"AI firms acquire and destroy physical books for LLM training data"
Large Language Model (LLM)
AI model that learns statistical text patterns from massive corpora to generate human-like output
Destructive Scanning
digitisation method that permanently dismantles a physical book by cutting its spine for high-speed scanning
The development of Large Language Models (LLMs) — the engine behind generative AI systems — depends fundamentally on vast quantities of high-quality, human-written text. As the NCERT Informatics Practices textbook notes, AI endeavours to simulate human intelligence by creating a knowledge base from which machines can learn statistical patterns; for LLMs, this knowledge base is built from text corpora of extraordinary scale. Early LLM training relied heavily on text scraped from the public internet — Common Crawl datasets, Wikipedia, and open-access repositories — because of their sheer volume and accessibility.
However, the internet-scraping model has encountered two compounding problems. First, as AI-generated content has proliferated online, the quality of web-scraped data has degraded — models trained on AI-generated text risk amplifying errors and hallucinations in a feedback loop. Second, large-scale web scraping has attracted sustained legal challenge under copyright law, with courts in multiple jurisdictions examining whether scraping constitutes infringement of authors' and publishers' rights. These twin pressures have pushed frontier AI companies toward an alternative source of clean, human-authored, long-form text: physical books.
The legal backdrop for book digitisation was set by the Google Books litigation of the 2000s, in which Google's project to digitise millions of library volumes sparked years of copyright litigation. Courts ultimately held that certain digitisation for search-indexing purposes could qualify as fair use, but the doctrine was narrowly construed and the precedent has not been straightforwardly extended to AI training — a purpose Google Books did not contemplate.
Large Language Models learn by processing enormous text corpora and optimising mathematical functions to minimise prediction error — essentially learning to predict the next word in a sequence with high accuracy across billions of examples. The quality, diversity, and cleanliness of training data directly determines model capability. Books are particularly valued because they contain long-form, edited, coherent prose across a vast range of subjects — qualities that short-form web content rarely offers. Older books published before the generative AI era are especially prized because they are guaranteed to be free of AI-generated content, making them what the industry calls 'clean' training data.
The process of converting physical books into machine-readable training data involves two steps. The first is digitisation: scanning physical pages to produce high-resolution images. The second is Optical Character Recognition (OCR), which converts scanned images into machine-readable text that models can ingest. Non-destructive scanning — photographing pages without dismantling the book — preserves the physical copy but is slower and can produce lower-quality images for brittle or tightly-bound volumes. …
Consider the following statements with respect to the AI Impact Summit, 2026 held in New Delhi:
1.The Summit's intellectual framework was based on three foundational Sutras: People, Planning, and Progress.
2.The Preamble of the Summit stresses Democratising AI Resources, which acknowledges the Charter for Democratic Diffusion of AI as a binding framework to support locally relevant innovation and strengthen resilient AI ecosystems while respecting national laws.
3.The New Delhi Declaration on AI Impact was structured around seven Chakras (Pillars), which included Access for Social Empowerment, AI for Science, and Secure and Trusted AI.
Which of the statements given above is/are correct?
Which of the following statements with regard to Large Language Models (LLMs) used in machine learning is/are correct?
1.LLMs assign probabilities to the next possible word and then pick the one with the highest probability.
2.LLMs process data through mathematical optimization to minimise prediction errors.
3.LLMs produce unbiased outputs.
Select the answer using the code given below:
With the present state of development, Artificial Intelligence can effectively do which of the following?
1.Bring down electricity consumption in industrial units
2.Create meaningful short stories and songs
3.Disease diagnosis
4.Text-to-Speech Conversion
5.Wireless transmission of electrical energy
Select the correct answer using the code given below:
Consider the following statements about Large Language Models (LLMs) and their training data requirements:
1.LLMs learn by assigning probabilities to sequences of words and minimising prediction error through mathematical optimisation.
2.AI-generated text proliferating online is considered a high-quality source of training data for frontier LLMs.
3.Physical books published before the generative AI era are valued as training data because they are free of AI-generated content.
Which of the statements given above is/are correct?
Which of the following best describes 'destructive scanning' as used by AI companies for book digitisation?
With reference to copyright law and AI training data, consider the following:
Assertion (A): Purchasing a physical copy of a book lawfully does not grant the purchaser the right to scan and reproduce its contents for AI training.
Reason (R): The first-sale doctrine permits resale of a lawfully purchased physical copy but does not transfer the copyright in the underlying text.
Which of the following is correct?
Which of the following statements about the Internet Archive's digitisation practices is correct?
Want the full note, Prelims Facts & MCQ answers?
Create a free account — takes less than a minute.
Create free account