Two narratives have dominated the conversation since November 2022: doom (extinction, x-risk, alignment) and hype (AGI, transformation, the race). They look opposed. They both serve the firms building AI — doom justifies consolidation, hype justifies capital and policy alignment.
What gets crowded out is the third frame — the harms with names, dates, dollar amounts, and victims. This investigation builds the documentary case for that frame across four acts.
Eighteen definitions, no consensus. The same word covers a thermostat, a regression model, and ChatGPT — depending on who is regulating, who is publishing, and who is selling.
A typical Gemini prompt = 0.24 Wh; a steak ≈ 570× a heavy AI user’s yearly chatbot footprint. The real concern is not joules per query — it is $680B in 2026 hyperscaler capex against $40–60B in revenue.
Most AI evaluation orgs share one funder. Mapping which trackers measure what, who pays them, and which findings should make it into your priors before any policy debate.
Doom and hype look opposed. Both serve the firms building AI. The harms with names, dates, and victims — labor, surveillance, deepfakes, deployed welfare algorithms — are crowded out of both narratives.
Acts I–III are research-complete and ship next. Act IV — the integrating thesis — is live now because it is the one that reframes the others.
The investigation argues that “AI” as a singular subject crowds out the actual harms. The map argues the same thing from the opposite side: it crowds out the actual work — the dozens of organisations across four continents publishing open weights under permissive licenses. Updated when the leaderboards move.
The full research notes — every claim with a citation — live in the repository at research-notes/. Each act page links the relevant deep-dive directly.
How this is maintained → — source hierarchy, the staleness-badge system, what is intentionally not in scope.
Sources flagged mid-research that deserve a deep dive before being folded into the four acts. A living queue — items move out once they become load-bearing citations.