AI SAFETY AND ALIGNMENT
Keep the benefit.
Question the permission.
A twice-weekly series by Adrian Dunkley. Each issue adds a method you can challenge, a tool you can use and a canvas for the decision.
New issues on Tuesdays and Fridays. The opening essay introduces my proposed 4S cycle. Later issues test its assumptions against economics, physics, history and safety engineering.
An uncontrolled acceptance of AI.
Measure the exposure between giving AI authority and stopping a harmful action.
Read the opening issueThe proposed 4S cycle
Specify the purpose. Stress-test the behaviour. Safeguard the action. Scrutinise what happens.
Original educational synthesis by Adrian Dunkley, drawing on established methods. Version 0.1 is a proposal to evaluate.
The direction of the series
These are planned issues. Their arguments and models will be checked against the evidence and previous editions before publication.
| Planned date | Issue | What it adds |
|---|---|---|
| 2026-11-03 | Lessons from Colonization: AI Misalignment. | Power, consent, value extraction and exit rights |
| 2026-11-06 | The Self Sharpening Blade: AI adoption | Feedback loops that improve or compound harm |
| 2026-11-10 | Trending Public Hate Private Addiction | The gap between expressed attitudes and behaviour; the headline poses a hypothesis |
| 2026-11-13 | Segregation of AI Policies: the fastest way of losing sight of AI use | Conflicting rules, reporting friction and hidden adoption |
| 2026-11-17 | Winnie Pooh lessons: the honey pot is your objective function | Reward hacking and incomplete objectives |
| 2026-11-20 | Brownian motion and the AI that wanders past your rules | Cumulative drift and first-passage thresholds |
| 2026-11-24 | Entropy will not tell you whether your AI is honest | Uncertainty measures and false reassurance |
| 2026-11-27 | Your safest AI may be the one you can afford to switch off | Irreversibility, dependency and option value |
| 2026-12-01 | The average passed. The severe failure disappeared. | Rare severe failures hidden by average performance |
| 2026-12-04 | Borrow the cockpit checklist. Leave the pilot myth behind. | Independent barriers and human workload |
| 2026-12-08 | A safety framework should survive an argument with its author | Falsifiable evaluation of 4S and revision criteria |
The historical comparison will examine institutions, consent and extraction without equating AI adoption with the human violence of colonial rule. Mathematical models will state their assumptions and limits.