Research map / Large language models / AI and society

Position: Current Model Licensing Practices are Dragging Us into a Quagmire of Legal Noncompliance

ICML 2025 · Large language models › AI and society · approach: Benchmarking and reproducibility practice

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.

Position paper arguing current ML model licensing practices (OSS, free-content, and model-specific licenses) cause widespread legal noncompliance, supported by an analysis of Hugging Face model trees, and proposes ModelGo Licenses (MGLs).

Problem

Model publishers choose licenses ad hoc, often with OSS or free-content licenses ill-suited to ML models, and derivative reuse (implicit dependencies, license proliferation, conflicts) creates legal noncompliance risks that existing license tooling cannot detect.

Approach

Argues that license mismatch, license proliferation, and license conflict on Hugging Face create a "quagmire" of noncompliance, and proposes a reusable set of ML-specific model licenses (ModelGo Licenses) with license elements (BY, NC, ND, RAI, OS) and a companion model sheet.

Key results

• Analysis of about 151K Hugging Face models finds the license mismatch rate consistently over 75%, and 30% of models lack a declared license.
• Among direct derivatives of Llama-3.2-3B-Instruct, 53% retain the original license and 28.7% declare no license.
• About 2.2% of models with licenses and edges exhibit detected license conflicts, with conflict ratios reaching 56.6% among qualified models.

Measured effects reported in the paper

Surprising findings

• Copyleft-style model licenses (e.g., Llama2/3 clause 1.v) can be triggered by implicit dependencies and propagate to derivatives, creating conflicts with GPL-licensed projects.
• Violations of pure copyleft licenses among derivatives are low (about 0.24%), while copyleft-style terms show license changes settling at 17.7%.

Limitations

• Self-reported HF dependency labels may not conform to actual practice.
• Conflict detection covers only four easily identifiable conflict types, missing commercial-use and disclosure conflicts.
• License clarity scoring is subjective and tied to the model publishing scenario.

Experimental setup

Datasets: Hugging Face model cards and model trees (top 8K downloaded models; 151,302 models and 146,818 relationships)

Metrics: license clarity score license mismatch rate copyleft violation percentage license conflict ratio

Baselines: Apache-2.0 GPL-3.0 CC-BY-NC-4.0 Llama2 license Llama3 license OpenRAIL-M Gemma license OPT-175B license

Builds on: Creative Commons licenses OpenRAIL Model Cards Model Data Licenses

Closest papers by meaning

• Position: Standardization of Behavioral Use Clauses is Necessary for the Adoption of Responsible Licensing of AI (ICML 2024 · similarity 0.86)
• Position: Machine Learning Models Have a Supply Chain Problem (ICML 2025 · similarity 0.82)
• Position: Require Frontier AI Labs To Release Small "Analog" Models (NeurIPS 2025 · similarity 0.81)
• SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (ICLR 2024 · similarity 0.81)
• Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content? (AAAI 2026 · similarity 0.81)
• Modeling the Economic Impacts of AI Openness Regulation (NeurIPS 2025 · similarity 0.80)