Elon Musk’s AI Predicts Americans Will Be $800 Richer Under Harris
My new research shows that OpenAI and (especially) xAI models forecast better economic, environmental, and poverty outcomes under a Harris administration.
In a new large-scale experiment, I asked leading AI models to forecast key metrics for 2025 under different administrative scenarios. Recent research validates that these models can make meaningful economic predictions when properly prompted.
Musk’s Grok AI produced the most optimistic projections for a Harris administration across each metric I tested:
- Air quality: PM2.5 pollution would be 1.51 µg/m³ lower than under Trump (compared to GPT-4o’s prediction of 1.26)
- Poverty: Supplemental Poverty Measure 3.42 percentage points lower (vs GPT-4o’s 1.76)
- Individual prosperity: Real GDP per capita $802 higher (vs GPT-4o’s $388)
Grok predicted stronger positive effects than GPT-4o by factors ranging from 1.2x to 2.1x, and the differences were statistically significant.
Why I did this
This project started at a Manifold Markets event in DC last week, where discussions about political forecasting got me thinking about how AI models process historical patterns to generate predictions about potential futures.
While we should take individual predictions with appropriate skepticism, I find the consistency across independently trained models intriguing. Even GPT-4o-mini’s more conservative estimates showed similar directional patterns.
To validate these findings, I ran 500 trials for each model and metric using their APIs. The results were statistically significant and remarkably stable across different prompting approaches.
The bigger picture
This experiment reveals something fascinating about how different AI models — trained on distinct datasets by different organizations — process economic and policy information.
At PolicyEngine, we build microsimulation models that precisely estimate how specific tax and benefit policies affect household budgets. These AI forecasts raise intriguing questions about how large language models might complement such tools — offering broader context while our models provide granular policy analysis.
You can explore my code and read the full working paper here: github.com/MaxGhenis/llm-presidential-outcome-forecasts
What do you think about using AI for policy forecasting? What other metrics should we test? Let me know in the comments.
Technical note: The study used narrative prompting techniques, asking models to describe scenarios as if from a future perspective. Full methodology and results are available in the working paper.
