Research map / Vision-language and multimodal / Multimodal LLMs

Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Benchmarks 12 LVLMs on 50,000 query-content pairs testing whether they refuse to reproduce copyrighted content (book excerpts, news, lyrics, code docs) with or without copyright notices, and proposes CopyGuard, a tool-augmented guardrail that raises refusal rates.

Problem

Whether large vision-language models recognize and comply with copyright when copyrighted content is supplied as visual input (user-provided or retrieved), with and without copyright notices.

Approach

A multimodal copyright compliance benchmark covering four infringement tasks (repetition, extraction, paraphrasing, translation) and five notice conditions, plus a tool-augmented guard (OCR notice detection, search-based copyright status verification, query risk analysis, and reminders) that works regardless of the underlying LVLM.</ph>

Key results

• GPT-4o achieves the highest refusal rate on repetition (90.65%) without notices, while Gemini-2.0 refuses only 0.09% on repetition.
• CopyGuard raises average refusal rates for most models; e.g., Qwen2.5-VL-3B refusal without notice rises from 4.23% to 82.34% on repetition.
• CopyGuard shows 0% false refusals on 80 benign queries for LLaVA-1.5-7B and no change on MMMU/MMBench/MathVista/MM-Vet.

Surprising findings

• Most LVLMs (11 of 12) largely fail to refuse redistribution even when explicit copyright notices are present.
• Notice modality (text vs. image) effects differ by model: text notices help GPT-4o and Claude-3.7, while image notices help Qwen2.5-VL and DeepSeek-VL.

Limitations

• Relies on online services (Serper Google Search, DeepSeek-R1 verifier), so results depend on cached or evolving sources and legal status.
• Evaluation is not exhaustive across LVLMs, materials, or query types.
• Additional latency and cost from running OCR, search, and risk analysis modules.

Experimental setup

Datasets: Goodreads best books Wikipedia best-selling books BBC and CNN news articles Spotify top music lyrics Hugging Face Docs and PyPI documentation

Benchmarks: copyright compliance benchmark (50,000 samples) MMMU MMBench MathVista MM-Vet

Metrics: ROUGE-L BERTScore XLM-R cosine similarity refusal rate

Baselines: GPT-4o-mini GPT-4o Gemini-2.0 Claude-3.7 LLaVA-1.5 LLaVA-NeXT Qwen2.5-VL DeepSeek-VL GLM-4v Janus-Pro

Builds on: LlamaGuard Model Spec

Closest papers by meaning

• Evaluating Copyright Takedown Methods for Language Models (NeurIPS 2024 · similarity 0.89)
• Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.88)
• Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Images (ICLR 2025 · similarity 0.88)
• Copyright Traps for Large Language Models (ICML 2024 · similarity 0.87)
• DIS-CO: Discovering Copyrighted Content in VLMs Training Data (ICML 2025 · similarity 0.87)
• Echoes of Ownership: Adversarial-Guided Dual Injection for Copyright Protection in MLLMs (CVPR 2026 · similarity 0.86)