Generative AI and intellectual property: a systematic literature review of possible issues and mitigations

TL;DR

This systematic literature review maps copyright issues and mitigations across the generative AI input and output pipeline under EU law, based on 24 high-quality studies published between 2019 and early 2025. It finds recurring concerns around opaque training data, web scraping and licence misalignment, memorisation and style mimicry, and reviews technical and legal mitigations including watermarking, TDM opt-outs, attribution, synthetic data, and provenance tracking.

Key findings

• The final SLR pool contained 24 sources, with input and output stages covered in a balanced way (52.8% input, 47.2% output). Input-stage mitigations explicitly mentioned included watermarking (38.2%), TDM opt-out (23.8%), protective perturbations (19.0%), synthetic data (9.5%), and blockchain (9.5%).
• Output-stage issues explicitly identified included transparency (33.3%), memorisation (25.0%), stylistic mimicry (20.8%), adversarial attacks (12.5%), RAG (4.2%), and unaware GenAI users (4.2%). Output mitigations included fingerprinting for authorship attribution (25.0%), attribution to originating training samples (25.0%), watermarking (22.4%), generation filters (10.8%), blockchain (5.6%), distilled models (2.8%), prompt rewriting (2.8%), C2PA (2.8%), and machine unlearning (2.8%).
• Specific reported results include: TPE truncation efficacy reached 99% in mathematics but dropped to 22% in coding; 93% of surveyed artists considered Glaze effective against generative style mimicry; approximately 80 million images were excluded from Stable Diffusion 3 training data; PREGen reduced direct-anchoring generation of copyrighted characters from 56.0–82.6% baseline to 0.6–6.6%, and indirect-anchoring from 26.6–27.4% to 0–2.0%.
• Other reported figures: fuzzy-fingerprint authorship attribution achieved 67.4% accuracy on 50 authors; Sakurai et al. achieved 88.4% accuracy on eight Japanese authors, with distillation reducing inference time from 20.33 to 3.38 s but decreasing accuracy by about 10%; DistilBERT has 40% fewer parameters and retains 97% of original performance; Lee et al. estimated nearly 1% of language model outputs are verbatim copies of training data.

Methods

• The authors conducted a systematic literature review following steps inspired by Garousi, Felderer & Mäntylä (2019) and quality-assessment guidelines from Kitchenham & Charters (2007), focusing on EU copyright law and the GenAI pipeline.
• Search strings covering GenAI and copyright were applied to ACM, IEEE, Springer, ScienceDirect, and Google Scholar, with a publication-year filter for 2019 to early 2025. Search bounding yielded 454 initial sources; after duplicate removal and inclusion/exclusion criteria, 44 sources remained.
• Two authors independently assessed source quality using tailored questionnaires; scores were normalised to 0–5, the average threshold was 3.48/5, and sources above this threshold were retained, resulting in 24 high-quality studies. Relevant information was extracted in structured form and organised around research questions for input and output stages.

Limitations

• The paper states that the topic’s fragmented nature made it harder to design an efficient and comprehensive search string; some recent rights-information-management initiatives such as JPEG Trust and Liccium’s Trust Engine fell outside its scope. It also states that audio and video remain underexplored, that the EU-only focus may lack global generalisability, and that the final corpus of 24 sources was smaller than that of a comparable secondary study.
• The authors state that legally charged keywords such as “copyright” likely biased ranking algorithms toward law-centric sources, limiting visibility of computer science literature, and that bounding strategies in Google Search and Springer Nature further restricted access to technical contributions. They also state that the review prioritised breadth over depth.
• Evident from the text: the review does not report new primary experiments; effectiveness figures are drawn from cited studies, and several mitigation results are described as limited by real-world transformations, model-specific vulnerabilities, or scalability constraints.