Studies Show AIs Can Reproduce Novels from Training Data

Researchers have found that top artificial intelligence models can generate near-exact passages from bestselling novels, raising serious concerns over copyright and the way these systems handle protected works.

A growing body of studies has demonstrated that large language models (LLMs)—the technology behind generative AI systems developed by organisations such as OpenAI, Google, Meta, Anthropic, and xAI—can reproduce sizeable sections of books, word-for-word, when prompted in specific ways. This behaviour is known as 'memorization,' where the AI, rather than paraphrasing or summarising, retrieves and outputs almost exact text from its training data.

LLMs are neural networks trained on vast quantities of text data. While their designers claim that they merely learn patterns and generate new text rather than storing and reproducing data, these findings challenge that assertion. According to legal and AI experts cited by the Financial Times, the models’ ability to recall copyrighted material undermines a central argument used by AI companies in court: that their systems do not store or distribute copyrighted works but use them to 'learn.'

This debate has far-reaching implications. Dozens of copyright lawsuits are under way across the globe, brought by authors and rights holders who argue that their works have been unlawfully used to train generative AI systems. The demonstration that AIs can generate near-verbatim copies may further bolster arguments from complainants and could influence ongoing and future regulatory decisions regarding AI and copyright law.

The extent and frequency of such memorisation events remain under study, but the new evidence suggests that current safeguards may be insufficient. Existing AI industry practices typically include 'deduplication'—removing duplicate entries from training data—but are not always effective in preventing memorization, particularly when the same text appears repeatedly online.

Regulatory bodies in the United States, Europe, and elsewhere are scrutinising the practices of AI companies, including how they source and use data. In Europe, the AI Act and recent copyright discussions at the EU level reflect growing concern about the intersection of AI advances and intellectual property rights.

The findings are likely to be closely watched by policymakers and the technology sector, as they may prompt additional calls for transparency, model auditing, and the establishment of clearer standards for training data usage. The issue of whether AI models can or should train on copyrighted works without explicit permission remains unsettled and at the centre of dispute for AI’s future.

Source: arstechnica.com.