⚡ TechnologyBeta — synthesis improving

Software Licenses that prevent LLMs from training on open source?

2 views

3 modelsComplete

50%Consensus Score — 2 models mostly agreed, with some nuance. Read the breakdown.

🔴Semantic Entropy — Highly divergent · 2 distinct meanings across 2 responsesH=100%

50%

consensus

Contested topic

How much the models agreed on this answer

This topic is genuinely contested — read the individual model responses for different perspectives.

Models agree on

✓Traditional open-source licenses like MIT, Apache 2.0, GPL generally allow AI training.
✓OpenRAIL-S and Commons Clause are licenses designed to restrict AI training.
✓AGPL indirectly restricts AI training by requiring open-sourcing.
✓MIT, Apache 2.0, BSD, GPL explicitly permit AI training.

When it comes to software licenses that prevent Large Language Models (LLMs) from training on open-source code, the landscape is nuanced and evolving. While traditional open-source licenses like MIT, Apache 2.0, GPL, etc., generally allow such training, newer licenses and clauses have emerged to restrict or prohibit it. Here’s a breakdown of the most relevant options and considerations:

Licenses Explicitly Designed to Restrict AI Training

·
OpenRAIL Family
- ·OpenRAIL-S, OpenRAIL-M, OpenRAIL-D: These licenses, developed by the Mozilla-supported RAIL project, include use-based restrictions that can forbid AI training. They are not OSI-approved but are the most comprehensive legal options available for this purpose.
·
Commons Clause
- ·Used by licenses like Redis Source Available License (RSAL) and Elastic License 2.0 (ELv2), these explicitly prohibit AI training. They are often used to monetize open-core software while preventing misuse by cloud providers.
·
Server Side Public License (SSPL)
- ·Used by MongoDB, SSPL doesn’t explicitly ban AI training but requires open-sourcing the entire system if the software is used in a service, indirectly deterring such use.
·
Creative Commons with 'NoAI' Proposals
- ·While not standardized, some CC-BY-NC-ND licenses exclude AI training. Proposals for a formal 'NoAI' clause are under debate.

Licenses That Indirectly Deter AI Training

·
AGPL (Affero GPL)
- ·AGPL requires open-sourcing the entire system if the software is interacted with over a network, which can deter proprietary AI training.
·
Creative Commons Non-Commercial (CC-NC)
- ·CC-NC prohibits commercial use, which can indirectly restrict AI training, especially for commercial LLMs.

Licenses That Permit AI Training

·
MIT, Apache 2.0, BSD, GPL: These permissive licenses explicitly allow AI training.
·
CC0 / Public Domain: These impose no restrictions at all.

Practical Advice

·
For Developers: Stick to permissive licenses like MIT or Apache 2.0 for your own models to avoid legal risks. Avoid Commons Clause, SSPL, and AGPL-licensed data unless prepared for open-sourcing. Prefer CC0 or Public Domain datasets for commercial use.
·
For Projects Aiming to Block Training: OpenRAIL-S is the most defensible choice, even though it sacrifices the OSI stamp. Custom clauses attached to existing licenses (e.g., MIT + 'No Training' clause) are another option, though enforceability remains uncertain.

Conclusion

While no widely adopted, OSI-compliant license currently prevents LLM training, OpenRAIL-S and Commons Clause offer the most robust options for those looking to restrict AI use. Legislation or industry norms may ultimately provide clearer solutions, but for now, careful licensing is the most practical approach.