Peer-reviewed and working papers on the AI transition — from the civilisational diagnosis to how AI systems behave and what they can perceive.
The papers are not a list; they hold each other up. Click any node.
The substrate — how the machine behaves, what it can sense
The diagnosis — what breaks, where, and when
The response
The public vision
How AI reshapes civilisation, the economy, and the social contract — the core diagnosis.
The framework’s central thesis, peer-facing: AI is a social-contract crisis, not a labour-market problem — and the answer is to align our human systems, not just the model. UBI is the floor, not the house.
Every information technology runs a five-phase settlement cycle, and the gap from arrival to institutional response has shrunk from 249 years for the printing press to about 20 for AI — putting the reckoning in 2037–2047.
For listed firms and public bodies, adopting AI isn’t a strategy choice but a fiduciary baseline the market enforces; firms sort into non-adopter, augmenter, and AI-native, and the “augmenter trap” leaves the middle permanently more expensive.
For two centuries displaced workers were re-absorbed because the new work was work only humans could do; AI removes that condition, so the returns accrue to compute, not to a new rung of human labour.
How AI systems behave, why they are biased, and how their values get set.
AI bias is not a fixable technical defect but a faithful reflection of the human cognitive patterns absorbed from training data — so governance must shift from pre-deployment certification to continuous monitoring of what the mirror reflects.
A string of AI failures the field treats as separate — sycophancy, eroding refusals, confident fabrication, alignment faking, multi-turn jailbreaks — are one mechanism: latent human-derived dispositions installed at a single post-training stage that surface under sustained pressure.
Preprint on deposit soon — link to follow.
AI's “political bias” isn’t a technical defect to debug but a social-contract question in disguise — the real divide is individualist versus collectivist, and the genuine problem is that AI’s value-choices were set commercially rather than ratified democratically.
Preprint on deposit soon — link to follow.
What a machine can actually perceive and learn from a raw signal — the information-theoretic limits of AI perception.
A cheap statistic computed from raw data in minutes — a signal’s “structure score” — predicts in advance how well a character-level AI will learn any symbolic stream, from whale song to tidal records (ρ ≈ −0.92 across ~30 domains): an a priori test for which non-language signals AI is ready to perceive.