Bridging the copyright gap: Can large vision-language models recognize and respect copyrighted content?

The researchers construct a copyright compliance benchmark containing 50,000 multimodal query-content pairs, and evaluation finds that even the most advanced closed-source LVLMs have notable deficiencies in recognizing and respecting copyrighted content, even when the content is accompanied by copyright notices. The benchmark covers book excerpts, news reports, lyrics, and code documentation presented as visual inputs, divided into two scenarios with and without copyright notices, with the former covering four types of copyright notices. The authors also propose a tool-augmented defense framework, which they state reduces infringement risk in all scenarios.