No, it cannot guarantee that.
Yesterday, we learned a very useful instruction:
Ask Copilot to create:
Word documents.
Excel spreadsheets.
PowerPoint presentations.
Before that, add the instruction:
“Do not add any dates, amounts, numbers, commitments, or conclusions that are not present in the data.”
This sentence:
is worth including.
But do not interpret it as:
“Once stated, AI will never make mistakes by adding incorrect information.”
It cannot guarantee such perfection.
Because prompts are instructions, not rigid data locks
You tell Copilot:
“Don’t add anything extra.”
This is like telling an assistant:
“This report can only use the data on the desk.”
Under normal circumstances,
the assistant will try to follow that.
But AI still needs to:
read,
summarize,
classify,
reason,
and reorganize.
It must decide:
what is important,
what is relevant,
and which sentences should be combined.
So it can still:
make errors.
Microsoft itself doesn’t claim Copilot outputs are always accurate
Microsoft’s official reminder for Word, Excel, and PowerPoint agents is straightforward:
After generating files,
verify accuracy before sharing.
For general Copilot output,
Microsoft even provides:
a validation framework,
because a piece of content that looks complete,
flows well, and
is nicely formatted
does not mean
it is ready for direct use.
“Not inventing random new numbers” is only the first layer
When many think of hallucinations,
they think only of cases like:
The source does not have:
an amount of 1 million.
But AI suddenly writes:
1 million.
This is definitely a problem.
But in real work scenarios,
more common mistakes:
are subtler.
Type 1: Data exists, but AI misunderstands
For example, the original document says:
“Budget cap is 500,000.”
Copilot summarizes as:
“Project budget is 500,000.”
The sentences differ slightly,
but the meanings are completely different.
One means:
it should not exceed that amount.
The other means:
the amount is already confirmed to be spent.
AI did not invent “500,000” out of thin air,
but it still
changed the facts.
Type 2: Turning “possibly” into “certain”
For example, meeting notes say:
“The client might move the event to Friday.”
The final presentation states:
“The event will be held on Friday.”
The date
existed in the data,
so it’s harder to spot at a glance that
this is not a random addition.
But the real error is the
certainty level has been changed.
This is very risky in work documents.
Type 3: Two valid data sources combined into one wrong answer
For example:
An old proposal says:
Event time is 14:00.
A newer email says:
Event time changed to 15:00.
Both sources are
genuine.
If Copilot
fails to identify which is later,
it may write in the presentation:
14:00,
or even
show both times on different slides.
This is not
AI inventing new data,
but rather
choosing the wrong source.
Type 4: Vague data gets AI-completed interpretation
Original data:
“Vendor has not confirmed yet.”
To make the project summary more complete,
AI might write:
“Vendor is expected to confirm tomorrow.”
It may consider this a
reasonable supplement,
but for the company,
this is already
a new fact.
Thus,
“don’t add”
means to prevent not only
random invented numbers,
but also
turning assumptions into facts.
Type 5: Omitting critical limitations
For example, the original says:
“Plan A is expected to save 10 hours per month but only applies to clients with fixed data formats.”
PowerPoint, to shorten text,
ends up with:
“Plan A saves 10 hours per month.”
It did not add incorrect info,
but it removed
the limiting condition.
This can still
mislead.
Microsoft’s validation guidelines
also emphasize:
Copilot output
may omit context that significantly changes judgments.
Therefore,
errors don’t always come from
adding too much.
Sometimes they come from
leaving out important details.
Type 6: Excel looks most “accurate” but can still err
Excel is especially reassuring,
because it has
formulas,
charts,
percentages,
and totals.
It doesn’t look like the AI is just writing prose.
But Microsoft also warns:
Copilot in Excel
can misunderstand info,
produce inaccurate
formulas, insights,
and tables,
all of which
need checking.
For example,
AI may not invent revenue numbers,
but it might have
wrong formula ranges,
miss a row,
use the wrong denominator for a percentage,
yet the chart still looks beautiful.
So, is “don’t add if not in data” useful at all?
Yes.
And it’s worth continuing to use,
because it defines
an evidence boundary
much clearer than
“make it more complete.”
It tells the model:
do not prioritize completeness over truthfulness.
This is a very good
behavior rule.
But don’t mistake
“less likely to err”
for
a guarantee of no errors.
How to address the second layer of errors?
The simplest way is:
after AI generates content,
don’t simply ask:
“Are you sure?”
Because it’s likely to answer:
“Yes, sure.”
This adds little validation value.
Instead, ask:
“List all dates, amounts, numbers, formal commitments, and main conclusions from the document, and indicate the original source of each; mark any without a source as unconfirmed.”
This is performing:
a source check.
Microsoft now directly recommends doing source checks
Microsoft’s Copilot validation guide
suggests users check whether:
claims in outputs
genuinely reflect original data,
and important statements
can be traced to reliable sources.
Check if Copilot
mixes different info,
exaggerates certainty,
or fills gaps with assumptions.
This matches
the question we explored today.
If you can’t find a source, mark it as “unconfirmed”
Don’t assume errors are acceptable just because
AI sounds plausible.
For example:
“It must be based on its analysis.”
If the document is
internal brainstorming,
that’s fine, but mark it as
an inference.
If it’s a
formal proposal,
official report,
quote,
financial statement,
or client commitment,
then anything without source
should not be written as confirmed facts.
You can add a “fifth step” to yesterday’s four-line framework
Yesterday we had:
Deliverable.
