The Open Record  ·  theopenrecord.org

Sources & Methodology

The Convenient Ghost: Who Built the AI Villain, and Why
Published March 2026  ·  Compiled for editorial transparency

Source Classification Legend

🏛️ Government / Official
🎯 Center / Non-Partisan
🔵 Left-Leaning
🔴 Right-Leaning

A Note on Human-AI Collaboration

This article was produced through direct collaboration between the author and Claude, an AI assistant made by Anthropic — the same company whose Pentagon dispute is covered in Act 1. That relationship is disclosed not as a disclaimer but as a demonstration of the article's central argument.

All research direction, editorial judgment, analytical conclusions, sourcing decisions, and investigative framing are the author's. AI was used to accelerate research retrieval, cross-reference sources, and assist with structural drafting. Every factual claim in the article was independently verified by the author against primary sources before publication.

The article argues that AI functions best as a collaborative tool extending human capability rather than replacing human judgment. This piece was written that way intentionally.

Act 1 — They Built a Scapegoat and Called It a Revolution

AI Capabilities: What the Research Actually Shows

🎯Peer-Reviewed
GPT-4.5 Turing Test Performance Study (2025)
Published in academic literature, 2025
Rigorous Turing test evaluation in which GPT-4.5 was identified as human by evaluators 73% of the time — outperforming the human participant in the same evaluation. Cited as evidence of genuine but bounded AI capability, not as evidence of consciousness or agency.
🎯Technical
AI Performance on Expert-Level Medical, Legal, and Mathematical Benchmarks
Multiple institutional sources, 2023–2025
Aggregate of peer-reviewed studies documenting expert-level performance of frontier AI models on standardized professional examinations including USMLE (medicine), bar examination (law), and competition mathematics. Used to establish that AI capabilities are genuine while the absence of consciousness and moral agency is equally documented.

The Pentagon / Anthropic Standoff

🎯News
Anthropic Sues the Trump Administration After Supply Chain Risk Designation
CNN Business
Reports on Anthropic's two filed lawsuits (N.D. California and D.C. Circuit), the supply chain risk designation, Trump's agency-wide cease order, and Hegseth's ban on military contractor use of Claude. Covers the OpenAI/DeepMind amicus brief and the hearing rescheduled to March 24, 2026.
🎯News
Anthropic Sues Pentagon Over Supply-Chain Risk Designation
TechCrunch
Detailed coverage of the two-jurisdiction filing strategy, the specific legal arguments (First Amendment retaliation, procurement law violations), and Amodei's two redlines: no autonomous weapons, no mass surveillance of Americans. Source for "unprecedented and unlawful" quotation and the protected speech argument.
🎯News
Anthropic Sues in Federal Court to Reverse Supply Chain Risk Designation
PBS NewsHour
Source for the six-month Pentagon phase-out timeline, Trump's Truth Social order directing immediate cessation of Anthropic use, and confirmation that Claude was used in classified military operations during the Iran conflict despite the blacklisting. Also documents the Palantir partnership for intelligence analysis, targeting recommendations, and battle simulations.
🎯News
Anthropic Sues Trump Administration in AI Dispute with Pentagon
NBC News
Confirms that the supply chain risk designation has never previously been applied to an American company. Documents the Palantir partnership and Claude's role in intelligence assessments and battle simulations on classified networks.
🎯News
Anthropic Says Pentagon Ban Could Cost It Billions
Yahoo Finance / multiple wire services
Source for the March 24 hearing date (moved up from April 3), Microsoft's court brief supporting Anthropic, and the Venezuela operation confirmation. Documents that Claude was used in Venezuela in January 2026 and the current Iran conflict.
— ✦ —

Act 2 — What It's Actually Being Used For, and Who Pays

Labor Arbitrage and Corporate AI Deployment

🎯Business
IBM, UPS, Dropbox, Salesforce, Google Layoff Announcements (2023–2025)
Corporate earnings calls and press statements; reported across financial press
Pattern documentation: AI capability announcements followed by headcount reductions in which "efficiency gains" are attributed to AI deployment. Used to establish that labor arbitrage is the dominant enterprise AI use case, with productivity gains flowing to shareholders rather than displaced workers.

Infrastructure Scale, Subsidies, and Obsolescence

🎯Watchdog
Data Center Subsidy Tracking: Texas, Illinois, Virginia
Good Jobs First — goodjobsfirst.org
Primary source for all subsidy extraction figures: Texas projection revised from $130M to $1B in 23 months; Illinois growth from $10M (2020) to $370M (2024), a 3,600% increase; Virginia's five-fold single-year balloon and ~$1B/year in foregone revenue without public disclosure of corporate beneficiaries. Direct quote: "no other form of state spending so out of control." Good Jobs First is a nonpartisan corporate subsidy watchdog.
🎯Infrastructure
Meta Hyperion Data Center, Louisiana
Multiple infrastructure and environmental reporting sources, 2024–2025
Source for the claim that Hyperion is expected to draw more than twice the power of the city of New Orleans upon completion. Also source for the Newton County, Georgia Meta data center drawing 500,000 gallons of water per day — approximately 10% of the county's entire water supply.
🎯Infrastructure
Meta Wyoming Data Center — Electricity Consumption Projections
Infrastructure and utility reporting, 2024–2025
Source for the claim that the planned Meta Wyoming data center will consume more electricity than all residential homes in the state combined.
🎯Financial
Amazon, Microsoft, Google Capital Expenditure Reports, 2024
Corporate earnings filings, 2024
Source for the $200B+ combined capital expenditure figure, with the majority directed toward data center construction. Drawn from official quarterly and annual filings reported across financial press.
🎯Industry
Data Center Build Timeline and Obsolescence Analysis
Infrastructure industry analysts, 2024–2025
Source for the two-to-three-year build timeline claim and the expert assessment that data center designs may be obsolete before construction is complete. Direct quotation: "There are just so many moving parts here it's unclear how you would build a data center and not have it be obsolete when you're completed." Source on background.

