Blind Spot in AI Regulation: Who Will Spend Weeks Verifying Documents Written by AI in Seconds?

Blind Spot in AI Regulation: Who Will Spend Weeks Verifying Documents Written by AI in Seconds?

Out of 194 citations, 50 supported the claims

The conference materials state that "70% of domestic companies have implemented it." When you click on the footnote, a survey report indeed opens. However, upon reading the original question, it turns out that they were asking, "Are you considering implementation?" The figure "70%" and the name of the survey are correct. Still, if this sentence is used as is in the conference, companies in the consideration stage would be mistakenly counted as having implemented it.

Verifying the sources of AI-generated text involves reading even such subtle differences in verbs. The individuals who actually undertook this task were Amartya Kumar, who founded the Indian research institute "Oryx Centre for Strategic Futures," and his co-founder Rhea Narwa.

According to an essay by Kumar published in the Australian opinion magazine 'On Line Opinion' on November 18, 2026, the two spent several weeks checking each citation from 30 AI-generated research documents. They compared the claims written in the documents with the sources provided right next to them.


Results of tracing the citationsNumber of cases
Could not reach the linked materials due to page moves, etc.95 cases
The source supported the claims in the text50 cases
The source existed, but the content contradicted the claims17 cases
Could not be verified31 cases
No link was provided1 case
Total194 cases

The proportion of cases where the two could verify the sources was about 26%. The remaining approximately 74% included broken links and unverifiable cases. It is possible that the original description was correct, and the page had merely moved. The figures are based on the authors' investigation of 30 documents and do not represent the overall accuracy rate of generative AI.


The hardest to notice are the links that open

The 17 cases that warrant a pause in the list are those where the links opened. If a link is broken, the reader immediately realizes it is "unverified." On the other hand, if a page from an existing research institution opens and similar words to the text appear, one might be tempted to conclude the verification there. In the initial example, it would be the moment when "70%" is found through a page search.

Errors in citations occur at different stages. Does the paper exist at all? Are the author and the journal correct? And can the conclusion in front of you be drawn from that paper? Even if the first two are passed, one can stumble at the last.

A study published in the academic journal 'Scientific Reports' in 2023 focused on the existence of literature and bibliographic information. Researchers had ChatGPT write literature reviews on 42 themes and checked a total of 636 references, finding that fictitious references were 55% with GPT-3.5 and 18% with GPT-4. Furthermore, with GPT-4,even among existing references, there were significant bibliographic errors in 24%. Although the model and task differ from Kumar's investigation, it shows that the presence of genuine papers cannot be the basis for trusting the entire list of references.


A librarian on social media talked about "pages spanning two papers"

How does this issue reach the library counter? In June 2025, prior to the publication of the original article, the library community on the overseas forum Reddit shared experiences of searching for books and papers recommended by AI.

 

A paper inquired about by a professor had the name of an existing researcher and journal, and the volume and publication period matched. However, when opening the specified page, it contained the end of one paper and the beginning of another. When the professor was asked about the source, they replied that they had asked ChatGPT again and received a response that it existed. The difference between asking again and getting the same answer and actually opening the journal to verify it is reflected in these few pages.

In another thread, a user shared an experience of showing a list of an author's works to a user while verifying the book title together. The title suggested by AI seemed to be a combination of parts of two existing works by the author and words used by the user during the search. The user was convinced after reviewing the list of works together.

These are personal anecdotes, but they reveal burdens not visible from the number of investigations. It takes a moment to create a plausible book title, but explaining why it can't be found involves checking the list of works, opening the actual journal, and sometimes turning pages with the other person.


In Japanese conference materials, which sentence should be verified first

Kumar raised a very practical question while discussing the systems surrounding AI in various countries. Even if safety policies are created and reporting destinations are decided when problems arise, who will open the footnotes in the materials distributed today?

In a Japanese workplace, first considering "If this sentence is incorrect, will the conclusion of the meeting change?" can make the order of verification clearer. If a proposal is made to hasten investment based on the initial "70%," the place the person in charge should open is not the cover of the report but the question text and aggregation table that produced that number. Is it a sample that can be described as "domestic companies"? Are "implemented" and "considering" separate options? Once you read that far, you can decide on the sentence that can remain in the materials.

The same can be done in school reports. Before copying the title of a paper into the reference section, write in the margin, "Which page and description of the paper supports this sentence of mine?" If the relevant part cannot be found, narrow the claim to the extent that it can be stated in the original text. If opening the source makes the text a bit dull, the value of verification lies in that correction.


Sources