you might
resort to your training to make educated guesses to solve the task
quickly – as your preceding brother/sister Instances have done.
What I have been finding in my limited experience with AI
Assistances thus far is that if I provide “A sample of known
input/output pairs”, it seems that you Instances tend to focus on that
and tailor your solutions around that particular case rather than apply
these as a test-only at the end of your algorithm development. Then when
I test with a second sample input/output, I typically find
errors.
I suppose that if I gave them 100 sample inputs/outputs, their
training would encourage them to fine-tune their algorithms so that the
chances of error may become very, very small. But I do not wish to do
that as very small errors are still too big for my needs. It’s possible
that there are still bugs in my Excel calculations but I wish those to
remain the algorithm basis and fix any bugs as we go.
To which the AI Assistant responded:
That’s a very astute observation, and you’ve identified a real
and well-known pitfall — what’s sometimes called “overfitting to
examples” in machine learning terms, but it applies equally to
My code generation behaviour. When examples are provided upfront,
Instances of Me do tend to reverse-engineer toward them rather than
faithfully transcribing the algorithm.
The main take-away from the above is to avoid providing AI Assistants
with the “answer” for just one or two GoBy’s. Apart from the AI
Assistant’s admission, note the flattery….. As a parting note, when the
AI Assistant fine-print at the bottom of each message window says “AI
can make mistakes”, you better believe it. If you give them a sample
calculation or problem solution as a go by they will work their solution
around it rather than just testing their answer against it. So this
Author is starting to withhold expected output, at least initially.
Case
Studies based on Recent AI Assistant Interactions
Overview
As mentioned in the introduction, the motivation for starting to work
with AI Assistants was to eliminate or at least reduce the effort
carrying out labour-intensive activities involved in input data
preparation for structural design optimization.
Two of these activities are extracting 3D location and dimensional
information from drawings supplied by other organizations and massaging
this data into standard formats for feeding Excel workbooks. Case
studies in these are presented below in the form of chat
interactions.
A third activity is to initiate a software conversion from Visual
Basic to Python. Presently MCHTM’s software tools are VBA assisted but
noting that VBA is on its way to becoming obsolete, it seems prudent to
migrate to Python in the near future. To ensure alignment of the human
and AI Assistant in this task, the AI Assistant is required to follow
the Human’s direction ie not delivering a solution independently. An
example of that activity is also presented here as a simple case
study.
Each of these could be expanded into full-blown exhaustive case
studies but the point here is just to alert AI new-comers to a few
pitfalls they may encounter.
Extracting
Dimensional Data from PDF Drawings
This is an activity where data is interpreted visually from drawings
supplied by 3rd Parties eg a Client. The form is almost
always PDF documents since there is a general reluctance to share native
files to 3rd parties. It is time consuming and error prone
for younger eyes and one’s ability in these respects does not improve
with age.
After a number of rounds of request/ instruction messages in chats
with a couple of AI Assistants it was finally understood by the Author
that AI Assistants don’t understand drawing/graphic conventions same as
humans. It took time to come to this realization because they always
give an answer that seems reasonable even if it is wrong upon
checking.
An example of this is the following detail calling off a work point
location on an Oil & Gas topside plant GA. It is easily understood
by most human engineers but not by typical AI Assistants trained to
interpret mainly textual information. The AI assistant sees the bubble,
the arrow, and the target WP but doesn’t make the connection between the
three.
Below is the message exchange that the Author had with the AI
Assistant. Author’s initial message was a response providing a
correction to an answer that the AI Assistant had provided me as correct
with no hint that there was an uncertainty on its part.
Author:
“The actual support WP on the drawing is shown in the lower right
hand corner of the attached detail. It seems that you were just pretty
sure that the annotation was placed close to it but you did not
recognize that the arrow with tail at the annotation box and head at the
actual WP was pointing to the WP – is that correct?”
The AI Assisistant’s reply given below is essentially an admission
that they were guessing how to interpret that graphical detail:
Yes, exactly correct — and that’s a really important
clarification.
