MIT isn’t “weak” because it allows LLM training; it’s weak because it puts zero obligations on the recipient.

Blocking “LLM training” in a license feels satisfying, but I’ve run into three practical issues while benchmarking models for clients:

1. Auditability — You can grep for GPL strings; you can’t grep a trillion-token corpus to prove your repo wasn’t in it. Enforcement ends up resting on whistle-blowers, not license text.

2. Community hard-forks — “No-AI” clauses split the ecosystem. Half the modern Python stack depends on MIT/BSD; if even 5 % flips to an LLM-ban variant, reproducible builds become a nightmare.

3. Misaligned incentives — Training is no longer the expensive part. At today’s prices a single 70 B checkpoint costs about \$60 k to fine-tune, but running inference at scale can exceed that each day. A license that focuses on training ignores the bigger capture point.

A model company that actually wants to give back can do so via attribution, upstream fixes, and funding small maintainers (things AGPL/SSPL rarely compel). Until we can fingerprint data provenance, social pressure—or carrot contracts like RAIL terms—may move the needle more than another GPL fork.

Happy to be proven wrong; I’d love to see a case where a “no-LLM” clause was enforced and led to meaningful contributions rather than a silent ignore.