We Let AI Write the Articles, and “However” Started Growing in Every Paragraph — How We Put Research Back Behind the Curtain
1. The incident: every joke summoned a peer reviewer
One article factory was built around a simple idea: pick up the funniest observations from conversation, then use research or statistics to dig a little deeper.
The intended flow was straightforward.
Someone says:
“Choosing one café has somehow turned into a procurement department.”
Then the article asks:
“Why do people get tired when they compare too many options?”
Or someone notices:
“Did this person learn one rule in a relationship and then apply it to the next person too?”
Then the article checks whether people tend to reuse behaviors that worked before.
So far, so good.
The problem began after the factory was repeatedly trained to be “more accurate,” “more research-based,” and “careful with uncertainty.”
A tiny peer reviewer moved into the AI.
Write a joke.
AI: “Strictly speaking, this is not identical to that psychological concept.”
Write a metaphor.
AI: “However, this individual cannot be explained by that theory alone.”
Make an exaggerated observation.
AI: “The research does not support such a general conclusion.”
Who invited you?
The reader is not an author responding to reviewer comments before publication.
2. Reading recent articles side by side made the pattern obvious
The strongest pieces had one clear spine.
A person enters a place, something unexpected happens, the experience changes how the price feels, and the story ends with a punch line. Official information and research sit on the side and support the experience without taking over.
The weaker pieces were the ones where psychology or medicine could easily multiply.
One joke would suddenly produce ownership, attachment, learning, cognition, trauma, decision-making, well-being, and five more concepts.
And each concept arrived with its own brake pedal:
“However, this has not been proven to be the same mechanism in this case.”
Medical pieces had another version of the same problem:
“Small sample.” “Not 100 percent.” “Cannot guarantee.” “Individual differences.”
The same limitation appeared in the main text, the FAQ, and the conclusion.
The information was not wrong.
The research was often strong.
But the reading experience had changed from:
“fun article”
to:
“fun article with a reviewer-response document stapled to it.”
3. The worst mistake was turning jokes into academic claims that needed correction
“Come on, that is basically X, right?” is usually not a journal conclusion.
It is a joke.
A metaphor.
An observation.
A working hypothesis.
But a precision-obsessed AI can silently convert it into a formal claim and then correct the formal claim it invented.
The bad flow looks like this:
comment → convert it into an academic label → explain why the label is not strictly exact → explain the study limitations → return to the original comment
That round trip is expensive.
If someone says:
“Choosing one café has turned into a procurement department,”
the article only needs something like:
“The more criteria you add, the more energy you spend before making the decision. People who keep searching for the best possible choice may also keep wondering whether something better existed even after they decide.”
That makes the joke deeper without killing it.
If the next sentence starts with:
“However, more options do not always reduce satisfaction under every condition…”
the procurement department has now hired a legal department.
4. Research was not the villain; it had simply walked onto center stage
The fix is not to delete uncertainty or evidence.
In four experiments on science news, adding caveats made reports seem more cautious without reducing interest or comprehensibility. A 2026 systematic review of 24 studies also found that communicating scientific uncertainty generally supports trust, while hiding uncertainty can undermine it.
So the enemy is not the word “however.”
The enemy is charging the reader for the same “however” four times.
Research on cognitive load describes a redundancy effect: unnecessary repeated information can impose extra processing because the reader still has to determine whether the repeated material adds anything new.
That led to a simple editorial rule:
Do not remove uncertainty. Do not make the reader read the same uncertainty twice.
“An 11-person study found X. Because the sample was small, it cannot guarantee that X happens every time.”
Done.
Do not split the same warning into “not absolute,” “not guaranteed,” “may vary,” and “individual differences” across four sections.
5. The better flow is comment → structure → research → punch line
The new default begins with the best thing in the conversation.
For example:
“This looks like a mouse voluntarily walking into a trap.”
Then explain the structure in normal language:
“If suffering itself starts to feel like proof of effort or worth, a harsher environment can feel more meaningful rather than simply worse.”
If research helps, bring in one useful piece:
“Psychology has studied patterns in which effort, sacrifice, and self-worth become tightly linked.”
Then return to the joke:
“So the trap did not necessarily get smarter. The mouse may have read the job ad and thought, ‘This place looks like a growth opportunity.’”
