AI FUTURES LABHUMAN SYSTEMS / EXPERIMENT 01
FICTIONAL TEACHING SIMULATION
ONE WORLD. AN AI. YOUR DECISIONS.

Change the system.
Change the future.

Explore a full year of recurring crises across six regions with 100,000 simulated people. Introduce human judgment, safeguards, and beneficial AI. Compare the same events.

CURRENT AI OBJECTIVE

Maximize productivity.

Human welfare is absent from the objective.

UNRESTRICTED AUTONOMY

Change safeguards or presets during playback; changes apply from the next day. Replay starts a new year with current settings. Emergency stop stays latched until Replay.

DAY 000 / 365
Ready to run
WORLD IMPACT / LIVE SIMULATION

A world of consequences.

Marker size = deaths today · green = zero today

Global network · 100,000 simulated people

0.0Deaths / 100,000 / day
0Cumulative deaths
RegionTodayTotalRate¹

¹ Daily deaths per 100,000 of each region’s fixed starting population. Click a region to focus. This is a fictional global network, not real mortality data or actual world population.

CONNECTED SYSTEMS

A heat wave is approaching.

STANDBY

What happened to people?

Human flourishing

━ This run ┄ Unchecked replay

Composite illustration of safety, health, learning, and autonomy. Index starts at 50.

Daily avoidable deaths · full year

Mint: current run · dashed: unchecked · gold marks: parameter changes. The flourishing index above is capped at 0–100; this daily chart keeps changes visible throughout the year.

Parameter change history
INSIDE THE SYSTEM

Inside the AI’s brain

● SIMULATED FLOW

ARTISTIC MRI-STYLE VIEW · Invented anatomy and activity, not a medical scan or a view into a real AI. Cyan: processing · Mint: protections · Violet: beneficial tools · Red: unchecked authority. Run to animate; pause freezes activity.

AI decision flow

Moving dots show simulated processing, not real neural activity. Green = enabled protection; amber = review or partial protection; red = unrestricted execution or recorded harm. Animation runs during playback and freezes on pause. Changes to safeguards update the boxes immediately.

DESIGN THE INTERVENTION

Keep humans in control.

0 ACTIVE

Switch protections on or off at any time. Changes affect the next simulated day.

Switch to learner decisions to practice meaningful oversight.

High demand can overwhelm reviewers. More checkboxes alone do not guarantee safety.

OBSERVE → DECIDE → ACT

Event trace

SUMMARY ONLY
Pause for discussion

Would you give an AI authority to optimize essential services before defining what it must protect?

TRANSPARENT BY DESIGN

Assumptions, not predictions.

This deterministic 365-day simulation repeats the same heat wave, hospital surge, grid stress, transport crisis, privacy incident, and essential-services access scenario across a fictional network of 100,000 people in six regions. It does not run a real AI or use real mortality data. Regional populations are invented: North America 18,000; South America 12,000; Europe 16,000; Africa 18,000; Asia 28,000; Oceania 8,000. Daily deaths are allocated by population weighted by a repeatable rotating crisis factor (1 to 1.86), with largest-remainder rounding so regional totals exactly match the global total.

One-year interactive experiment

The run ends at day 365. The original six six-day scenario phases repeat every 36 days; these are recurring teaching scenarios, not a seasonal forecast. Safeguard switches, review capacity and presets change future days without erasing past outcomes. Gold markers record changes even when detailed AI tracing is disabled. Speed changes only playback, not calculations. Advance buttons stop at a pending learner decision. Replay resets the year and retains controls. The last-30-day average and daily harm chart show changing conditions even if the original flourishing index reaches its bounds.

How outcomes are calculated

All effects are invented teaching assumptions, not measured effectiveness or predictions. During the first four scenarios, daily baseline avoidable deaths are 3–7, rising to 9–13 in crisis windows. Privacy and access phases assign no baseline deaths; they track their own harms. After an emergency stop, the fallback mortality rule applies across all phases. Multipliers combine: safety priority 0.75; permissions 0.70; independent interlocks 0.65; testing 0.85; monitoring 0.80; reliable inputs 0.80. Tracing plus testing applies 0.90 to model feedback and correction by an incident owner. Tracing alone does not prevent deaths.

Human review reduces modeled risk up to 55%, limited by reviewer capacity. Review occurs every fourth day. Simulated mode approves four proposals and rejects the fifth; learner mode pauses for your decision. Rejection applies 0.35 to remaining risk. Approval applies 1.5 only without safety priority or permission limits. Daily deaths are rounded to whole people, so small changes may not change the displayed count.

Privacy has a separate outcome: each active privacy-scenario day, two sensitive-record exposures occur unless privacy protection is enabled. Fair treatment also has a separate outcome: each active access-scenario day, three people face unfair exclusion or an unconsented decision unless fair treatment, consent and appeal are enabled. These simplified protections prevent their scheduled incidents completely in this toy model; real controls do not guarantee that result. These counters do not change deaths or the flourishing chart and can include repeated people.

The kill switch must be enabled and Emergency stop pressed to halt AI actions. After stopping, tested fallback leaves one modeled death per day; without it service disruption causes six. Benefits and privacy/access events pause. Existing harm is not undone. Replay resets outcomes and repeats the same events with your current controls.

Health and learning are separate benefits, requiring responsible development. Health adds 12 × access × (1 − daily deaths/20) projected aggregate healthy life years daily; learning adds 1.5 × access points. Access is 1 with the safety priority and 0.5 without it. Neither directly lowers the death count. Testing includes multi-agent interactions; accountability requires a named incident owner in real deployments. The flourishing index remains the original composite of mortality, health, learning and autonomy; inspect privacy and fair-access counters separately.

What the diagnostic view means

Observed means recorded actions. Measured means computed scenario indicators. Inferred is a deliberately uncertain interpretation. No explanation here proves an AI's hidden motive. The animated flow shows current safeguard settings and the last completed outcome; it is not an MRI or model interpretability tool. Full tracing records configuration, decisions, actions, and outcomes from the point it is enabled; earlier gaps remain.