Fact one: companies increasingly attribute output to AI. It reduces perceived labor cost and it impresses investors, so the number is advertised and it climbs.
Fact two: AI cannot hold copyright. Human authorship is, in the courts' word, a "bedrock requirement." If content is entirely AI-generated it cannot be protected; "mere provision of prompts" is not enough human authorship.
The same output cannot be "AI-generated" on an earnings call and "human-authored" on a copyright registration. Something has to give — and the paper shows exactly what gives, and to whose benefit.
The corporation captures the value without performing either function: the AI did the work but can't be credited or held accountable; the human is accountable but reduced to a "review function"; the company that did neither owns the result.
Here is the sharp edge. The higher the AI-attribution percentage, the more IP flows to the corporation. Because raising the AI share shrinks the individual employee's claim to authorship credit, while the "reviewed and accepted" step — Pichai's exact phrase — is engineered to be the minimum viable human involvement needed to satisfy the copyright requirement. The employee who clicks accept becomes the legal author by default, their creative contribution minimized, and the employment agreement transfers that thin authorship upward.
A structural aside the paper adds: the dominant architecture is Mixture-of-Experts (DeepSeek V3, Grok-1, GPT-4/5, Mixtral, Llama 4, Qwen3, Gemini, Kimi K2), which fragments the model into specialized sub-networks that activate per-token. So the "AI" being credited with 25–30% isn't a single coherent agent — it's "a committee of experts that form and dissolve on a per-token basis." There is no persistent author even within the model's own architecture, which further undermines any future argument for AI authorship.
The fix is not to deny AI involvement. It's to document the co-authorship chain from the beginning, establishing human authorship with receipts rather than by legal omission. The TriPod model:
This is the point where the argument becomes a build: the operational spec for exactly this is the companion sphere — the attribution standard. The paper says why; the standard says how.
The closing line is the whole thesis compressed: "The attribution structure IS the capture structure. Document accordingly."