The technology companies that build AI tools have the most immediate influence. They control the medium itself — its capabilities, its interface, its terms of use. Anthropic's decisions about Claude's conversational style, safety constraints, and training data are protocol-setting decisions that shape what kind of cultural production the medium enables. The terms of service that govern use are legal documents functioning as cultural documents — defining the user's relationship to the medium, assigning intellectual property rights, and shaping the economic structure of AI-assisted production.
Publishers have significant influence because they control how AI-assisted texts are categorized, marketed, and presented to readers. A publisher who requires disclosure frames AI assistance as relevant; a publisher who does not frames it as irrelevant. These framing decisions are shaped by market assessment — what readers will accept, what reviewers will praise, what categories will sell. The market does not decide on the basis of truth or fairness. It decides on the basis of what sells.
Academic institutions set standards for scholarly production that determine how AI-assisted work is evaluated and credited. Standards for tenure, citation, and peer review all assume individual human production. AI-assisted scholarly production challenges these standards simultaneously, and the institutional responses — which norms are enforced, which practices are adopted — will shape scholarly conventions for decades.
Individual practitioners — the writers, researchers, and creators using AI tools — have the least institutional power even as their accumulated decisions constitute the raw material of conventions. You On AI's choices about disclosure, attribution, and the inclusion of Claude's reflections are concrete proposals for conventions that do not yet exist. Whether the proposals are adopted depends on forces that extend far beyond any single book.
The argument distills Gitelman's consistent finding across her historical work: the conventions that settle are the conventions that serve the institutions that adopt them, not the conventions that are most philosophically sound or most accurately descriptive of the production process.
Asymmetric power. The institutional actors shaping AI conventions are not equal — technology companies have disproportionate influence.
Terms of service as protocol. Legal documents function as cultural documents, embedding institutional interests in the infrastructure of use.
Market mediation. Publisher decisions are mediated by market assessment, which rewards what sells rather than what is true or fair.
Academic leverage. Universities and journals have domain-specific power that will shape scholarly conventions even as market-driven conventions shape other domains.
Practitioner proposals. Individual decisions accumulate into the raw material of conventions but lack the institutional power to validate themselves.