Research map / Large language models / AI and society
Position: Near to Mid-term Risks and Opportunities of Open-Source Generative AI
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.
A position paper arguing for responsible open-sourcing of near- and mid-term generative AI models, introducing a five-level openness taxonomy applied to 40 LLMs and weighing risks and opportunities.
Problem
Calls for tighter regulation of generative AI may stifle open-source development; the paper asks how the openness of gen AI models trades off risks and benefits in the near to mid-term.
Approach
Defines a point-based openness taxonomy (C1-C5 for code, D1-D5 for data) based on license permissions, applies it to the pipeline components (training, evaluation, deployment) of 40 high-impact LLMs, and uses the findings to argue for responsible open sourcing with specific recommendations.
Key results
- Openness taxonomy applied to 40 high-impact LLMs (2019-2023) shows a balance between open and closed deployed components (inference code and weights).
- A notable skew toward closed source exists in training data (fine-tuning and alignment) and in safety evaluation code and data.
- Licenses are scored on five permissions; e.g., LLaMA-2 license scores 3 (moderately restrictive, C3/D3), MIT/Apache 2.0 score 5 (restriction free, C5/D5).
Surprising findings
- Closed models are not clearly safer: closed models can be fine-tuned to be as unsafe as open ones and are vulnerable to jailbreaks.
- Mistral's 70B model was reportedly leaked by an early customer, undermining the claim that closed access can always be restricted.
Limitations
- Evidence is partial and incomplete given rapid AI change and limited insight into industrial applications, as the authors acknowledge.
- Taxonomy classifies component openness rather than reproducibility; GPT-4 and GLM-130B, for example, are classified C1 despite differing documentation.
- Risk assessment rests on fundamental assumptions that may be violated by disruptive changes.
Experimental setup
Datasets: The Pile LAION RedPajama
Metrics: openness score (0-5 license points) openness level per pipeline component
Baselines: GPT-4 Claude Gemini PaLM-2
Builds on: DPO RLHF LLaMA-2
Closest papers by meaning
- Position: A Safe Harbor for AI Evaluation and Red Teaming (ICML 2024 · similarity 0.86)
- Position: Generative AI Regulation Can Learn from Social Media Regulation (ICML 2025 · similarity 0.84)
- Modeling the Economic Impacts of AI Openness Regulation (NeurIPS 2025 · similarity 0.84)
- Position: Require Frontier AI Labs To Release Small "Analog" Models (NeurIPS 2025 · similarity 0.84)
- On the Consideration of AI Openness: Can Good Intent Be Abused? (AAAI 2025 · similarity 0.84)
- Position: Open-Endedness is Essential for Artificial Superhuman Intelligence (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.