The Language Behind the Hype

TechCrunch published its living AI glossary in April 2026 — a useful signal, not because the definitions are surprising, but because they reveal exactly how unsettled the vocabulary still is. Three organizations, three definitions of AGI. “AI agent” explicitly flagged as meaning different things to different people. For a trade-press audience, that ambiguity is a story. For an architect specifying an AI-enabled coordination tool, or a BIM manager evaluating an Autodesk or Graphisoft agent integration, it’s a liability surface.

←TODAY: The EU AI Act has been in force since August 2024, with phased application running through 2026–2027 — and it carries its own legally binding definitions of “AI system” and “general-purpose AI model” that do not map cleanly onto vendor terminology.
→3012: In a Zurich where building-permit logic runs on autonomous reasoning chains, imprecise terminology in a procurement clause becomes enforceable ambiguity in an SIA contract.
Fulcrum: The gap between marketing vocabulary and regulatory vocabulary is where professional exposure lives — and closing it starts with reading both.

Here is what the glossary actually tells a practitioner, term by term — and where each definition breaks down at the desk.

AGI is the easiest to dismiss as irrelevant and the hardest to ignore. OpenAI’s Sam Altman describes it as “the equivalent of a median human you could hire as a co-worker.” OpenAI’s own charter defines it as systems that “outperform humans at most economically valuable work.” Google DeepMind sets the bar at “at least as capable as humans at most cognitive tasks.” These are not minor variations — they imply entirely different capability thresholds and, consequently, entirely different procurement risk profiles. When a vendor claims their tool approaches AGI-level performance on code generation or clash detection, which definition are they using? Ask.

AI agents are where AEC exposure is most immediate. The TechCrunch definition — an autonomous system that draws on multiple AI models to complete multi-step tasks — is functionally accurate. But as the glossary itself concedes, “infrastructure is still being built out to deliver on envisaged capabilities.” Autodesk and Graphisoft are both shipping features marketed as agentic; what they actually deliver today is closer to orchestrated prompt chains with partial autonomy. That distinction matters when you are deciding how much human review to build into a specification workflow.

Chain-of-thought reasoning is the most technically stable term in the set. When a reasoning model — derived from a large language model and optimized via reinforcement learning — breaks a problem into intermediate steps before outputting an answer, it takes longer but produces more reliable results in logic-heavy tasks. BIM coordination, clash resolution logic, code-compliance checking: these are exactly the multi-step reasoning tasks where chain-of-thought models outperform their faster, single-pass counterparts. The trade-off is latency and compute cost. Per ISO/IEC 22989:2022, the international AI vocabulary standard, this maps to what the standard calls “inference with intermediate representation” — a framing that carries more contractual weight than “chain-of-thought” in a European procurement document.

Hallucinations — referenced in the TechCrunch title but not fully defined in the excerpt — deserve the sharpest treatment here. A hallucination is a model output that is confidently stated and factually wrong. In a consumer context, this is annoying. In an AEC context, it is a professional liability event: an AI-generated SIA 118 reference that cites a non-existent clause, a structural load calculation with a plausible-looking but incorrect intermediate step, a fire-protection specification drawn from the wrong cantonal ordinance. No glossary entry reduces that risk; only a verification protocol does.

Atelier: In PAZ cohort workflows, we treat AI-generated technical outputs as draft-zero material — always subject to a named human sign-off before any document reaches the BEP or Bauleitung stage. Chain-of-thought models improve the draft; they do not replace the review. Build that gate explicitly into your team’s Atelier-Code before the tool ships to a live project.

The EU AI Act’s official vocabulary — available at ec.europa.eu — is the reference frame Swiss and European practitioners actually need alongside industry glossaries. “General-purpose AI model” under the Act is not the same as AGI under OpenAI’s charter. If you are procuring, specifying, or deploying an AI system in a regulated context, read both documents side by side and flag the gaps to your Bauleitung and legal counsel now, not after go-live.

Pull up the EU AI Act’s Article 3 definitions this week. Compare three terms — AI system, GPAI model, and high-risk AI — against whatever vocabulary your current AI tool vendor uses in their documentation. The mismatches you find are your next agenda item.

Source: TechCrunch