The record
Evidence Ledger
Claims are cheap. Entries below carry a date, a tag, and a primary source. Corrections welcome — the Ledger has a public changelog.
Showing 12 of 12 entries · sorted newest first
RSS feedThe Model Capability Initiative captured employee keystrokes and screenshots to feed training pipelines; over a thousand staff petitioned; the same week, 8,000 were laid off.
Reported in press coverage and an internal staff petition, 2026
Pixar's structured mechanism for candid, hierarchically unfiltered feedback: a loading device that stores social potential energy through resistance. The films are the projectiles.
Reported in Pixar case studies and interviews
Replaced roughly 700 agents with a chatbot handling 2.3 million conversations; satisfaction eroded, repeat contacts climbed, and the company began rehiring. Cost had been, in its chief executive's words, a too predominant evaluation factor.
Reported in company statements and press coverage, 2024–25
Orgvue's survey of 1,163 C-suite leaders: 39 per cent had laid off staff to deploy AI; 55 per cent of those now call it the wrong decision; 34 per cent report staff quitting as a direct consequence.
Orgvue survey of 1,163 C-suite leaders, 2025
Stanford Digital Economy Lab finds a 13 per cent relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations — the apprenticeship rungs disappearing first.
Stanford Digital Economy Lab, 2025
Microsoft Research and Carnegie Mellon: in roughly 40 per cent of AI-assisted tasks, workers reported exercising no critical thinking at all.
Microsoft Research & Carnegie Mellon University, 2025
Anthropic's AI fluency research: in fewer than nine per cent of interactions do users pause to verify model output — systemic overconfidence in numerical form.
Anthropic, AI Fluency research, 2025
Salesforce routes routine queries to AI at an 83 per cent autonomous resolution rate while steering complex, trust-laden cases to humans — and discovers customers ask AI what they were too embarrassed to ask people, a new class of knowledge.
Salesforce product data and case studies, 2025
Granola enriches the human note-taker's own jottings rather than replacing them: machine text in grey, human judgement in black, a verifiable hybrid record.
Granola product design writeup, 2025
Shumailov and colleagues in Nature: models trained on model output degenerate as the distribution narrows; rare knowledge and dissenting formulations — the tails — vanish first.
Shumailov et al., Nature, 2024
With widespread reliance on AI-generated summaries, modelled public beliefs deviate 2.3 times further from ground truth as perspective diversity narrows.
Peterson, 2024
Dell'Acqua and colleagues at Harvard Business School: consultants using AI inside its frontier improved by around 40 per cent; outside the frontier, performance fell 19 per cent — discernment, not access, decides the sign.
Dell'Acqua et al., Harvard Business School, 2023