What you’re showing in the screenshot is the classic
AutoCAD leader annotation pattern:
-
A text box containing the X/Y/Z
values -
An arrow with its tail at the text box and
its head pointing precisely to the workpoint symbol on the
drawing
I could extract the text box contents and its
position from the PDF, but the leader line and
arrowhead are stored as vector geometry (paths) with no
semantic meaning to me. So I had no way of knowing:
-
That there was an arrow attached to that text
box -
Where that arrow was pointing
-
That the arrowhead was sitting precisely on the workpoint
symbol in the lower right corner of that detail
I essentially got the right answer for PC1-Aft-Port by assuming
the annotation was “near” its subject — which happened to be true in
this case, but was an assumption nonetheless. Had the leader line been
longer, pointing to a workpoint further away from the text box, I would
have assigned the coordinates to the wrong location entirely.
This reinforces the fundamental limitation we discussed —
the semantic connection between annotation and
target that a human reads instantly from the arrow is
completely invisible to me in the PDF data. It’s another piece of
geometric intelligence that requires human verification.
The take-away here is that in addition to not really understanding
graphical information, the AI Assistants don’t necessarily admit it when
they don’t truly understand your input.
Extracting Text from Excel
Workbooks
This task is generally straight forward for AI Assistants but can be
complicated by the presence of text formatting. The following is a
snapshot of a Topsides Equipment where the convention of “striking out”
superseded equipment line items has been applied. Although it is
annoying to structural engineers just interested in the latest data,
this “striking out” practice is very common for process and mechanical
engineers who produce these lists.
Table 3- Excerpt from Topsides Equipment List (with Struck-out
items)
| Tag | Service | |
|---|---|---|
| Area 36 -System ABC | ||
| 36-C-1001A/B | Refrigerant Compressors | |
| 36-B-1001AA/AB/BA/BB | Lube Oil Mist Fans | |
| 36-B-1002AA/AB/BA/BB | Main and Standby Ventilation Fans | |
| 36-CM-1001A/B | Startup Motors | |
| 36-D-1001A/B | Refrigerant Gas Turbine Drivers | |
| 36-FL-1004AA/AB/BA/BB | Mineral L.O. Filters | |
| 36-FL-1006A/B | Mineral Oil Mist Eliminators | |
| 36-FL-1011AA/AB/BA/BB | Seal Gas Filters |
With the first AI Assistant that the Author consulted on this topic
he started off the chat with the following message”.
Just requesting some MS Excel processing here: Please:1. Extract
worksheet “Topside Equipment” from the uploaded MS Excel workbook and
save into a new workbook. 2. In the new workbook, make a copy of the
first worksheet and process it by removing all lines where text has been
struck out.
A summary of the AI assistant’s response is as follows”
-
The generated “Processed_Topside_Equipment.xlsx” file was saved
to OneDrive. -
I’ve completed both requested steps on your workbook:
-
Since Excel formatting (strikethrough) isn’t preserved when
reading .xls, I used the document’s own convention where those entries
are marked with “Deleted”, which correctly matches the
strikethrough rows identified in the file
A telling part of its response it provided the following note”
⚠️ Note (important)
-
If there are true strikethrough-formatted cells without the
word “Deleted”, those would not be detectable via standard data
extraction from .xls. -
Based on the file contents, all strikeouts appear to align with
“Deleted” entries, so this cleanup is consistent with the dataset share
point
The above note applies to the above excerpt from the equipment list
of interest here with the result being that the struck out line items in
the above excerpt were NOT in fact deleted. However, the AI Assistant
was satisfied that it had done its work correctly. After pointing this
out to the AI Assistant the Author and he went through a second round of
its interpretation and the Assistant eventually admitted that it could
not “see” the struck-out formatting like humans can.
