LLMs and Memorization: On Quality and Specificity of Copyright Compliance

LLMs vary dramatically in copyright compliance: The study found "huge differences in copyright compliance, specificity, and appropriate refusal among popular LLMs," with models like GPT-4, GPT-3.5, Alpaca, and Luminous performing significantly better than others like OpenGPT-X at avoiding reproduction of copyrighted material.

Current memorization poses legal risks: LLMs readily reproduce copyrighted content from training data, which the researchers note "may violate existing copyright laws as well as the European AI Act," making this a pressing legal and regulatory concern for AI developers and deployers.

How models refuse matters legally: The paper provides "a first legal assessment" of different refusal behaviors, finding that models don't simply refuse equally—some hallucinate, others refuse outright—and these different responses have different legal implications under copyright frameworks like the German Copyright Service Provider Act.

A practical legal threshold exists: The researchers applied a 160-character threshold (borrowed from German copyright law) to identify potential infringements, providing a concrete, legally-grounded standard for what constitutes problematic memorization rather than relying on arbitrary measures.

Real-world testing reveals the problem: Unlike previous studies, this analysis evaluated instruction-finetuned models "in a realistic end-user scenario," suggesting that copyright violations may be more prevalent in how people actually use these systems than laboratory conditions might suggest.