The most consequential AI policy fights of 2026 are not happening in congressional hearings or corporate boardrooms. They are happening in glossaries.
Consider the term "open source." For decades it meant something precise: code that anyone could inspect, modify, and redistribute. Then Meta released Llama with restrictions on commercial use and called it "open." Mistral followed suit. Now regulators drafting AI legislation must decide whether "open" means what it always meant or what trillion-dollar companies want it to mean. The EU's AI Act uses "open source" as a carve-out for lighter compliance burdens — so the definition isn't academic. It's worth billions.
The vocabulary of evasion
Language has always been a tool of corporate positioning, but AI's novelty gives incumbents unusual power to shape the lexicon before critics can organize. "Hallucination" is a masterpiece of deflection: it implies the model is having a vivid experience rather than simply generating false information. "Alignment" sounds like yoga. "Foundation model" suggests bedrock stability rather than a system that confidently invents Supreme Court citations.
The term "artificial general intelligence" deserves special scrutiny. OpenAI's corporate structure hinges on whether AGI has been achieved — at which point Microsoft's commercial license reportedly terminates. So who decides? OpenAI's board. The company has every incentive to keep the definition as fuzzy as possible for as long as possible. "AGI" is not a scientific threshold; it's a contractual tripwire dressed up as philosophy.
When glossaries become law
Regulators are catching on, slowly. The EU's AI Act defines "high-risk" systems with specificity that took years to negotiate. The US has been less rigorous. Executive orders reference "dual-use foundation models" without precision, leaving enforcement to agencies that lack technical staff and institutional memory. China's approach is characteristically direct: the Cyberspace Administration publishes approved definitions and expects compliance.
The stakes extend beyond compliance costs. If "synthetic media" is defined narrowly, deepfakes of political candidates may escape disclosure requirements. If "personal data" excludes inferred attributes, the entire behavioral-advertising ecosystem remains untouched. Every glossary entry is a policy choice masquerading as neutral description.
The missing glossary
What the industry needs — and conspicuously lacks — is a glossary written by people without equity stakes in the outcome. Academic efforts exist but lack authority. Standards bodies move too slowly. Journalism defaults to whatever terminology the press release uses. The result is a discourse where the most powerful actors also control the dictionary.
This is not a call for some impossible linguistic neutrality. It is a recognition that when a company says its model "reasons," it is making a claim that deserves scrutiny, not transcription. When a regulator writes "frontier AI" into law without defining it, they are outsourcing policy to whoever fills the vacuum.
Our take
The glossary is never just a glossary. It is the first draft of regulation, the frame through which journalists explain stakes to the public, the vocabulary that venture capitalists use to justify valuations. The AI industry's rapid lexical innovation is not a sign of intellectual vitality — it is a land grab. The companies that define "safety," "openness," and "intelligence" will shape the rules that govern them. Everyone else is arguing in a language they did not choose.
