One day,
you spent a long time writing an article.
Not asking ChatGPT to write it.
Not having Claude rewrite it.
But:
You thought it up yourself,
wrote it yourself,
deleted parts yourself,
and rewrote it yourself.
Then you put the article into an AI detector,
and the screen shows:
100% AI.
How would you prove that:
"This was really written by me"?
This year, a prize-winning short story faced exactly this dilemma.
What happened in the end was even more striking:
An AI claimed the text looked very much like AI-written, but the human organizers examined the creative evidence and chose to believe the author.
A Prize-Winning Story Suddenly Tagged as "100% AI"
The story is titled:
The Serpent in the Grove.
The author is Jamir Nazir from Trinidad and Tobago.
The piece first won the Caribbean regional award in the 2026 Commonwealth Short Story Prize.
It later secured the overall grand prize for the entire competition.
But shortly after the regional winners were announced,
suspicion quickly arose online:
Was this story AI-written?
People noticed its sentence structures,
its metaphors,
its rhythm
seemed very similar to text commonly generated by large language models.
Someone then ran the whole text through an AI detection tool called Pangram.
The result:
100% AI-generated.
The situation escalated from "readers feel something is off"
to:
"Even the AI detector says 100%."
But what does 100% really mean?
Seeing "100% AI,"
people naturally interpret it as:
The system has determined:
Every single word was written by AI.
But that’s not the case.
When Nature recently investigated the new generation of AI detection tools,
they interviewed Max Spero, founder of Pangram.
He explained that Pangram first splits the article into different segments,
then judges each segment as more likely:
written by a human,
AI-written,
or a mixture of both.
The final overall result is based on the proportion of segments classified as AI-written.
So:
100% AI does not mean the tool knows the author of every word was AI.
More precisely,
each analyzed segment is predicted by the model to be "more like AI."
This is a classification outcome,
not a video recording,
not a digital fingerprint,
and definitely not the detection of ChatGPT typing the text in person.
This is like facial recognition saying "looks similar," not a court verdict
AI detectors fundamentally perform pattern recognition.
They might analyze:
sentence arrangements,
which words frequently co-occur,
paragraph rhythm,
language structure,
semantic patterns,
and compare these against vast amounts of:
human writing
and
AI-generated text.
The final answer is:
"Which side does this piece resemble more?"
These tools can be very useful,
and sometimes quite accurate.
But "highly accurate" doesn’t mean "never wrong."
Even if a system’s error rate is very low,
scanning hundreds of thousands of student essays,
millions of job applications,
and numerous submissions will inevitably produce false positives affecting real individuals.
The person mistakenly flagged faces not:
"On average, this system is accurate,"
but rather:
"How do I now prove I didn’t cheat?"
Commonwealth Foundation chose not to fully rely on the AI detector
The key takeaway from this incident lies not in how accurate Pangram is,
but in how the contest organizers handled the situation.
The Commonwealth Foundation said they spent about a month investigating claims of AI use.
They intentionally did not treat:
"What the AI detector said"
as final evidence.
They held detailed discussions with all regional winners about their creative process,
reviewed:
drafts,
documents with timestamps,
notes,
and evidence showing how the story developed step by step.
On June 22,
the Commonwealth Foundation publicly stated
that after consulting judges and considering all available information,
they believed AI was not used to compose the winning works.
Jamir Nazir retained his regional title,
and on June 30 was officially awarded the overall 2026 Commonwealth Short Story Prize.
This created a stark contrast:
AI detector: 100% AI.
Thorough human review: accepted the author’s creative proof.
So who should we trust?
This isn’t a simple matter of:
"See, AI detectors are garbage."
That’s not accurate either.
Newer-generation AI detectors have made significant progress.
Nature’s August 25 investigation showed that better-performing tools,
like Pangram and GPTZero,
outperform earlier models in recognizing purely AI-generated text.
In some tests, they nearly detect all AI-written content.
Therefore:
AI detection is not worthless.
The real question is:
What do we use it for?
Using AI detection as a clue versus as a verdict are entirely different
If a teacher receives an essay flagged as highly likely AI,
a reasonable next step is to:
review the student’s past writing,
ask how the article was developed,
request drafts,
have the student explain their arguments,
and verify editing records.
In this case, the AI detector serves as:
a reminder.
But if the process instead becomes:
AI detector shows 95%,
so the essay is simply given zero,
then the tool shifts from:
providing clues,
to
delivering a verdict.
These two approaches carry very different risks.
AI and human writing are increasingly blended
There’s also a more complicated aspect:
Previously, we could imagine:
Category A:
Entirely human-written.
Category B:
Entirely AI-written.
Then detectors just distinguish A from B.