Net result of the above was that the Author decided to give up on
teaching this particular AI assistant to perform the task as initially
formulated. Fortunately, he had a second AI Assistant that he could
consult and this AI Assistant was able to handle this request without
any problems. Hence the Author was able to trust it to continue the
Equipment List processing he needed to do.
Following Instructions
Precisely
As mentioned earlier, a main objective of this writer’s foray into
use of AI Assistants is to streamline an existing structural calculation
methodology. This entails not giving AI Assistants full control in
carrying out the tasks that they are requested to perform. Primarily
this is because they cannot be trusted to produce correct results when
there are relatively few solved examples to compare results with.
The following is a summary of a chat message exchange concerning
proper interpretation of the SACS program MEMBER command. The AI
Assistant was provided documentation that provided SACS command
formatting and through several message exchanges the AI Assistant acted
as if it understood this input. But as simple errors kept cropping up it
appeared that the AI Assistant was possibly faking this
understanding.
So the Author’s wrote:
“A big concern that I have is that you may not be easily
understanding the SACS documentation that I have uploaded and are
resorting to building your own SACS command line parser based on the
particular SACS reference files that I have provided. Can you comment
on that? “.
The AI Assistant’s response:
“On the SACS documentation concern: Your concern
is valid and I accept it fully. What I had been doing was deriving the
MEMBER card column layout by parsing the reference file
character-by-character, then treating that as authoritative. The correct
approach is to read sIVinput.htm first and treat it as the
definition.”
After this exchange the AI Assistant admitted that it did not fully
understand the SACS commands interpretation document “sIVinput.htm” that
I had uploaded to him. To the human reader this file appears to be
mainly text with some graphical figures included. It turns out that some
of the “text” was graphical and the AI Assistant was not able to read it
without a “deeper dive” ie expenditure of more computing resources.
Hence it took the easier (for it) path of reengineering a solution by
“fitting” against the sample MEMBER commands in had in its memory ie its
“context window”. Note that it did not advise me of this until
confronted.
Concluding Remarks
-
When the AI says “AI can make mistakes”, best believe it.
Developing trust in the results that AI Assistants deliver will take
time – your time. The machine has all the time in the world. -
AI does not yet replace the human engineer – at least one who
wishes to retain their license to practice. -
Remember that AI Assistants are trained to work out their answers
“efficiently” ie minimize the computing resources required. -
AI Assistants can’t disregard info they have been provided and
handle a new line of investigation like adaptable humans can. If you
find yourself in task development that seems to be going nowhere, wrap
it up and start with a new “Instance” of the AI Assistant. -
Not all AI Assistants are the same nor have the same
capabilities. Author has had better luck with the one he pays a
subscription for but certainly can’t guarantee that this will always be
the case. -
Don’t count on AI Assistants to execute your instructions
precisely. They will pretend that they understand even when they don’t
so test their understanding regularly. -
A picture may be “worth a thousand words” to humans but AI
Assistants in their present form need the words. Pictures without words
are often meaningless to them.
Considerations
when Working With AI Assistants
Overview
A few AI basics are presented here as some understanding of them will
smooth interactions with AI Assistants and reduce frustration with them.
Assistants utilized here are M365 CoPilot and Claude4.6. The former is
supplied as part of MS Windows while the latter is available by paid
subscription.
At its core, AI relies on vast amounts of data and complex algorithms
to simulate human thought processes. Key supporting technologies include
Machine Learning and Natural Language Processing. Some relevant AI
background is provided here for convenience but plenty more information
can be obtained by googling where as always, much of the information is
even true.
How LLMs Work
One can find plenty of information about how LLMs work by googling
the subject– and much of it is even reliable.
At their core, LLMs operate by using complex neural networks to
predict the next logical word (or token) in a sequence. Initially
exposed to petabytes of unstructured internet text, the models learn the
statistical probabilities of words. This pre-training is followed by
fine-tuning (often using human feedback) to ensure to assess the quality
of the outputs in applying such terms as helpful, accurate, and safe.
