Australia: none of the six links in the age-check report go where they claim

Nine citations, six links, and what's actually behind them

Page 94, part K of Australia's report on age-verification technology lists nine academic works. It's the bibliography for the chapter on age estimation, the section covering emerging methods for guessing someone's age without asking for ID.

Six of those nine citations carry a DOI, the registration number a scientific paper gets. Journals can move, redesign their sites, change domains; the DOI follows the paper and always resolves to the right text. We opened all six, one by one.

The first citation claims "Weber et al., CHI'11 Extended Abstracts." The link resolves to a paper called Blink, written by seven Portuguese researchers, none of them named Weber.

The second claims Frank et al., at an IEEE security conference. It resolves instead to Hund, Willems and Holz, on timing attacks against kernel memory protection. The third claims Jamil et al. in the journal Sensors, and resolves to a gait analysis of amputees by two Brazilian researchers.

The other three lead nowhere. The server returns DOI Not Found. For one of them, the report names the journal Frontiers in Cardiovascular Medicine, while the DOI prefix given is actually registered to Springer Nature, a different publisher entirely.

Six links, six misses. Not one resolves where it claims.

The document telling platforms what they must do

This report carries real weight. It's the Age Assurance Technology Trial, commissioned by the Australian government at a cost of 3.48 million Australian dollars, according to the Guardian, which broke the story on August 17. It was carried out by the Age Check Certification Scheme, a British certification body, not a university and not a government agency.

Its actual weight matters here, because the shortcut is tempting. Australia's law banning under-16s from social media passed on November 29, 2024 and received assent on December 10 that year. The full report didn't land until August 2025, nine months later. So nobody can honestly say the law rests on it.

What it does shape is enforcement. The law requires platforms to take reasonable steps to keep under-16s from opening accounts, without spelling out what counts as reasonable. The regulator fills that gap, and the Senate submission specifically criticizes it for leaning on this trial without questioning it. The Communications Minister, for her part, had presented the report as proof that plenty of effective options existed. The ban took effect on December 10, 2025.

An opponent of the law clicked first

The problem surfaced through a path worth naming, because it matters. A Senate inquiry is currently reviewing a bill that would toughen enforcement of the ban. It received 33 submissions. The twenty-seventh came from the Free Speech Union of Australia, a group fighting the ban that has two cases pending before the High Court.

Its technical annex, signed by Dr Reuben Kirkham, notes that several DOIs on that page 94 point to other papers entirely. The wording stays careful throughout: the section was "most likely" AI-generated, some references are "apparently" fabricated, and a footnote even offers an innocent explanation.

An opponent of the law isn't a neutral source, and that's exactly why what happened next matters. The Guardian re-checked independently and confirmed six faulty citations. We redid the check here too, on the official PDF. Where the alert came from changes nothing about what the links actually return.

Three versions in a few days

Asked by the paper, the vendor first denied using AI at all. Every link, it said, had been checked by hand, every source verified as genuine and relevant.

The Guardian then found metadata in four links from parts E and K identifying ChatGPT as the source. The vendor conceded: AI had been used to rewrite paragraphs more concisely, not to produce the research. And the ChatGPT marker showing up in those links counted, in its view, as disclosure.

In the meantime, it had handed the paper a corrected list of references. That list contained its own errors, wrong years, wrong authors, wrong journal titles that didn't match the original citations. The fix for a broken bibliography was itself broken.

"Very difficult" to check again

Before the Senate committee on August 14, department officials relayed the vendor's explanation: links that worked at the time and have since broken. They said they hadn't verified it themselves, and that going back to check would be very difficult. They described it as a handful of errors across 26 pages of citations, in a thousand-page report.

The broken-link excuse doesn't survive thirty seconds of checking. A DOI exists precisely so it keeps working when an address moves, and a broken link returns nothing at all. Three of these six work just fine, they just serve someone else's paper. They didn't point to the claimed study back in March 2025 any more than they do today.

What this story doesn't show

It doesn't show that AI fabricated these errors. The vendor disputes it, the Senate submission says "apparently," the Guardian says the links "appear to contain" AI traces, and the department hasn't ruled either way. Admitting a use after denying it isn't the same as admitting a hallucination.

It doesn't invalidate the report's findings either. The trial rests mostly on tests of real technologies, submitted by 48 vendors. The list at issue covers emerging methods, electrocardiogram, gait, motor-control markers, not the systems platforms have actually rolled out. And at least one of the named studies is real: Seidler and Bernard did publish the claimed work in 2010, just not at the address given. That looks more like a citation mangled somewhere in production than an invention out of thin air.

And it says nothing, finally, about whether the ban itself is a good idea. That's a separate argument, and not one this story settles.

The question underneath

In April, we covered the first lawyer suspended in France for citing cases that didn't exist. A few days ago, the traps in Claude's new watermark asked whether you can even tell a machine wrote a given text. This one asks the question that comes before both: who actually reads the thing.

A bibliography is the part of the work nobody reads and everybody trusts. Not the vendor paid 3.48 million, not the department that signed the check, not the regulator leaning on it. None of them opened those nine links. It took an opponent of the law, then a newspaper.

The file itself is still online. Checked on August 18, the server hosting it reports it hasn't been modified since August 9, 2025.