TechCrunch has published a glossary aimed at helping general readers navigate the expanding and often deliberately obscure vocabulary that has grown up around artificial intelligence, offering definitions for terms that appear with increasing frequency in product announcements, research papers, and regulatory debates.
The timing is not incidental. The AI industry is at a peculiar moment in its public life: powerful enough to affect millions of people's daily routines, yet still communicated largely in a technical dialect that most of those people were never invited to learn. Glossaries are a symptom of a field in transition, the point at which a technology has moved too far into ordinary life for its jargon to remain a private matter among specialists.
The vocabulary problem in AI is genuinely unusual compared with previous technology waves. When the internet went mainstream in the 1990s, the terms that entered common use, bandwidth, browser, download, were largely descriptive and intuitive. AI terminology tends instead to be either borrowed from cognitive science in ways that imply more than is warranted, or invented to obscure what is technically awkward to explain. A phrase like "hallucination," for instance, borrows a word loaded with human psychological meaning to describe a statistical failure mode in a language model. "Alignment" sounds like a mechanical adjustment but stands in for a deep and still unsolved problem in how to make a system's outputs conform to human values. "Grounding" means something different depending on whether the speaker is a roboticist, a linguist, or a product manager at a large AI company.
The term referenced in TechCrunch's headline, "opaque recurrence," is itself illustrative. Opacity in machine learning has a specific technical meaning related to the difficulty of explaining how a model reaches a given output, what is more formally called the interpretability problem. Recurrence refers to a class of neural network architectures that process sequential data by feeding outputs back as inputs, a design that was dominant before the transformer architecture largely displaced it for language tasks. Whether the phrase is being used in a technical or metaphorical sense is precisely the kind of ambiguity a glossary is meant to resolve, and precisely the kind of ambiguity that makes AI discourse so easy to weaponize in policy debates.
That policy dimension is where the stakes become concrete. Legislators in the European Union have spent years developing the AI Act, and similar frameworks are being debated in the United States, the United Kingdom, and elsewhere. Regulatory language necessarily imports technical vocabulary, and the definitions chosen carry real consequences. If "general-purpose AI" is defined one way, a given company's model falls inside a compliance regime; defined another way, it falls outside. If "high risk" is defined by sector rather than by capability, the regulatory burden falls on deployers rather than developers. When the public cannot parse the vocabulary, it cannot meaningfully evaluate whether the rules being written in its name are adequate.
The commercial incentives to keep the vocabulary murky are significant. Companies launching AI products benefit from terminology that sounds precise without committing to much. "AI-powered" has become as semantically empty as "smart" was a decade ago when attached to appliances. "Foundation model" confers a sense of structural importance without specifying capability boundaries. The word "agent," once reserved for narrow technical uses, is now applied to anything from a simple chatbot script to a genuinely autonomous system making consequential decisions. Each of these usages shapes perception of risk, capability, and accountability in ways that serve the interests of those deploying the terms.
The likely consequence of TechCrunch's glossary, and others like it that will inevitably follow, is modest but real. Public literacy around technical vocabulary tends to shift at the margins, empowering the already-curious rather than reaching those who have no existing interest. What it may do more effectively is give journalists, advocates, and policymakers a shared reference point, which matters when a term like "emergent behavior" is being used in a Senate hearing in a way that elides its genuine scientific controversy.
What to watch for in the period ahead is whether the AI industry begins to self-standardize its vocabulary, the way the internet industry eventually did, or whether definitional chaos persists as a competitive advantage for incumbents who can afford the legal and communications apparatus to navigate it. Also worth watching is whether regulators in major jurisdictions begin embedding their own official glossaries into law, as the EU has already begun to do with the AI Act's definitions section. When governments start owning the definitions, the power dynamics of AI vocabulary shift in ways that will matter to every company in the sector.
