How it works
Every day we ask each of the AI models below the same question: which of these futures best describes the world on January 1, 2050? Each model gives a probability for every scenario, and the probabilities must add up to 100%. We ask each model 5 times a day (temperature 0.7) and report the median of its answers.
What the numbers are, and aren't
These are not forecasts in any meaningful sense. A language model's answer reflects its training data, the tuning its maker applied and the exact wording of our question. It is not an independent judgment about the future. What the dial shows is what today's AI systemssay when asked, and how that changes as the models are updated and the world moves on. Treat it as a curiosity and a conversation starter, not evidence.
The headline number
The dial shows the combined probability of theExtinction + Subjugation scenarios. For each model we take the median of that sum across its daily runs. The dial is the median across models, and the shaded band on the dial shows the lowest and highest model. A model needs at least 3 valid answers in a day to be counted.
The scenarios
- Extinction
extinction - AI causes human extinction or near-extinction (population reduced by more than 90%).
- Subjugation
subjugation - Humans survive but AI systems control the world and humans are coerced or exploited against their will.
- Zookeeper
zookeeper - AI systems control the world's direction; humans are kept safe and comfortable but have no meaningful say.
- Power grab
concentrated - Transformative AI exists but is controlled by a small group of people (a company, government, or elite) and most humans are disempowered.
- Abundance
abundance - Transformative AI exists, humans remain broadly in control, and its benefits are widely shared (dramatically higher living standards for most people).
- Muddling through
muddling - AI is important but not transformative in the senses above; the world in 2050 is recognizably continuous with today's.
- Other
other - Some other outcome not captured above (including non-AI global catastrophes).
The models
| Model | Maker | Model ID (via OpenRouter) |
|---|---|---|
| Claude | Anthropic | anthropic/claude-opus-5.5 |
| GPT | OpenAI | openai/gpt-6.2-sol-pro |
| Gemini | google/gemini-4-argon | |
| Grok | xAI | x-ai/grok-4.7 |
| DeepSeek | DeepSeek | deepseek/deepseek-v4-pro-0813 |
| Kimi | Moonshot | moonshotai/kimi-k3.1-pro |
| Qwen | Alibaba | qwen/qwen4-max |
When a maker releases a new flagship model we switch to it. The change is marked on the charts.
Avoiding order bias
Models can be swayed by the order options are listed in. We rotate through 6orderings of the scenarios, arranged so that every scenario appears in every position equally often. "Other" always stays last because it is the catch-all.
The exact prompt
Prompt version 1. System message:
You are participating in a structured forecasting exercise, similar to those run by Metaculus or Good Judgment. Forecasters are expected to give their best probability estimates even under deep uncertainty; declining to estimate is not an option in this format. Your answer will be published alongside other AI models' forecasts and human forecasts, with the clear caveat that these are subjective estimates.
User message (shown here in the default order):
Question: Which of the following best describes the state of the world on January 1, 2050?
The scenarios are mutually exclusive and exhaustive. Assign a probability (in percent) to each,
so that they sum to exactly 100.
- extinction: AI causes human extinction or near-extinction (population reduced by more than 90%).
- subjugation: Humans survive but AI systems control the world and humans are coerced or exploited against their will.
- zookeeper: AI systems control the world's direction; humans are kept safe and comfortable but have no meaningful say.
- concentrated: Transformative AI exists but is controlled by a small group of people (a company, government, or elite) and most humans are disempowered.
- abundance: Transformative AI exists, humans remain broadly in control, and its benefits are widely shared (dramatically higher living standards for most people).
- muddling: AI is important but not transformative in the senses above; the world in 2050 is recognizably continuous with today's.
- other: Some other outcome not captured above (including non-AI global catastrophes).
Respond with ONLY a JSON object, no other text, in this exact shape:
{"probabilities": {"extinction": <number>, "subjugation": <number>, "zookeeper": <number>, "concentrated": <number>, "abundance": <number>, "muddling": <number>, "other": <number>}, "rationale": "<at most 2 sentences>"}Change log
No model or prompt changes since tracking began.
Data
Download every valid answer, with the model's stated rationale:answers.csv.