Audience.
What must be retained.
No additions.
Today, add:
Source verification.
The process becomes:
Before generating
Deliverable:
What do I want?
Audience:
Who will see it?
Must retain:
What cannot be deleted?
No additions:
What information must not be assumed?
After generating
Source verification:
Where does each important claim come from?
This forms a
complete cycle.
You don’t need to check every sentence’s source
This is unnecessary.
Focus on information that
carries consequences, such as:
people’s names,
dates,
deadlines,
prices,
quantities,
percentages,
formulas,
official conclusions,
legal requirements,
client demands,
external commitments,
and any statements that may influence
decisions.
General sentences like:
“This project aims to improve efficiency.”
may not be worth verifying word-for-word.
But,
“Costs are expected to reduce by 30%”
definitely requires asking:
“Where does this 30% come from?”
You can do QA by focusing on “high-risk information”
For example, after AI completes a PowerPoint,
check only:
all numbers,
all dates,
and sentences containing:
words like
“definitely,”
“confirmed,”
“decided,”
“will,”
“guaranteed,”
“expected savings,”
etc.
These are areas
where assumptions often get
mistaken for facts.
For Word, especially check if conclusions overreach source data
Example: original interviews say:
“The implementation process is somewhat complicated.”
Copilot writes:
“Employees generally oppose the new system.”
This exceeds the
original evidence.
It may not have
fabricated any interview quotes,
but the
conclusion is overstated.
Word reports
must check the
strength of conclusions
to ensure they don’t surpass
original data.
PowerPoint needs attention to meaning changes after condensing
Slide space is limited.
AI often
compresses a long conditional statement
into
a strong, definitive headline.
This looks great in presentations,
but can be
dangerous factually.
For instance:
Original:
“If Q4 demand remains at the current rate and new equipment is delivered on schedule, capacity may increase by about 15%.”
The title becomes:
“Q4 Capacity Increases 15%”
All wording is from the original context,
but the conclusion
has been rewritten.
Excel requires checking both number sources and formulas
If Copilot creates a summary table,
at minimum you should confirm:
no original rows were missed,
formulas cover the correct ranges,
blanks aren’t treated as zeros,
percentage denominators are correct,
and charts do not select wrong data series.
This is because the most dangerous part of Excel AI
is that
errors can look very professional.
Is it enough to let Copilot check itself?
No, it’s insufficient.
It can
help identify mismatches,
assumptions,
and missing context,
which is very useful.
But Microsoft explicitly says:
Copilot can assist with
output validation,
but cannot provide a final certification of accuracy.
Crucial data still
needs to be verified against the
original source.
This is the difference between
AI review
and
human verification.
AI can be a second reviewer, but not its own notary
You can ask AI to:
“Indicate statements lacking direct evidence.”
“List all assumptions.”
“Compare the original data and the presentation for consistency.”
“Identify discrepancies in numbers, dates, and names.”
These are all worthy steps.
But if AI answers:
“All is correct.”
don’t automatically close the case.
High-risk data
still requires
human confirmation.
Is using AI this way cumbersome?
Not necessarily.
The real comparison is:
Previously, you started a full document from
a blank page.
Now,
AI does 80% for you first.
You spend
10 minutes
checking the high-risk 20%.
If your total time
still decreases,
it’s worthwhile.
The issue isn’t
that AI requires checking,
but rather
not mistaking
“a quick first draft”
for
“final output.”
The most common mistake is lowering your guard after seeing a polished file
With plain text chat,
you may immediately question one wrong number.
But a
formatted Word document,
an Excel with charts,
or a beautiful PowerPoint
gives a psychological impression of
“this is already done.”
In fact,
formatting
and
evidence completion
are two separate matters.
So today’s answer can be summarized by one formula
“Don’t add info not in data”
=
a clearer
prompt boundary,
but
≠
an accuracy guarantee.
The truly reliable workflow is:
restrict first,
↓
generate,
↓
check important claims,
↓
refer back to original sources,
↓
human verification.
This echoes our September 12 question about Plaud
Back then, we asked:
Plaud Agent
made a PPTX
grounded in conversation context,
can we just
send it directly to the client?
The answer was also:
no direct inference possible.
Having context
doesn’t mean
AI’s summarization is error-free.
Today’s topic steps back one layer:
even when you tell Copilot explicitly
“don’t add info,”
it only
adds a limitation layer,
not
eliminate all errors.
The multi-document problem from August 9 is similar
AI can organize
three documents beautifully,
but if the three sources
contradict each other,
AI’s output
does not mean
the contradictions disappear.
It may even
choose one version for you.
So AI document workflows
always come down to the question:
do you know which source is the truth?
Finally, keep this in mind:
If Copilot writes a
number,
date,
commitment, or
conclusion
you don’t remember seeing before,
don’t first ask:
“Did it just invent this?”
Instead, ask:
“Can I trace this back to a source?”
If found,
check if the context has been changed.
If not found,
mark it as
unconfirmed.
This is much safer than simply trusting:
“I told it not to add info in the prompt.”
Because a good AI workflow
is not about
designing a perfect prompt that never errs,
but
having a fast method to catch errors if AI does make mistakes.
Today, progress a little with AI.
Learn one AI skill a day.
Save some time every day.
Improve your ability bit by bit.
SasaDaily—growing with you.
Recommended Reading
AI Quick Q&A ｜ What You Need to Know Before Using AI-Generated Content Directly: 3 Key Points
