The 5-second answer: AI doesn't just make writing cheaper
When you use AI to produce articles, the time spent writing goes down. But that's not where the real savings are.
Editing, translation, quality checks, internal links, publishing, checking the live site, re-inspection and anomaly monitoring can all shift from "work a person does every single time" to "a process that runs by rules." Even while the humans sleep, the factory's night shift keeps going.
The catch: the easier mass production gets, the more another problem shows up. A thin article written by AI gets read by an AI, rephrased by yet another AI, and the source vanishes somewhere along the way. The internet risks turning into one giant game of telephone.
That's where what this article calls an "information family register" comes in. (In Japan, the koseki is the official family register that records where a person comes from. The idea here is the same thing for information.)
Where did this information come from? When was it checked? What was the original data? What was changed? Who, or what, verified it?
The goal is to make all of that traceable. But you shouldn't make readers wade through a full certified copy of the register from line one of the article. Keep it short for readers, deep for those who want details, and exact for machines. Three layers.
What AI changes is how you run the process, more than how fast you write
In ordinary article production, a person walks the piece from step to step.
Research
↓
Write
↓
Edit
↓
Check sources
↓
Translate
↓
Check each language
↓
Add internal links
↓
Publish
↓
Check after publishing that nothing broke
If you use AI only a little, all you get is faster writing.
Automate the process itself, though, and the story changes.
Set the quality rules
↓
AI processes the article
↓
Automated checks
↓
Checks for meaning
↓
Fix anything that fails
↓
Re-check
↓
Only what passes moves on
↓
Keep monitoring after publishing
This looks less like a blog and more like a small manufacturing line.
In software development, the system that automatically tests every change is called CI (continuous integration), and the system that delivers what passes to the live site is called CD (continuous delivery). Applied to articles, CI is the inspection line and CD is the shipping line.
QA means quality assurance. It doesn't only look at the finished product; it also asks whether the way you make things makes defects unlikely in the first place. QC (quality control) is closer to the individual inspections inside that.
So an AI article factory isn't about hiring one "AI writer." It's about moving the work of an editorial desk, a translation team, a quality team, web operations and the publisher, little by little, into equipment.
What's expensive by hand is the surrounding work, not the text itself
Article production costs a lot for reasons beyond the hours spent typing.
Take 1,200 articles rolled out into 11 foreign languages. That's 13,200 articles x languages.
Even if a person spends just 10 minutes checking each translation:
13,200 x 10 minutes = 132,000 minutes = 2,200 hours.
At 20 minutes each, it's 4,400 hours.
And that doesn't yet include editing the Japanese, research, internal links, checking the published page, sending things back for fixes, or update work.
Outsourcing translation makes the money easier to see. According to the order-price guidelines published by the Japan Translation Federation, Japanese-to-English translation of a computer manual runs 20 yen per Japanese character. For 5,000 characters, that's a guideline of 100,000 yen for English alone. Of course this varies a lot with the field, volume, deadline and quality, and it isn't a quote you get by simply multiplying by 11 languages.
In short, a multilingual article isn't done once you've written the text. If you want to keep quality up, the surrounding work snowballs.
Turning the jobs AI takes over most easily into equipment, as a value pack
List the tasks AI can easily replace, and there are quite a lot.
- Drafting
- Tidying up the text
- Checking typos and structure
- First-pass translation
- Checking translations for awkward phrasing
- Help with verifying sources, URLs and numbers
- Suggesting internal links
- Pre-publication checks
- Regular monitoring after publication
- Re-checking the same anomaly
Each one is small on its own. Bundled together, they become a department.
What's interesting here is that, rather than "replacing" humans with AI, you package the work itself and turn it into equipment.
Nobody has to ask "what do I do next again?" every time. The next step starts by looking at the previous step's certificate of passing. Ideas I picked up building game automation and simulators, like "make the data the source of truth," "record the state," and "reproduce the same process," can turn straight into a business factory. I was building a hobby robot, and an editorial department sprouted next to it. How did this happen?
Running while you sleep isn't magic. It works because the rules came first
"The AI keeps working while I sleep" is half right.
Leave AI alone without deciding anything, and you don't get a night shift. You get a stray-cat meeting.
Unattended operation needs humans to decide a few things first.
- What counts as passing
- What must never be changed
- How to protect sources and numbers
- What to do if two machines fix the same article at the same time
- Where to restart if something fails midway
- How many failures before an article gets quarantined
- What to check after publishing
Once those are settled, the human's job shifts from "doing the work every time" to "improving the rules of the equipment."
Using it to come up with four article ideas while out for a walk works the same way. Humans supply the ideas, the gut feeling that something is off, and the judgment. AI handles research support, structuring, drafting, translation and re-checking. The labor moves from "typing out words" to "deciding what to make."
The problem after mass production: AI slop and "copies of copies"
Being able to mass-produce will itself stop being rare.
Flooding the web with low-quality AI-generated content is commonly called AI slop. The problem isn't "it's bad because AI wrote it." The problem is that text with no evidence and no added value gets copied in huge quantities until nobody can tell which one is the original information.
Google also treats mass automated generation done only to win search rankings, rather than to help people, as a problem, and advises asking "who made it, how, and why." When AI or automation is used heavily, explaining how the content was made can sometimes help readers.
Research offers a different warning. A study published in Nature in 2024 showed "model collapse": when generated data is used recursively and indiscriminately to train the next generation of models, information at the edges of the original distribution gets lost.
But this is not a study proving that "publishing one AI article breaks the internet." It's research about recursive use of training data.
The lesson is still easy to see.
If you're going to make more copies of copies, don't erase the signposts back to the original data.
