What may the studio do with AI-written content?

The question

The studio uses several AI services, and trains small models of its own. For each provider, which of four uses do the published terms allow: shipping output as a commercial asset, training on it, evaluating with it, and running code it wrote? The maintainer’s rule applies: a use is allowed unless the provider’s contract or licence restricts it, and the search for restrictions is documented.

What we found

• Google (consumer terms, 2026-07-30) bars using AI-generated content “to develop machine learning models or related AI technology”. Read by its anti-distillation purpose, that covers training data and evaluation items, not ordinary code no model trains on. Outside that, Google disclaims ownership and restricts nothing that applies, so commercial use is allowed. Flow Music’s terms grant it outright.
• xAI leaves output with you (“you retain your ownership rights”) and requires “Written with Grok” attribution. Its Acceptable Use Policy bars “distilling model data or Outputs”, so no training on Grok output.
• Anthropic (Help Center 12326764, 2026-03-16) allows Claude’s outputs to train non-competing “specialized classifiers and tools”. It bars general chatbots, open-ended text generation, and outputs as training targets.
• Moonshot’s Kimi app is non-commercial only. Through Ollama Cloud, Ollama’s terms leave content with you, and the Kimi K3 and K2.6 licences are MIT-style.
• Open weights: free unless the licence restricts outputs. OpenMDW-1.1, Muse-Glimmer’s synthetic-data intended use and DeepSeek-R1’s distillation grant say so outright. Plain Apache-2.0, MIT and NVIDIA’s Open Model License are silent, so they’re free too. Muse-Spark and Llama 3.3 add conditions.

What it changed

Stage 1’s 810 training rows and 210 apply items are being regenerated by licence-proven open models. Generator code written by Antigravity or Grok is being rewritten. Interlock’s Grok and Kimi code stays, with credit.

Caveats

• This is a reading of published terms, not legal advice. Terms change, so every clause carries its page date.
• The student verifier counts as a “specialized tool” by the maintainer’s judgement. An AI support agent at Anthropic agreed, but that isn’t a human review.
• A use is barred only by a clause that can be named (2026-10-17). A two-tier “strict” version was withdrawn the same day.
• The rule was first worded so that silence counted as a bar. It was reworded the same day, and these verdicts follow the reworded rule.

Receipts

mcp-tool-shop-org/rnd:

• experiments/generated-content-terms/README.md: each provider’s clauses, re-read with URLs and dates;
• experiments/stage1-regeneration/PLAN.md §1: the open-weight licences;
• experiments/stage1-regeneration/PROVENANCE.md: the inventory, and the code and terms readings.