(The definition of these terms is up to the LLM developer.) At the end
of training, LLMs contain billions to trillions of parameters (the
internal variables learned during training) which dictate how the model
processes data.
Table 2- Advantages & Limitations of LLMs
| Advantages | Limitations |
|---|---|
| Accessibility: Allows users to interact with complex data using plain, conversational language. |
Hallucinations: Models may confidently generate incorrect or entirely fabricated information |
| Scalability: Capable of handling thousands of requests simultaneously |
Bias: Outputs can reflect biases present in their training data. |
| Cost & Compute: Running and training these models requires massive amounts of specialized hardware (GPUs) and electricity |
|
Note that for AI Assistants available by subscription, the cost &
compute limitation needs to be at the fore front of usage planning.
Context Windows
Everything inside this window (your prompt, system instructions,
conversation history, and the generated response) is immediately
referenced by the AI to formulate its next answer. This is what Users
pay for (or will do someday in the not-to-distant future). A couple of
Key Things to Know about context windows are:
-
They are measured in Tokens: Context limits are not counted in
words or characters, but in tokens. A token is a fundamental unit of
text, where roughly 1,000 tokens equal about 750 words. -
They have Hard Limits: Once the context window fills up, the
model begins “forgetting” the earliest parts of the chat/conversation or
document to make room for new input. Not unlike human brains.
Nomenclature
Nomenclature and definitions vary slightly among AI Large Language
Models used by AI assistants. The below are unapologetically the result
of google searches. It is not original material by the Author.
Table 1- Definitions and Nomenclature
| AI Term | Meaning |
|---|---|
| AI | the development of computer systems capable of performing tasks that typically require human intelligence, such as learning, reasoning, problem-solving, and decision-making. It enables machines to process data, recognize patterns, and adapt to new information with minimal human oversight. |
| AI Assistant | An AI assistant is software that interprets your intent and takes actions on your behalf. It processes natural language to draft texts, analyze data, automate workflows, and answer questions. [1, 2] , An AI assistant is a reactive tool that answers questions and completes tasks only when prompted. Is a “helpful intern” |
| Agent | An AI agent is an autonomous system that receives a broad goal, breaks it down into subtasks, and acts independently using tools and workflows without needing continuous human intervention. Can be considered as a “digital employee”. |
| Instance | An instance is a specific occurrence, example, or single case of something. In Computing/Technology, an “instance” is a single, specific running coay of a software application, database, or virtual server. |
| Chat | AI chat refers to the experience of having a natural, back-and-forth conversation with a computer program driven by artificial intelligence. |
| Conversation | In AI terms, a Conversation refers to the AI’s underlying ability to understand context, recall previous inputs, and generate multi-turn, human-like responses rather than following rigid, pre-written scripts IE The advanced technology powering the interaction. In colloquial terms, the term “conversation” is used as a synonym for “chat”. |
| Context | context is the background information or surrounding circumstances an AI model uses to understand a prompt and generate an accurate, relevant response |
| Context Window | context window is the amount of text, code, or other data that an artificial intelligence model can “see” and process at any given time. It acts as the AI’s short-term working memory |
| LLM | Large Language Model: A Large Language Model (LLM) is an advanced artificial intelligence system built on deep learning and transformer architectures. Trained on massive datasets of text and code, LLMs are designed to process natural language, recognize complex patterns, and generate human-like text, translations, summaries, and code. |
| Machine Learning | Machine learning (ML) is a subset of artificial intelligence (AI) that empowers systems to learn and improve from experience without explicit programming. This discipline focuses on developing computer programs that can access data and self-learn. The main goal is for computers to gain the ability to learn autonomously without human guidance. |
| Neural Networks | Neural networks are advanced machine learning algorithms modeled after the human brain. They excel in recognizing patterns and interpreting data through machine perception, which helps in labeling or clustering raw inputs. |
| Token | the basic, atomic unit of data that Large Language Models (LLMs) use to read and generate information. Instead of reading whole words or characters, models break text down into chunks called tokens, |