The next quality standard is the "information family register"
The W3C's PROV describes provenance, meaning what entities, activities and people produced a piece of data or other work, and says it can be used to judge quality, reliability and trustworthiness.
Academic publishing has an idea called Crossmark, too. With one button, readers can see whether a paper is still valid, whether there are corrections, retractions or updates, and what editorial information exists.
For digital assets such as images and video, C2PA has standardized Content Credentials, which prove where something came from and how it was edited.
They all have one thing in common.
Don't keep only the finished product. Keep "how it got here."
For a general-audience article, it works well to keep the following as its "information family register."
- First publication date
- Date the content was meaningfully updated
- Date the information was last verified
- The date the data is as of
- Primary and official sources
- Original data and data version
- The date each source was retrieved and checked
- Which claim in the text each source supports
- How AI was used
- Whether a human expert reviewed it
- How the automated QA and AI meaning-level QA were done
- Important change history
What matters here is not to overstate "who verified it."
In Schema.org, reviewedBy is the field for the Person or Organization that confirmed a web page's accuracy and completeness. If only AI checked it, write "AI quality check" separately. Don't summon an imaginary expert. Keep the magic circle closed.
Showing everything makes it harder to read, so use three layers
Say you get carried away with the idea and line up the SHA, DOI, retrieval timestamps, Git commit and QA rule version right under the title. What happens?
Reader: "Where's the article?"
The W3C's guidance on cognitive accessibility recommends clear words, short sentences, short blocks of text, clear headings, summaries and white space.
So make the register information three layers too.
Layer 1: a "short card" everyone sees right away
Near the title or the opening, put only what can be read in 5 seconds.
Evidence and update info for this article
Last verified: 2026-08-27
Data as of: 2026-08-27
Main primary/official sources: 9
How it was made: AI-assisted + source checking
Independent expert review: none
Key latest update: first version
[See sources, how it was made, and change history]
You can use the catchy name "information family register" if you like. But put a heading that says what it is in plain words first, such as "Evidence and update info for this article."
Layer 2: "details" that only interested readers open
Here you show how sources match the text, retrieval dates, research limits, AI use and change history.
Source S8
Type: research paper
Claim it supports: recursive use of generated data and model collapse
Original title: AI models collapse when trained on recursively generated data
Published in: Nature 631, 755–759 (2024)
Publication date: 2024-07-24
Checked on: 2026-08-27
Caution: not a study proving that publishing a single AI article breaks the internet
In the body text, give the plain explanation first. Keep the original title, DOI, authors and URL in the details for checking. Don't delete the hard information; change the order it's read in.
Layer 3: "structured data and internal evidence" that machines read
Humans don't need it, but machines need exact identifiers.
datePublisheddateModifiedcitationversion- An accurate
author reviewedBy, only when a person or organization really did review it- Internally: content SHA, source SHA, quality-rule version, the commit that made the change, and so on
Google's Article structured data also treats datePublished as the first publication date and dateModified as the date of change, separately. Schema.org has citation, version and reviewedBy.
Internal identifiers are handy, but there's no need to publish private repository names, personal accounts or secrets. Transparency is not the same as streaking.
One date isn't enough
If an article only says "Updated: today," nobody knows what "today" refers to.
Is it the day it was rebuilt? The day the text was rewritten? The day the sources were re-checked? The as-of date of the statistics?
At the very least, separate these.
| Item | Meaning |
|---|---|
| First publication date | The day it was first shown to readers |
| Content update date | The day the meaning, conclusion or evidence of the text changed |
| Last verified date | The day sources and current status were last re-checked |
| Data as-of date | The point in time that numbers or rules refer to |
| Source retrieval date | The day that source was checked |
Avoid a case where fixing only some CSS makes it look like "this article was updated today." Readers don't want a report that the server cheerfully rebuilt itself.
In 12 languages, translate the "explanation" but not the source's "ID card"
Sources are where multilingual work most often goes wrong.
The explanation in the body should be reworked so people in that language understand it naturally.
But the following are kept as the identifying information of the original.
- Original paper title
- Author names
- Journal name
- Year
- DOI
- PubMed ID
- URL
- Dataset name and version
For example, in the Japanese text, you first explain in Japanese what the study looked at, its main results and its key limitation. Then you put the English original title after that.
Chinese and Korean work the same way. Explanations for readers are in the local language; the source's ID card stays as in the original.
That way you get readability and traceability together.
What keeps its value in the AI era is mass production that can trace back to evidence
The fact that AI can write text will quickly become ordinary.
Then "we can make 100 articles a day" alone is hard to stand out with.
What makes the difference is the following.
- Each of the 100 articles has its evidence
- You can trace back to the sources
- You can tell when each source was checked
- You can tell what AI did
- You can see the revision history
- Sources don't break across languages
- You can re-verify when things get old
- When you find a mistake, you can apply the same fix rule to every article
Before AI, the more carefully you worked, the more labor cost grew.
After AI, the contest is whether you can build carefulness into the equipment as rules.
After the "factory that mass-produces articles" comes the factory that keeps a family register attached to its information while sending it out in volume.
Mass-producing text is something a photocopier can do.
Make articles that can walk around carrying their sources, dates and reasons for change. Only then does the AI article factory move toward being "information infrastructure" and away from being a "web pollution machine."
This article's information family register
- First version: 2026-08-27
- Last source check: 2026-08-27
- Data as of: 2026-08-27
- How it was made: points pulled from a conversation; public standards, official materials and research papers researched with AI assistance, then structured, written and localized into 12 languages
- AI use: research support, summarizing, structuring, writing, translation, consistency checks
- Independent human expert review: none
- Important change history: 2026-08-27 first version
- Sources: the shared Source Registry at the end ~
