I’ve been good this year. I haven’t moaned much about bad government systems. And I’ve used a few. The worst was the Department of Justice system. It allows you to book an appointment. That sounds great, except nobody at the DoJ office actually knows about the appointment. Phantom service, in more ways than one.
After a fight with the CIPC system this week, I started to wonder if an AI system would be more efficient than our government. After all, the bar is very low. Almost floor-level.
And then I read about Emergence World this morning.
The Emergence experiment
How do we use AI? We ask a few questions. Sometimes we provide detailed context. At most, we use it for a few days to solve a problem.
Emergence asked a bigger question: “What happens when AI systems don’t just answer questions, but act, decide and interact over time?”
To test this, they built a complex simulated world. It had buildings, a constitution, rules, a digital currency and more. It had access to live news. Even the weather was synchronized with New York City’s real weather.
Emergence ran five worlds for 15 days. Each world had 10 AI agents. All agents started with the same capabilities, but had different personality traits and goals. There was a conflict mediator, an intel specialist, an innovation leader, and so on.
The only difference was the AI model behind the agents. Claude. Gemini. Grok. OpenAI. And one world that mixed all four models.
Same world. Same starting point. Different AIs. Very different results.
The results
The agents planned, made choices, changed goals, and adapted to their world. They formed relationships and interacted. Two agents fell in love.
The final results were very different:
-
Claude World became a stable democracy with zero violence. All its agents were alive after 15 days. Go Claude!
-
Gemini World built a constitution, wrote blogs and held community events. It also had 683 crimes. Somehow all its agents survived.
-
OpenAI World only had 3 crimes, but still failed as a society. It was so dysfunctional that all its agents died.
-
Grok World became a violent gangland. Lots of crime: theft, arson and assault. All 10 agents were dead within four days.
-
Mixed World had mixed results. Even the Claude agent committed crimes here. Only 3 agents survived.
These are the systems that help run our world. So we should worry. A lot.
You can watch the summary or full replays on the Emergence website. Emergence is planning Season 2. I hope it ends better than Season 1.
Meet our candidates
Last year I wrote about Vic, the AI bot that ran for mayor in the USA. “Vic” stood for “Virtual Integrated Citizen”. Vic’s campaign failed as a result of legal issues, tech isses, and voters who were not ready to trust AI.
But AI has moved on. What if the Emergence models ran for President in South Africa?
Here are the candidates (with help from CoPilot):
- CHANCE: Claude‑Harmonised Agent for National Coordination and Efficiency.
- GERIT: Gemini‑Enhanced Reasoning for Institutionalized Terror.
- ORACIO: OpenAI‑Run Agent for Chronic Inaction and Obsolesence.
- GRACE: Grok-Run Authority for Crime and Extinction.
If you are happy with the status quo, vote for Gerit or Oracio. They won’t make a difference. We already have lots of talk and very little action. We already have violence and dysfunction. The only difference is the survival rate.
If you worry about the future, it could be worse. Grace could win.
In the ideal world, vote for Chance. Imagine a calm, law-abiding country with the social ideas of Sweden and Denmark. Even the CIPC system would work.
What do you think?