To check a library collection for AI-generated books, combine acquisition records, publisher checks and a close reading of sample text. Automated detectors alone are not reliable, so the final decision should rest with a person who has read the book.
The BBC reports that the Tamil Nadu government has ordered an internal audit of library collections across the state. The order follows a scandal over claims that a new large library in India had acquired books generated by artificial intelligence. The source does not say how many titles are involved, how they were bought or who supplied them, and none of that is known here. The questions behind the story apply to any library, school or archive, though. This guide sets out a practical, cautious way to review a collection when there are concerns about machine-generated or low-quality synthetic content.
Gather what you need before starting
An audit like this depends on records more than on software. Before you begin, collect the following:
- Acquisition records. Purchase orders, invoices, supplier names, dates of acquisition and the budget lines used. These show how a book entered the collection.
- Catalogue data. ISBNs, publisher names, author names, edition details and subject classifications for each title under review.
- A written definition of the concern. Decide in advance whether you are looking for books that were entirely machine-generated, books with undisclosed AI assistance, books with factual errors, or books bought through irregular procurement. These are different problems and need different responses.
- Reviewers with subject knowledge. People who read the language the book is written in, and ideally know its subject well, are far better placed to judge quality than any tool.
- A neutral recording template. A spreadsheet or form where each title gets the same fields, so findings can be compared and challenged later.
Define the scope of the review
Do not try to examine every item at once. Start with a defined group, such as titles acquired in a particular period, from a particular supplier, or in bulk purchases. If the concern started with specific titles, begin with those and work outwards to related acquisitions. Write the scope down so that anyone reviewing the audit later can see what was and was not examined.
Pull and compare acquisition records
Many collection problems turn out to be procurement problems. Match each title under review against its purchase record. Look for patterns that call for a closer look: many titles from one previously unknown supplier, large numbers of books by the same author released in a short space of time, unusual pricing, or orders that skipped the usual selection process. None of these proves a book was machine-generated. They tell you which books to read first.
Check publisher and author details
Look up each publisher and author in independent sources such as national ISBN agencies, national library catalogues and established trade databases. Signs worth noting include:
- A publisher with no verifiable address, history or other catalogue.
- An author with an implausibly large output across unrelated subjects.
- No author biography, or a biography that cannot be confirmed anywhere.
- ISBNs that are missing, duplicated or registered to a different publisher.
Record what you found and what you could not confirm. A missing record is a reason to investigate further. On its own it does not prove anything.
Read sample passages closely
This is the most important step and cannot be automated away. Reviewers should read the opening chapters, a passage from the middle and the ending of each flagged title. Text produced with little human editing often shows recognisable features: repetitive phrasing, generic statements with no specific detail, invented or garbled references, factual errors stated with confidence, inconsistent names or dates, and chapters that repeat each other. Translations and regional-language books need reviewers who are fluent in that language, because awkward phrasing can come from poor translation rather than generation.
Use detection tools only as a secondary signal
AI text detectors are widely available, but their accuracy is disputed. They can label human writing as machine-generated, especially text by non-native writers or in formal styles, and they can miss edited synthetic text. Their performance in languages other than English is less well understood. If you use them, treat the result as one input among several, record which tool and version you used, and never remove a book on a detector score alone.
Record findings and decide on action
For each title, record the evidence from every step and a reviewer’s overall judgement, for example: no concern, quality concern, suspected undisclosed AI generation, or procurement irregularity. Possible actions range from keeping the book with an added catalogue note, to moving it to a reference-only section, to withdrawing it. Procurement findings should go to whoever handles finance or governance, kept separate from judgements about content. Keep the full record so that decisions can be reviewed and, where necessary, reversed.
Avoid the mistakes people actually make
- Trusting a single detector score. This is the most common error. It leads to false accusations against human authors and gives a false sense of certainty.
- Confusing low quality with AI generation. Poorly written, badly edited or hastily translated books existed long before generative AI. The response to a weak book may be the same, but the explanation matters when suppliers or authors are involved.
- Ignoring procurement. Focusing only on the text can miss how the books were selected and paid for, which is often where the real problem lies.
- Auditing without a written scope. Without one, results are hard to defend and easy to dispute.
- Naming authors publicly before review is complete. Unverified suspicions can cause reputational harm and may later prove wrong.
- Withdrawing books silently. Removing items without any record weakens trust in the collection and makes errors impossible to correct.
Know when this approach is the wrong choice
This manual, evidence-based method suits targeted reviews of a defined set of titles. It is the wrong choice in some situations:
- Very large collections with no specific trigger. Reading samples from every item is impractical. Improving acquisition controls for future purchases may be a better use of effort.
- Suspected fraud or corruption. If the concern is about payments, kickbacks or fake suppliers, it belongs with auditors, investigators or the relevant authorities, not a collection review.
- Disputes over legitimate AI-assisted works. Some books openly use AI tools in writing or translation. Whether to collect them is a policy question for the library’s governing body, and a quality audit cannot settle it.
- Languages without qualified reviewers. If nobody available can read the text fluently, judgements about its quality will be unreliable, and outside expertise should be sought.
Frequently asked questions
How can you tell if a book was written by AI?
No single test can confirm it. Reviewers usually look for repetitive phrasing, vague generic statements, invented references, confident factual errors and inconsistencies in names or dates. They also check whether the author and publisher can be verified independently. Taken together, these signals can support a reasonable judgement, but only a careful reading by someone who knows the language and subject can give a dependable assessment.
Are AI text detectors reliable for checking books?
They are not considered reliable enough to justify decisions on their own. Detectors can wrongly flag human writing, especially formal prose or text by non-native writers, and they can miss AI text that has been edited by a person. How well they work in languages other than English is less established. They are best used as one secondary signal and recorded alongside other evidence.
What did the Tamil Nadu government order?
According to the BBC, the Tamil Nadu government has ordered an internal audit of library collections across the state, following a scandal over claims that AI-generated books had been acquired. Details such as the number of titles involved, the suppliers concerned and the timetable for the audit are not given in the source material used for this article, and are not known here.
Should libraries remove AI-generated books?
It depends on the library’s collection policy. Some libraries may accept books that openly disclose AI assistance and meet quality standards, while others may exclude them. Books that are inaccurate, misleading or bought through irregular procurement are a separate concern. Any removal should follow a documented review so that the decision can be explained, checked and reversed if it turns out to be mistaken.
Sources and further reading
- BBC News technology coverage, which reports the Tamil Nadu audit order behind this article
- National library and ISBN agency guidance on cataloguing and verifying publisher records
- Academic and library-sector research on the accuracy and limits of AI text detection tools
- Professional library association guidance on collection development and deselection policies
Surfaced from the rss:bbc_tech signal “library AI-generated books audit”. AI-assisted draft, editorially reviewed.