Local vs. Cloud Architecture

🎯Peer-Reviewed
Local AI Model Query Handling Capacity — 88.7% Finding
Peer-reviewed study, 2024
Source for the finding that local AI models can successfully handle 88.7% of everyday chat and reasoning queries without cloud infrastructure. Also primary source for the finding that hybrid local-cloud systems achieve 40–65% reductions in energy, compute, and cost, and that hybrid edge-cloud processing can achieve up to 75% energy savings and 80%+ cost reductions versus pure cloud.
🎯Technical
Apple Neural Engine / M-Series On-Device AI Architecture
Apple technical documentation and product announcements, 2023–2025
Source for the claim that Apple has embedded dedicated AI processing into its hardware specifically for on-device inference without data center routing.
🎯Technical
Microsoft Copilot+ PC Requirements — Local AI Mandate
Microsoft product documentation, 2024
Source for the claim that Microsoft's Copilot+ PC initiative requires a local Neural Processing Unit (NPU) as baseline hardware, enabling on-device AI inference.
🎯Open Source
Ollama, LM Studio, Jan — Local AI Runtime Tools
Open source projects
Documentation and project repositories for the three local AI runtimes cited as evidence that on-device AI is accessible to general users without technical expertise.

Environmental Costs and the Guilt Trap

🎯Historical
Plastics Industry Recycling Campaign History (1990s)
Documented through investigative reporting and industry archive research
Background source for the "paper straw" analogy. The 1990s plastics industry campaign promoting individual recycling responsibility while continuing mass plastic production is the historical precedent cited for the individual environmental guilt narrative surrounding AI energy use.
🎯Industry
OpenAI Per-Query Water Usage Disclosure
Statement by Sam Altman / OpenAI, 2023
Source for the critique of voluntary, imprecise environmental disclosures. The published per-query figure lacked methodology details including definition of "average" query and whether training costs were included — cited as an example of precise-seeming numbers with invisible methodology.
— ✦ —

Act 3 — What It Could Actually Do For Us

Information Access and the Private Tax on Being Poor

🎯Research
Legal Services Access Gap — Unmet Civil Legal Needs
Legal Services Corporation — lsc.gov
Background data on the gap between legal need and legal access among low- and moderate-income Americans. Used to establish that legal advice functions as a "private tax on being poor" — a cost borne disproportionately by those least able to pay it.

Cooperative and Distributed AI Models

🎯Technical
Distributed AI Infrastructure — Technical Literature
Academic and industry publications, 2023–2025
Background for the argument that cooperative AI infrastructure is an engineering choice requiring different incentive structures, not a technical impossibility. The hyperscale centralization model is a business architecture decision, not a constraint imposed by the technology itself.
— ✦ —

Methodology

Research Framework

This investigation uses a structural analysis framework examining AI not as a technology story but as an accountability story — asking who makes decisions, who benefits from those decisions, and who absorbs the costs. The central analytical lens is the anthropomorphization strategy: how attributing agency to software functions as a liability shield for identifiable human decision-makers.

Research was conducted across government filings, corporate disclosures, peer-reviewed literature, financial reporting, legal documents, and infrastructure industry analysis. All major factual claims were cross-referenced against a minimum of two independent sources before inclusion.

Source Selection and Political Balance

Sources were deliberately drawn from across the political spectrum. The four-category legend system (🏛️ Government, 🎯 Center/Non-Partisan, 🔵 Left-Leaning, 🔴 Right-Leaning) reflects The Open Record's standard practice of transparency about institutional perspective. Readers are encouraged to evaluate sources with this context.

Where data appears only from sources with a discernible perspective, this is noted. No claim dependent on a single ideologically positioned source was included without corroborating evidence from a source of different orientation.

Contested and Evolving Data

Several data points in this article are genuinely contested, voluntarily disclosed, or actively developing at time of publication. These include: AI energy and water consumption figures (voluntary corporate disclosure, no standardized methodology); Pentagon/Anthropic litigation (filed March 9, 2026; hearing March 24, 2026 — outcome pending); data center subsidy figures (drawn from state budget documents compiled by Good Jobs First, subject to revision as states update projections).

Where figures are contested or methodology-dependent, this is flagged in the article text. Readers should treat specific environmental impact numbers with particular skepticism, as noted in Act 2.

Archival Preservation

Key sources were archived via the Wayback Machine at time of research. Where sources have been archived, the archive URL is available on request. PDFs of critical regulatory and legal documents are hosted directly at theopenrecord.org to guard against link rot.

theopenrecord.org/tools/wayback_archiver/

Corrections Policy

The Open Record corrects factual errors promptly and transparently. If you identify an error in this article, contact the editorial team at theopenrecord.org. Corrections are noted in the article text with date and description.