But the real world is no longer that clear-cut.
Some people:
write the first draft themselves,
then use AI to fix grammar.
Others:
do their own research,
then use AI to organize structure.
Some:
translate interviews with AI,
then rewrite themselves.
Others:
use AI to brainstorm,
then write the entire story themselves.
Still others:
use AI to write the first draft,
then do extensive rewriting themselves.
So the real question now might not be:
"Human or AI?"
but rather:
"To what extent was AI involved?"
This is much harder than a simple binary choice.
Nature encountered an interesting case
When Nature recently tested Pangram,
they scanned a research article about publishing.
The older version of Pangram classified it as:
100% AI.
Author Elena Vicario said
the original draft and core ideas were hers alone,
and AI was only used for text polishing.
The article then underwent human editorial revisions.
This shows that even if detectors pick up
"AI traces in the text,"
they still cannot precisely determine:
How much of the content was AI-generated?
Pangram’s founder admits
there is much room for improvement in handling
texts created through mixed human writing, AI polishing, and human editing.
Minor changes can significantly alter detection scores
Nature also mentioned a phenomenon called:
jitter.
This means:
for some documents with mixed human and AI edits,
small modifications can cause detection results to shift noticeably.
Also, analyzing a single sentence in isolation may yield different results from analyzing it within the full article.
This is because detectors do not just assess:
"Does this sentence look AI-written?"
They are influenced by:
contextual paragraphs,
how text is segmented,
and the overall article pattern.
Therefore, claiming:
"This sentence is 87% AI-written,"
can easily create a false impression of precision beyond the tool’s actual capability.
The impact on students may be greater than on writers
Novelists can at least:
respond publicly,
accept interviews,
and present creative records.
Imagine a college student,
applying to graduate school,
spends weeks writing a personal statement,
only for the school’s AI detector to flag it as AI-generated.
Nature reported in July that some U.S. graduate application portals warn:
If personal statements are deemed AI-written,
applications may be rejected outright.
In such cases, a detection error is not just a lost grade,
but a:
lost chance at admission.
So tool accuracy is only part of the equation.
An even bigger question is:
Does the affected person have any chance to appeal?
Fairness requires preserving the creative process
Jamir Nazir’s case was reconsidered largely because:
The organizers didn’t rely solely on the final work.
They looked back at:
drafts,
timestamps,
notes,
and the creative process.
This sends a practical reminder to everyone using AI today.
If your work highly values:
originality,
academic integrity,
authorship,
and the creative journey,
the best proof may not be battling with detectors at the end,
but rather keeping from the start:
version histories,
drafts,
research notes,
sources,
and revision records.
Because those truly answer:
"How was this content created?"
Not just:
"Who does this text look like it was written by?"
Writing "too much like AI" might influence human writing styles
If this trend continues,
it could lead to a strange side effect.
Suppose we all learn that:
certain sentence patterns,
structures,
and writing habits
are more likely to be flagged by detectors.
What might humans do?
They could deliberately avoid those sentences,
write in irregular ways,
or intentionally make their writing look less like AI.
Nature Human Behaviour warned this year:
If schools and publishers over-rely on fixed linguistic features to detect AI,
they could end up changing the nature of human writing.
In other words:
We’re not teaching AI to mimic humans;
instead,
humans start changing their style just to prove they aren’t AI.
This would be a very strange world.
The greatest danger is forgetting AI detection is a judgment, not a fact
The biggest illusion AI detectors create is their:
slick,
confident,
seemingly certain percentages.
87%,
96%,
100%.
Those numbers look objective,
but remember,
there’s still a model behind them.
It has not seen:
you sitting at your desk late last night writing.
It hasn’t seen:
which sentences you deleted.
It doesn’t know:
why that metaphor popped into your head.
It only sees:
the final pattern of the text.
And then it classifies.
So a mature approach isn’t to:
entirely reject AI detection tools,
but rather understand:
what they can answer,
and what they cannot.
They can say:
"This part deserves further review."
But if the outcome affects:
grades,
admission,
reputation,
awards,
or job opportunities,
the final decision cannot just be:
"Because AI says you look like AI."
What’s truly needed are:
more evidence,
full context,
and a human process willing to reconsider.
Because in a world where AI increasingly mimics humans,
we may face a paradox:
One day, the biggest problem for humans may not be being impersonated by AI,
but:
having to prove what they wrote isn’t AI-generated.
Today, let’s improve alongside AI.
Learn an AI skill daily.
Save a little time every day.
Enhance abilities bit by bit.
SasaDaily, growing with you.
Recommended Reading
Your Resume Is for Companies to See, so Why Does AI Sometimes Score It First Before Humans Read It?