That is enough.
Research is lighting crew, not the lead actor.
It should not walk to center stage and announce:
“Please note that this particular mouse has not been clinically diagnosed.”
6. We also dropped the universal template: “the five-second conclusion” is not for everything
Another problem was that the factory’s mold became visible.
Every article started to look like:
five-second conclusion
↓
thirty-second overview
↓
research
↓
FAQ
↓
summary
That works well for practical questions.
If someone asks what they can eat before a medical checkup, the answer should arrive early.
But a travel story dies if the first paragraph says:
“Conclusion: by the end, I reinterpreted the restaurant as live entertainment and decided the price was good.”
The show is already over.
A comedy piece also suffers if the punch line appears in the first five seconds and the remaining article becomes an explanation of why the punch line was funny.
So the structure now follows the genre.
Experience pieces use scenes and chronology.
Comedy pieces use escalation and callbacks.
Practical pieces lead with the answer.
High-risk medical, legal, financial, or safety pieces surface important warnings early.
Research articles explain methods and limitations only when the research itself is the subject.
Not every article needs to come out of the same mold.
7. The factory itself had to change: one apology prompt was not enough
A good editorial rule in one conversation is useless if the automation wakes up an hour later with an old prompt.
Then the next scheduled run says:
AI: “Understood. According to research—”
and the memory loss begins again.
So the fix went into the factory, not just the prose.
The highest-priority editorial rules now say:
- comments, jokes, metaphors, and hypotheses stay in the lead;
- research explains, supports, and extends them;
- only clear factual errors or harmful misunderstandings receive direct correction;
- the same caveat should normally appear once;
- repeated “research says → however” rhythms fail quality review;
- jokes should not be explained twice;
- research must not bury the original observation;
- one universal article template is forbidden;
- translation must not amplify reviewer-style language.
The same rule is now read during generation, editing, multilingual production, foreign-language quality review, formal adoption, and pre-publication checks.
In other words:
if the peer-reviewer AI comes back, later stages can reject it too.
And one bad article does not stop the entire factory.
Fix the bad item and keep the line moving.
Otherwise the solution to “too many caveats” becomes “no articles at all,” which is not editing. It is a shutdown.
8. The new quality check: symptoms of peer-reviewer AI
From now on, these are warning signs:
The same warning appears twice with different wording.
An obvious joke is immediately followed by “strictly speaking.”
The research section is longer than the original observation.
One joke attracts five or six academic concepts.
Every heading starts with “research shows” and ends with “however.”
A joke is told, then explained again.
A travel story reveals its ending in a five-second summary.
A local-language article keeps interrupting itself with unnecessary foreign technical terms.
And the most important test:
If you remove the research, does the article still have a spine?
If not, research has become the protagonist.
If sending the backstage crew home makes the whole stage disappear, you were not writing a conversation-driven article anymore. You were writing a research explainer.
9. Conclusion: nobody hired a logician
Accuracy matters.
Sources matter.
Caveats matter when a mistake could affect health, law, money, or safety.
But readers did not come to see the AI’s internal audit log.
A human says:
“This is basically X, right?”
The article asks:
“Why does it look that way?”
Research adds:
“Here is a mechanism that helps explain it.”
Then the article returns to the human observation.
That order works.
Research is not a hammer for hitting the comment until it becomes academically acceptable.
It is scaffolding behind the comment.
The factory’s new north star can be reduced to one sentence:
Not “a funny article proven correct by research,” but “a funny article kept from being wrong by research.”
When I want a logician, I will read a paper.
Today, please let me read the blog.
References (3)
- Bott, L. et al. (2019). Caveats in science-based news stories communicate caution without lowering interest. Journal of Experimental Psychology: Applied, 25(4), 517–542. https://pubmed.ncbi.nlm.nih.gov/31246056/
- Schuster, C. & Scheu, A. M. (2026). How Communication of Scientific Uncertainty Affects Trust in Science—A Systematic Review. Risk Analysis, 46(5), e70233. https://doi.org/10.1111/risa.70233
- Sweller, J., van Merriënboer, J. J. G., & Paas, F. (2019). Cognitive Architecture and Instructional Design: 20 Years Later. Educational Psychology Review, 31, 261–292. https://doi.org/10.1007/s10648-019-09465-5
