Two narratives, opposite in posture, identical in effect.
Doom and hype look opposed. Both serve the firms building AI — doom justifies consolidation, hype justifies capital. The harms with receipts are the ones not being discussed.
Since November 2022 the AI conversation has been organized around two framings. Doom warns of extinction, of unaligned superintelligence, of catastrophic misuse. Hype promises imminent AGI, compressed centuries of progress, abundance.
They look opposed. Read closely, they are the same instrument. Doom justifies consolidation — only the firms with the most compute can build it “safely.” Hype justifies capital and policy alignment — the race is real, resistance is futile, the brakes must come off. Both outputs serve the same set of firms. Both push the same category of concern out of the room: the harms that are already happening, with names and dates and dollar amounts and victims.
The remainder of this page is the documentary case for the figure above. Each load-bearing claim is wrapped in a card you can expand for the receipt.
A small donor network, a coherent ideology, an outsized policy footprint.
AI x-risk discourse is the institutional output of a small, ideologically coherent subculture — funded by a handful of tech-fortune donors — that captured the policy frame in 2023.
Inevitabilism is a balance-sheet item.
"AGI is coming, the only question is who builds it first" is rhetorical infrastructure that unlocks capital, policy concessions, and attention — and the actors deploying it have direct, measurable financial interest in its believability.
The harms with names, dates, and victims.
Each cell in the central box of the figure is a real event with a documentary trail. None of the people involved consented; none of the deployments had a prior public debate.
Below: a few of the cases shown above, expanded with the primary-source receipt. The page would run to thirty ClaimCards if every pin got its full court file. The atlas lets you click each pin for the same level of detail, in place.
Two different fears wearing the same word.
The cleanest empirical signature of the pincer is the gap between what publics actually report worrying about and what elite policy discourse covers.
Path dependency. Defaults set by procurement.
Every quarter the buildout proceeds without addressing concrete harm, those harms become integrated infrastructure that requires winning a political fight to reverse.
ImmigrationOS, Lavender, Robodebt — each is harder to remove than it would have been to prevent. Democratic deliberation has not happened on AI in welfare, courts, hiring, healthcare. The defaults are being set by procurement decisions, not legislation. The shape of governance — EU AI Act enforcement, US executive orders, China's algorithm rules — is being written now. What is missing from the frame will be missing from the law.
Consent was never given. ~500K Australians (Robodebt). ~26K Dutch families (toeslagenaffaire). ~37K Palestinians (Lavender). 47M views in 17 hours of Taylor Swift deepfakes. None of them consented; none had a meaningful prior public debate.
What meaningful AI accountability would look like.
AI accountability is not a future problem awaiting a superintelligence. It is a present problem about wages, consent, due process, and concentration.
- 01
Wages, classification, and trauma support for data workers as a precondition of model deployment.
RLHF labor recognized as employment; psychological-health benefits at the level provided to Meta and YouTube content moderators; right to organize; transparent supply chains the way conflict minerals are now disclosed.
- 02
Use bans, not just disclosures, in welfare, child protection, immigration, courts, and policing.
Until per-deployment audits show non-discriminatory error rates lower than the human baseline they replace. Robodebt and toeslagenaffaire as the precedents.
- 03
Per-deployment consent for biometric capture — facial, gait, iris, voice.
Criminal liability for noncompliant scraping (the Illinois BIPA model, scaled).
- 04
A binding international ban on lethal autonomous weapons that select and engage humans without meaningful human control.
161+ states have already supported this at the UNGA.
- 05
A binding norm against AI in nuclear command, control, and communications.
Joining the Biden–Xi Nov 2024 statement and the Nov 2025 UN resolution rather than opposing it.
- 06
Statutory liability for non-consensual intimate imagery and for foreseeable mental-health harms to identified users.
The Raine and Character.AI cases as wedge precedents.
- 07
Compute and cloud antitrust that breaks the NVIDIA → hyperscaler → frontier-lab → application stack vertical.
Treating frontier compute the way telecoms and electricity are treated — common-carrier obligations, capex transparency, no exclusive deals.
- 08
Training-data provenance with an opt-in regime and statutory damages for unconsented inclusion.
Suno/Udio settlements as the market-based floor, not the ceiling.
- 09
Worker codetermination over workplace AI.
Algorithmic pacing, sentiment scoring, deactivation made bargainable subjects in unionized workplaces and disclosable in non-union ones.
- 10
Election-period synthetic-media provenance requirements with platform takedown obligations.
Romania's annulment as the warning. Democratic oversight of recommender amplification.
- 11
Public, independent compute — sovereign or academic.
Sufficient to allow non-corporate evaluation of frontier models, so independent science exists about the systems being deployed.
- 12
Procurement-as-policy.
Any AI deployed in public services subject to algorithmic impact assessment, public registry, redress mechanism, and sunset review — defaults set by legislatures, not vendor RFPs.