The Report Was Checked by Nobody
The government officials said it would be very difficult to go back and check.
I want to begin there, at the end of the story, because it is a perfect little piece of office furniture. A report costing $3.48 million helped clear the way for Australia to keep children under sixteen off social media. The report contained citations that led nowhere, citations that led to the wrong place, names attached to papers those names had never touched, and research described in ways the research did not recognize.
The people who had paid for it were told the links had worked at the time.
Perhaps they did. Perhaps the dead came back briefly to answer the telephone. Perhaps a journal article published in June was already available in March, like a bottle of wine ordered by a man who has discovered time travel and still wants to expense lunch.
Or perhaps nobody checked.
There are two ways to read a footnote.
One way is to see it as a small iron nail holding a claim to the floor. You say something about children, surveillance, age checks, public policy, and you show the reader where the weight is coming from. The reader can bend down, pull at the nail, see whether the boards are solid. It is slow. It is boring. It is the sort of work people call administrative until the roof lands on their head.
The other way is to see a footnote as a tiny costume. Put enough of them under a paragraph and the paragraph looks qualified to enter the building. A DOI. A surname. A year in parentheses. The little black marks line up like men in cheap suits outside a courthouse. Nobody has to be inside the courthouse. Nobody has to know what happened there. The suits are doing the work.
The machine understands the second kind beautifully.
Ask it for sources and it has the eager panic of a junior clerk who has been told the vice president needs an answer before lunch. It does not know the article. It knows the shape of an article. It knows that serious things are written by people with two surnames, in journals with names like International Review of Something Important, in years that have already happened. It can make a citation the way a counterfeiter makes a twenty: close enough to pass from one tired hand to another.
This does not make the machine malicious. A broken adding machine is not a criminal mastermind either. It is just a bad thing to use for accounting.
The nasty part is the human being who sees the number, hears the bell, and hands it to the cashier anyway.
The company behind this thousand-page report now says ChatGPT was used only to rewrite paragraphs more succinctly, not to originate research or generate cited materials. The links, it says, were checked and verified by humans. That is a fine defense. It has the same clean smell as a restaurant telling you the soup was inspected by the man who dropped the cigarette ash in it.
A human verification does not become human merely because a human was in the room.
I have watched this kind of work get done. Somebody copies a number from one sheet to another. Somebody else initials it because the first somebody has a job title. A supervisor checks the box because the deadline is hungry. The report travels upstairs. By the time it reaches the man who signs it, everybody involved is protected by the existence of everybody else.
No villain. No scene. Just a chain of hands passing an empty envelope toward power.
The children in this story are almost beside the point, which is what makes it worse. Their names are on the policy. Their faces are on the posters. Adults say they are trying to protect them from the chemical factory of social media: the compulsive feeds, the predators, the little private humiliations that become permanent before a child has learned how to drive.
Maybe the ban is wise. Maybe it is foolish. Maybe age checks will protect a fifteen-year-old or build a new gate through which every citizen must crawl. These are real questions. They deserve real arguments, real evidence, real people who can be asked where they got a fact and who will feel their neck warm when the answer is bad.
Instead, we get a government report on the reliability of age-assurance technology whose own sources may be costume jewelry.
The joke has no bottom.
To establish whether young people can be safely identified online, a contractor produces claims that may not survive identification. To persuade a government that the technology has been tested, it offers a document that cannot pass its own test. Then, when asked to verify the foundation, officials explain that verification would be difficult.
Yes. That is why it was supposed to happen before the report became a bridge people were invited to walk across.
There is a phrase people use now: human in the loop. It sounds reassuring. You imagine a careful woman with a pencil seated beside the roaring machine, catching the bad answer before it reaches daylight.
But the human in the loop is often a man with forty tabs open, a supervisor asking whether he is finished, a contract measured by delivery, and a machine that produces ten plausible lies before he can find the first library database. He is not a guardian at the gate. He is a tired fellow jogging behind a truck that has already left the depot.
Then comes the grand escape hatch: the error was inadvertent. Of course it was. Most serious errors are. Nobody wakes up wanting to put a nonexistent article in a public report. They wake up late. They have too much work. They trust a tool that talks like the college roommate who always knew the answer until the exam paper came back. They see something that looks right and decide looking right is the same as being right.
That is not an excuse. It is the entire danger.
The old fraud at least had the decency to forge a signature with a pen. This new fraud can be accidental, scalable, polite, and wrapped in a workflow diagram. It can arrive with the cheerful language of assistance. It can make every participant feel less responsible because the machine did not mean anything by it.
Neither did the dead link.
But somewhere, a minister made a decision. Somewhere, a parent heard the word safety and assumed somebody had done the dull work. Somewhere, a kid will be asked to prove an age to a system justified in part by a document whose authors could not reliably prove the age of their own sources.
I am not asking for a ban on machines. I am asking for the old ugly standard: if your name is on the report, read the damned report. Follow the footnote. Open the link. Find the paper. Ask whether it says what you claim it says. If there are a thousand pages, hire enough people to do it or write fewer pages.
Do not tell the public that truth was checked by a human when the human was merely standing nearby while the machine dressed it for work.
The officials say they may engage with the contractor as appropriate.
I can see the meeting already. Water glasses. A careful apology. Somebody using the word process. Somebody else promising improvements. The report will remain on a server, enormous and official, its bad citations sleeping quietly in the dark.
Outside, a child will be asked to enter a birthday.
The machine will check it.